The NEXT LAST R600 Rumours & Speculation Thread

[Skinner]

Why is no one committing about the IQ?

[bdotobdot2]

Why is no one committing about the IQ?

Cause its hard to see thru smoke and mirrors?!? ;)

Those that might be putting up legit reviews will wait til the NDA is up and put up full fledged reviews they can take credit for.

[Shtal]

http://bp2.blogger.com/_BabjUDZIqPw/RjIc5GGObEI/AAAAAAAAACM/NsWoRGaAG-0/s1600-h/013.jpg


16 texture units = Grown-Man trying to ride /accelerate using pedals on baby bicycle :lol:
Oh! come on....

-Or-
16 texture units = You have a huge engine but trying to fit in a small car. Good Luck with that :)

I hope ATI/AMD do realize that in order to boost FPS you need 32+ texture units and 20 or 24+ ROP's "Just hope that 65nm R650 will be a big surprise for me"

320 stream processor units running on lower MHz Frequency, I rather have 128 stream processor units but running at high Frequency. It will be very hard to utilize all 320 stream running low MHz frequency.
But I still like ATI. :)

[Farhan]

320 stream processor units running on lower MHz Frequency, I rather have 128 stream processor units but running at high Frequency. It will be very hard to utilize all 320 stream running low MHz frequency.
Care to explain why it's harder?

[Shtal]

Care to explain why it's harder?

If you have a very big pipe but the water flow is low or not enough in order to push the water flow to the max, creating pressure that pores the water out in/at the great speed that it will wash you away if you stand in its way.

Big pipe = 320 stream processor units
Water = MHz

[Razor1]

hmm doesn't really work that way :wink:

[Anon Lamer]

That's an old, CPU-limited benchmark suite.
Hardly relevant to DX10-era GPU's.

But 3dmark03 is known to be very little influenced by CPU.

[BRiT]

If you have a very big pipe but the water flow is low or not enough in order to push the water flow to the max, creating pressure that pores the water out in/at the great speed that it will wash you away if you stand in its way.

Big pipe = 320 stream processor units
Water = MHz

But you're still moving the same volumn of water!

Do not think of the speed of the water, think in terms of the volumn.

[Shtal]

hmm doesn't really work that way :wink:

Different comparison: but: comparison R600 = 7 lanes on freeway traffic driving at 50MPH Vs. G80 3 lanes traffic on freeway driving at 120MPH.

[Razor1]

work done is what you want to see at the end at a given frequency at a given time.

[Shtal]

But you're still moving the same volumn of water!

Do not think of the speed of the water, think in terms of the volumn.

But the original question was to utilize 100% 320 stream processors units, which will never happened!

Just like R580 w/ 48 pixel shader processors.

[silent_guy]

If you have a very big pipe but the water flow is low or not enough in order to push the water flow to the max, creating pressure that pores the water out in/at the great speed that it will wash you away if you stand in its way.

Big pipe = 320 stream processor units
Water = MHz
You've broken the secret code of chip design!

[Razor1]

But the original question was to utilize 100% 320 stream processors units, which will never happened!

Just like R580 w/ 48 pixel shader processors.


Think your confusing a bottleneck with ultilization.

[Shtal]

You've broken the secret code of chip design!

Please don't make fun of me, I'm trying to find/use the best English words to my knowledge.

[Farhan]

Please don't make fun of me, I'm trying to find/use the best English words to my knowledge.
It is not your english that is amusing, trust me.
Your analogy about freeway lanes doesn't quite cut it either. What you have described is more like the difference in latency for a single instruction, and even then, you are making the big assumption that the pipeline depth (the length of the freeway) is the same for both.

[CMAN]

Which modern games are heavily texture dependent? These would probably be the best to show any defenencies due to having only 16 texture units. Hopefully, Rys will have that under control though. :smile:

In addition, is there any analysis that shows at what resolutions 16 ROPs may become a bottleneck?

[Jawed]

I would expect those numbers from a non scalar architecture
Disappointed, huh?

The bizarre thing is that in the same time it can do a vec4 MAD (vector or scalar makes no difference) it can do any of the following:

2 scalar MULs
2 vec3 MULs
1.6 Vec4 MULs
4 scalar ADDs
3.2 vec3 ADDs
2.5 vec4 ADDsBut not knowing how many texture fetches there are and other such quibbly details, I dunno if these are worth much investigation. The GPUBench guys had to keep fixing their code in order to get meaningful results...

Jawed

[CMAN]

Care to explain why it's harder?

Granted, I don't know how the "320 Stream Processors" relate to the 128 Nvidia utilizes, but I'd assume it would require more transistors to schedule and keep more parallel stream processors full (since ATI probably doesn't use scalar ALUs like Nvidia does it's difficult to make a correct comparison and there is a bunch of hand waving and assuming here :wink: ). Thus, less parallel processors could require less transistor overhead and allow that space to be utilized for other more productive uses.

I'd also assume though, that ATI engineers would do a complete analysis on such usage before starting, so I'm not sure how relevent this is in reality.

[Shtal]

I would like to say sorry to everyone on this forum for my comparison, it is well possibly my message is not accurate and correct because of my lack understanding of the GPU design. :(

[Brian118]

I would like to say sorry to everyone on this forum for my comparison, it is well possibly my message is not accurate and correct because of my lack understanding of the GPU design. :(

lol, It's ok, at least you are brave enough to try. I have no idea what some of you guys are talking about half the time.

[Pete]

"What...flavor gum is that? (http://en.wikipedia.org/wiki/Korgoth)"

Seriously, that's all I've got. R600 has thrown me for too many loops for me to keep churning away at my little clown cycle. (http://en.wikipedia.org/wiki/Homie_the_Clown) Someone wake me when B3D's analysis emerges on the other side of the NDA.

[silent_guy]

Please don't make fun of me, I'm trying to find/use the best English words to my knowledge.
As Farhan said, it's not the words but the metaphor itself: it doesn't come close to describing why an slow clocked architecture with more ALU's is harder than a fast clocked architecture with less ALU's, because the initial premise is wrong.

Fundamentally, there is no reason when one should be harder to get performance out than the other. So if they are different, it's because of implementation details: schedule efficiency, thread batch size, register files size, etc.
Since we have zero knowledge about the shader internals of R600, and contradictory performance data, it's pointless to start comparing the merits of either architecture.

[Jawed]

Hmm, looking at GPUBench results for R580 (I've deliberately linked to the results in numeric form):

http://graphics.stanford.edu/projects/gpubench/results/X1900XTX-5534/instrissue_all.jgr

ADD 44.1676
SUB 44.1837
MUL 30.1375
MAD 30.1312
EX2 29.6869
LG2 29.6879
POW 15.1881
FLR 20.5105
FRC 30.1374
RSQ 29.6881
RCP 29.6878
SIN 12.1333
COS 12.1339
SCS 28.0166
DP3 30.1374
DP4 30.1380
XPD 28.8265
CMP 30.1361

http://graphics.stanford.edu/projects/gpubench/results/X1900XTX-5534/instrissue_scalar.jgr

ADD1 42.7038
ADD4 42.8175
SUB1 42.8165
SUB4 42.8213
MUL1 29.5816
MUL4 29.5220
MAD1 53.9185
MAD4 29.5811

I'm really struggling to see how anyone can decode the architecture from the instruction issue rate. ADD rate is ~1.5x the MAD rate? Eh?

Why does R520 show scalar speed-up on all instructions:

http://graphics.stanford.edu/projects/gpubench/results/X1800XT-5340/instrissue_scalar.jgr

ADD1 16.2333
ADD4 13.0924
SUB1 14.4938
SUB4 10.1468
MUL1 14.4940
MUL4 10.1478
MAD1 16.2334
MAD4 10.1477

while R580 doesn't (only on MAD)?

Jawed

[Farhan]

Granted, I don't know how the "320 Stream Processors" relate to the 128 Nvidia utilizes, but I'd assume it would require more transistors to schedule and keep more parallel stream processors full (since ATI probably doesn't use scalar ALUs like Nvidia does it's difficult to make a correct comparison and there is a bunch of hand waving and assuming here :wink: ). Thus, less parallel processors could require less transistor overhead and allow that space to be utilized for other more productive uses.

I'd also assume though, that ATI engineers would do a complete analysis on such usage before starting, so I'm not sure how relevent this is in reality.
Pipeline depth is also very important. Deeper pipelines are harder to schedule for. G80 is about 10 stages, and needs 12 cycles to hide a dependent instruction IIRC. I hope someone measures this latency for R600 and I have been wondering how previous architectures like R5xx and G7x compare as well.

[CMAN]

Pipeline depth is also very important. Deeper pipelines are harder to schedule for. G80 is about 10 stages, and needs 12 cycles to hide a dependent instruction IIRC. I hope someone measures this latency for R600 and I have been wondering how previous architectures like R5xx and G7x compare as well.

As I said, there was a lot of hand waving and assumptions in my post. :razz:

I think it comes down to whether the deeper pipeline of the G80 (and we don't know how much deeper at the moment) or the larger amount of units are more difficult to schedule for. Unfortunately, I am not exactly sure.

Another question that has been nagging at me is how ATI is going to combat the 8800 GTS with 320 MB of RAM. While the smaller amount of video RAM can affect the card's performance in certain areas, will it still be close enough to the 2900XT to cannabilize the sales? Obviously, this assumes the 2900XT is similar in performance to the 8800 GTS cards. Does anyone know the sales ratio of 8800GTS cards with 640MB to cards with 320MB?

[Brian118]

When is that Chinese guy gonna wake up and gives us more benchmarks! F5! F5!

[Sound_Card]

hmmm, who should I side with?

A blog site and a fud reporter whose name is fud? Or Kombatant, CJ, and Gibbo? :!:

Kombatant seems to think thier is a properganda viral marketing stinch in the air. He has intercepted some Nvidia emails and seems to suspect foul play going on. Then we have Gibbo and CJ saying the exact same thing as Kombatant. Their are some stuff out there that is true, then on the other hand, their are some stuff out their that is totally off base.


Smoke screens and mirrors are everywhere, and danm this is getting intresting. Nice turn of events I must say.

[Shtal]

As Farhan said, it's not the words but the metaphor itself: it doesn't come close to describing why an slow clocked architecture with more ALU's is harder than a fast clocked architecture with less ALU's, because the initial premise is wrong.

Fundamentally, there is no reason when one should be harder to get performance out than the other. So if they are different, it's because of implementation details: schedule efficiency, thread batch size, register files size, etc.
Since we have zero knowledge about the shader internals of R600, and contradictory performance data, it's pointless to start comparing the merits of either architecture.

In your opinion what lacks in R600 for - outperforming 8800GTX.

Their are few advantages over G80 R600XTX has.

A. Memory Bandwidth
B. Stream processor units
C. Clock frequency of the GPU
D. 80nm tech "More space"
E: Amount of memory, plus including GDDR4 type memory

But looks to me these advantages does not help R600 right now.
So what lacks in R600 to win significantly over G8800GTX?

http://www.dailytech.com/article.aspx?newsid=7052

Thanks :)

[silent_guy]

Pipeline depth is also very important. Deeper pipelines are harder to schedule for.
A 10 stage pipeline doesn't have to mean a 10 cycle penalty in case of a data hazard if your machine supports forwarding (as is the case for all CPU's.) Maybe that's too much complexity for a GPU? And not strictly necessary because the threads...

(BTW, Any pointers about where those 10 and 12 numbers come from?)

[Geeforcer]

Kombatant seems to think thier is a properganda viral marketing stinch in the air. He has intercepted some Nvidia emails and seems to suspect foul play going on.

Just make sure you don't put words in anyone's mouth.

[silent_guy]

In your opinion what lacks in R600 for - outperforming 8800GTX.
The whole idea of my post was to point out that there's currently not enough data to start discussing the relative advantages of one over the other.

So what lacks in R600 to win significantly over G8800GTX?
42. Without a doubt.

[Sound_Card]

Just make sure you don't put words in anyone's mouth.

Who else "may" be running a fud campaign?

it smells like a FUD campaign to me

[Shtal]

42. Without a doubt.

?? example / describe please.... Thanks :)

[mao5]

Ask him what driver he used too.

Tester told me he used 8.361-070401a-046675C-ATI
ATI Technologies Inc.

btw, "070401a-046675C" will leak the tester id? I think use xxxx to relace it.

[3dcgi]

?? example / describe please.... Thanks :)
silent_guy was being sarcastic as the previous sentence said there's wasn't enough information to make a comparision.

[Farhan]

A 10 stage pipeline doesn't have to mean a 10 cycle penalty in case of a data hazard if your machine supports forwarding (as is the case for all CPU's.) Maybe that's too much complexity for a GPU? And not strictly necessary because the threads...

(BTW, Any pointers about where those 10 and 12 numbers come from?)

10 stages is the ALU itself, got this from reading the ALU patent, this was a LONG time ago, before the G80 was even launched. And 12 cycles is based on some calculations from CUDA documentation, which adds up nicely. Couple of cycles longer for scheduling or whatever.

My thoughts about forwarding:
1. It would only make sense where you have a lot of dependencies and not enough independent instructions to hide them. Don't think a GPU would have much trouble with that...
2. GPUs have hundreds of ALUs, so adding forwarding logic to each pipeline and multiply that a hundred times... and your scheduling also becomes a lot more complex... plus given (1), the benefits are probably not that great.
3. Also, the problem is that the ALUs on GPUs are deeply pipelined compared to CPUs, and you can't forward from the ALU until the result is complete, so even if it saves you a couple of cycles (forwarding only saves you from writing back to the register file and reading from it again), it's probably not that big a difference compared to a CPU where the ALUs are only like 2 cycle FP ADD and 4 cycle FP MUL, so saving a couple of cycles via forwarding cuts down latency a lot.

[trinibwoy]

Kombatant seems to think thier is a properganda viral marketing stinch in the air. He has intercepted some Nvidia emails and seems to suspect foul play going on.

Lol are you kidding me? Kombatant has intercepted Nvidia emails? I think you've watched one too many spy movies :lol:

That's not what he said - he said he saw some emails and he has a feeling there's something amiss. He seems to be a good guy but he has as much credibility as any ATI employee at this point. CJ is saying to ignore the DT numbers which is fair, but can the conspiracy bullshit until we have proof of something shady going on ok?

If Nvidia is stupid enough to be involved in any R600 misinformation campaign it won't do them any good. The official reviews will be here soon and there's nothing they can do about it.

[mao5]

http://bp2.blogger.com/_BabjUDZIqPw/RjIc5GGObEI/AAAAAAAAACM/NsWoRGaAG-0/s1600-h/013.jpg


16 texture units = Grown-Man trying to ride /accelerate using pedals on baby bicycle :lol:
Oh! come on....

-Or-
16 texture units = You have a huge engine but trying to fit in a small car. Good Luck with that :)

I hope ATI/AMD do realize that in order to boost FPS you need 32+ texture units and 20 or 24+ ROP's "Just hope that 65nm R650 will be a big surprise for me"

320 stream processor units running on lower MHz Frequency, I rather have 128 stream processor units but running at high Frequency. It will be very hard to utilize all 320 stream running low MHz frequency.
But I still like ATI. :)

16 texture units not = 16 texture units

[Farhan]

16 texture units not = 16 texture units

Care to explain what that means? :wink:

[silent_guy]

silent_guy was being sarcastic as the previous sentence said there's wasn't enough information to make a comparision.
The opportunity was too good to let it pass: the question, the lack of info, and the .sig of Shtal, all leading into "the Answer to Life, the Universe, and Everything (http://en.wikipedia.org/wiki/The_Answer_to_Life,_the_Universe,_and_Everything)". :wink:

[trinibwoy]

Care to explain what that means? :wink:

It could mean that ATI isn't that dumb after all?

[Sound_Card]

I see a different way, but hey, time will tell wont it?

http://rage3d.com/board/showthread.php?p=1334915537#post1334915537

[mao5]

When is that Chinese guy gonna wake up and gives us more benchmarks! F5! F5!

I'am waking up,and prepare my brunch

[mao5]

Care to explain what that means? :wink:

you guys know why there is L1/L2/L3 Cahce in CPU?

[Farhan]

you guys know why there is L1/L2/L3 Cahce in CPU?

Yes, but i don't see how that is relevant to the number of texture units... G80 has L1/L2 caches too. So what are you trying to say?

[trinibwoy]

I see a different way, but hey, time will tell wont it?

Yeah I would love to see those emails too and draw my own conclusions. But, as I'm sure you know, jumping to conclusions and screaming bloody murder in public without presenting the underlying evidence only breeds paranoia and spreads its own brand of FUD.

[mao5]

Yes, but i don't see how that is relevant to the number of texture units... G80 has L1/L2 caches too. So what are you trying to say?

I can not say more, but, maybe my describing is not very appropriate

[trinibwoy]

I can not say more, but, maybe my describing is not very appropriate

Yeah can you come up with another riddle? I didn't get the cache connection :)

[mao5]

Yeah can you come up with another riddle? I didn't get the cache connection :)

a car with 16cylinders, anothe car with 16 cylinders and 1xx overdrive facilities, which one is faster?

[Farhan]

a car with 16cylinders, anothe car with 16 cylinders and 1xx overdrive facilities, which one is faster?

Depends on whether you're equating the 1xx overdrive facilities with cache?

[Russell]

a car with 16cylinders, anothe car with 16 cylinders and 1xx overdrive facilities, which one is faster?

So, as was mentioned is necessary for R600 to dream of competing, its 16 "cylinders" are in some way beefed up from those of the R580. If so, at least that verifies that ATI's engineers aren't stupid. Thank god for that.

[mao5]

Depends on whether you're equating the 1xx overdrive facilities with cache?

no, they are 1xx small men rolling inside the engine, have you seen the helix ad?

[Sound_Card]

Yeah I would love to see those emails too and draw my own conclusions. But, as I'm sure you know, jumping to conclusions and screaming bloody murder in public without presenting the underlying evidence only breeds paranoia and spreads its own brand of FUD.

"may" and "seems" are the key words here. So no conclusions are drawn up. This is a roumor mill.

[Farhan]

no, they are 1xx small man rolling inside the engine, have you see the helix ad?

Are those 1xx small men involved in filtering the oil? ;)

[Geeforcer]

a car with 16cylinders, anothe car with 16 cylinders and 1xx overdrive facilities, which one is faster?

C, the one with 32 cylinders?

[Shtal]

you guys know why there is L1/L2/L3 Cahce in CPU?

You could have faster L1/L2 Cache latacy memory speed, but you cannot have faster texture units, 16 is still 16.... "No hard feelings :) :)"

[mao5]

Are those 1xx small men involved in filtering the oil? ;)

can not say more, sorry

[Twinkie]

Maybe he means that ATi has 16 "super" TMUs that can do trilinear, anisotropic, and FP16 filtering as if there was 32 TMUs something like what the G80 does.

[mao5]

keep on disclosering:

If you want to oc XT with 8.361 drv, your psu must have 6pin and 8pin line, or a 6pin-8pin converter line, otherwise, you can not oc 1MHz with 8.361 drv.

[Unknown Soldier]

what about the 6pin-6pin configuration? You've only mentioned 6pin-8pin?

US

[mao5]

what about the 6pin-6pin configuration? You've only mentioned 6pin-8pin?

US

6-6 works on the drv below 8.361, but 6-6 lost its oc function on 8.361. I think it will also lost oc function above 8.361

[mao5]

freshly heard rc drv is out, default XT totally smash default GTX

[Cuthalu]

keep on disclosering:

If you want to oc XT with 8.361 drv, your psu must have 6pin and 8pin line, or a 6pin-8pin converter line, otherwise, you can not oc 1MHz with 8.361 drv.

6pin->8pin -converter? Isn't that supposed to be disallowed by standards? Or do you mean from regular 4pin to 8pin -converter?

[memberSince97]

freshly heard rc drv is out, default XT totally smash default GTX

What? I cant hear that well...

[epicstruggle]

freshly heard rc drv is out, default XT totally smash default GTX
/faint

My heart cant take all this up and down. :razz:

[trinibwoy]

My heart cant take all this up and down. :razz:

Lol, tell me about it. It's like R600 is evolving day by day. Maybe it's made out of DNA :shock:

[BRiT]

Maybe he means that ATi has 16 "super" TMUs that can do trilinear, anisotropic, and FP16 filtering as if there was 32 TMUs something like what the G80 does.

No, no. They are 16 Extreme TMUs.






:runaway:

[mao5]

http://forums.amd.com/forum/messageview.cfm?catid=12&threadid=68294&STARTPAGE=19&FTVAR_FORUMVIEWTMP=Linear

[Bjorn]

freshly heard rc drv is out, default XT totally smash default GTX

If that's true (talking about overall performance here, not 3D Mark only) then i would like to know the reason for the 400$ pricetag.

I also want to know where i can order it if i want it next week :-)

[Farhan]

If that's true (talking about overall performance here, not 3D Mark only) then i would like to know the reason for the 400$ pricetag.

I also want to know where i can order it if i want it next week :-)

Well if the performance we've seen so far is a lie, perhaps the $400 price tag is a lie too? :-P

[mao5]

If that's true (talking about overall performance here, not 3D Mark only) then i would like to know the reason for the 400$ pricetag.

I also want to know where i can order it if i want it next week :-)

I sure you can buy one right now, but....

[trinibwoy]

Yep, IMO everything we (think) we know about R600 says it should be faster than G80 except the pricetag.

[Skrying]

Well if the performance we've seen so far is a lie, perhaps the $400 price tag is a lie too? :-P

According to my sources the price is actually $401.

[mao5]

Well if the performance we've seen so far is a lie, perhaps the $400 price tag is a lie too? :-P

yes, it's 399$, and u should subtract several games price in the XT retail box

[mao5]

According to my sources the price is actually $401.

2 bucks less.

[Geeforcer]

http://forums.amd.com/forum/messageview.cfm?catid=12&threadid=68294&STARTPAGE=19&FTVAR_FORUMVIEWTMP=Linear

Don't believe everything you read on the forums.

[Unknown Soldier]

According to my sources the price is actually $401.

lol

[mao5]

I don't think ATI will make money on XT, XT financial loose wil be compensated by AMD

[ChrisRay]

Don't believe everything you read on the forums.

Call me slow here. But what "exactly" was being linked too?

[Bjorn]

I don't think ATI will make money on XT, XT financial loose wil be compensated by AMD

Compensated how, with money from the CPU side ? I'm guessing that your joking if that's what you mean.

Sorry, AMD needs to make money and the R600 is released so that they can make money from it. Releasing the XT at a 400$ pricetag seems like a bad decision businesswise if it performs like the GTX. Unless there's some caveats with it. Lets say that it only performs like the GTX in DX 10 games, or in some but not all games.

[Skrying]

Compensated how, with money from the CPU side ? I'm guessing that your joking if that's what you mean.

Besides the recent confirmation of a price point of $401. I was also told of a rc drv that makes the R600 much improved in performance. The only issue is that the rc drv currently only works on rc drv AMD certified platforms. These rc drv AMD certified platforms are going to be extremely regulated containing the following:

1x AMD Semporno X3 29000+
2x AMD Semporno-mate X3i motherboard
1.5x 2GB AMD Semporno-rapid RAM
5x AMD Semporno-hard drive in 1TB, 2TB, and 1MB varients

You heard it here first.

[gunblade]

That is really surprising. AMD could not afford to lose any money now. They are burning their cash right now and if they have to lose money in order to sell R600, you can surely bet that there won't R700 anymore.

[mao5]

Compensated how, with money from the CPU side ? I'm guessing that your joking if that's what you mean.

Sorry, AMD needs to make money and the R600 is released so that they can make money from it. Releasing the XT at a 400$ pricetag seems like a bad decision businesswise if it performs like the GTX. Unless there's some caveats with it. Lets say that it only performs like the GTX in DX 10 games, or in some but not all games.

sorry, the truth is AMD can not make a cent on XT.

[Geeforcer]

Call me slow here. But what "exactly" was being linked too?

I can only presume this:


Before I go to a "special place" and also getting a new system built, so I'll be offline until after work tomorrow.. unless someone builds it for me in due course....

Don't believe those figures, just yet, even if you be intellectually inspired. There are 4 guys I know, 2 independent and 2 AMD workers, who have clearly told me OPPOSITE news to that just this week and they own a HD 2900 XT, with 2 having the XTX. They're not immature kids or online geeks even, they work for corporations and are professionals.

I've just talked with an AMD employee, who was also my dad's friend. (God bless him)

He called it utter shens!

No fanboism, no sheepism, no blindness or stupidity, no kidding. He ran HIS HD 2900 XT while I was on the phone with him and he also owns a XFX 8800 GTX XXX at 655/2100 watercooled.

He said he can't say anything other than it's very impressive, quiet silent in a rig and nothing you would be disappointed with.

Hoping to not give away too much, he also said those figures are BADLY low compared to the actual card he has with RC drivers that will be final.

I didn't ask him what could be wrong with that bench, he just said it's complete BS and that the final release card murders the G80 GTX. http://forums.amd.com/forum/i/expressions/face-icon-small-smile.gif


I love the whole "Didn't say anything other then...." followed by the highlighted part... and the fact that the person has been relying on Fudzilla/INQ for his information up to this point, so this tsunami of insider access must be a very recent development.

Anyway, I will make this bet to anyone willing to take it:

Once Hexus publishes their review, anyone who accepts the bet will get $5 (USD) from me for every test XT beats GTX in (both clocked at default). Conversely, the bettor will owe me $5 for every test that GTX wins. The bets are to be settled within a week after review's publication, via PayPal. And no, Rys can't particiapte :wink:

[Farhan]

sorry, the truth is AMD can not make a cent on XT.

If it really smacks the GTX like you say, then why can't they price it accordingly?

[epicstruggle]

Besides the recent confirmation of a price point of $401. I was also told of a rc drv that makes the R600 much improved in performance. The only issue is that the rc drv currently only works on rc drv AMD certified platforms. These rc drv AMD certified platforms are going to be extremely regulated containing the following:

1x AMD Semporno X3 29000+
2x AMD Semporno-mate X3i motherboard
1.5x 2GB AMD Semporno-rapid RAM
5x AMD Semporno-hard drive in 1TB, 2TB, and 1MB varients

You heard it here first.
Wow, you werent kidding about it being extremely regulated.

[mao5]

I can only presume this:



I love the whole "Didn't say anything other then...." followed by the highlighted part... and the fact that the person has been relying on Fudzilla/INQ for his information up to this point, so this tsunami of insider access must be a very recent development.

Anyway, I will make this bet to anyone willing to take it:

Once Hexus publishes their review, anyone who accepts the bet will get $5 (USD) from me for every test XT beats GTX in (both clocked at default). Conversely, they bettor will owe me $5 for every test that GTX wins. The bets are to be settled within a week after review's publication, via PayPal. And no, Rys can't particiapte :wink:

Why Hexus?

[Geeforcer]

Because I have high confidence in the quality of their reviews.

Why, you have alternative suggestions?

EDIT: LOL, I completely forgot that B3D also does the comparative testing in their reviews now :) That's also acceptible, provied that artilce doesn't get delayued by 4 months.

[clement]

Can it be a preview? :D

'cause if they do HD-dvd/blu-ray tests you might have some problems...;)

[mao5]

Because I have high confidence in the quality of their reviews.

Why, you have alternative suggestions?

EDIT: LOL, I completely forgot that B3D also does the comparative testing in their reviews now :) That's also acceptible, provied that artilce doesn't get delayued by 4 months.

come on, we should get a list of review sites. and vote for the best bet review site.

[Geeforcer]

Can it be a preview? :D

'cause if they do HD-dvd/blu-ray tests you might have some problems...;)

LOL, we are talking about tests related to 3D rendering.

[mao5]

Can it be a preview? :D

'cause if they do HD-dvd/blu-ray tests you might have some problems...;)

two more, dx10 gs test and MSAA+MRT test, seems I win 10 bucks or 70 chinese dollars.

[mao5]

LOL, we are talking about tests related to 3D rendering.

give some credit, you say "for every test XT beats GTX"

[Geeforcer]

come on, we should get a list of review sites. and vote for the best bet review site.

So what's wrong with Hexus or B3D?

[mao5]

So what's wrong with Hexus or B3D?

I didn't say them wrong, I think since you want sb else to participate your bet, you should give more choice, and we all person in this forum vote for one.

[Geeforcer]

give some credit, you say "for every test XT beats GTX"

:lol: I can see it now: "As the last 137 audio acceleration tests show, GTX still does not support it, unlike R600".

The terms: 3D rendering, B3D or Hexus, whichever posts their review first, $5 per test. Take it or leave it.

EDIT: LOL, I think B3D had over 50 tests in their review. Once XT "murders" GTX, someone will become fabulously wealthy.

[mao5]

:lol: I can see it now: "As the last 137 audio acceleration tests show, GTX still does not support it, unlike R600".

The terms: 3D rendering, B3D or Hexus, whichever posts their review first, $5 per test. Take it or leave it.

LOL

[mao5]

:lol: I can see it now: "As the last 137 audio acceleration tests show, GTX still does not support it, unlike R600".

The terms: 3D rendering, B3D or Hexus, whichever posts their review first, $5 per test. Take it or leave it.

EDIT: LOL, I think B3D had over 50 tests in their review. Once XT "murders" GTX, someone will become fabulously wealthy.

50x5bucks? you are kidding

[BlizzardOne]

No, no. They are 16 Extreme TMUs.






:runaway:

That's what I was gonna say!!:lol: :lol:

[Geeforcer]

50x5bucks? you are kidding
No, I 100% serious.

[mao5]

No, I 100% serious.

50x5bucks will make someone become fabulously wealthy?

[satein]

50x5bucks will make someone become fabulously wealthy?

At least, you might finally get a 8800GTX or HD2900XT for almost free... like a donated money :wink:

[neliz]

If you want to lose money, go ahead and engage in the bet ;)

[satein]

If you want to lose money, go ahead and engage in the bet ;)

No, the bet is not my taste, but I just offered him an opinion.

I hope the NDA will be lifted next week, so that we will see what exactly happened for that pass week....

[mao5]

The tester will check NVIDIA dx10 demo when guys in NYC have theire breakfast

[satein]

Credit to R300King posting @Rage3D... the diagram for HD2600 and HD2400.
http://rage3d.com/board/showpost.php?p=1334917155&postcount=51

HD2600
http://img215.imageshack.us/img215/3317/hd2600wg2.jpg

HD2400
http://img214.imageshack.us/img214/5963/hd2400xl4.jpg

I supplied it for comparison sake with that HD2900 earlier posted to Jawed and others to analyse (I know noting in order to interpret these diagrams).

[Bjorn]

Can it be a preview? :D

'cause if they do HD-dvd/blu-ray tests you might have some problems...;)

IMO, that's pretty much only interesting for HTPC's and then you'd likely want a 8500 or 8600 GT instead of a 8800. And those look pretty good afa HD-DVD/Blue-Ray goes:

http://www.anandtech.com/video/showdoc.aspx?i=2977&p=4

They do miss VC-1 decoding though which perhaps AMD will support. Dunno how important that will be though.

[fellix]

Credit to R300King posting @Rage3D... the diagram for HD2600 and HD2400.
http://rage3d.com/board/showpost.php?p=1334917155&postcount=51
Rather poor pixel output rate for both designs. :roll:

p.s.: And the HD2400 is missing a colour compression apparatus and the multi-level Z-Hierarchal buffer... ?

[Bouncing Zabaglione Bros.]

Guys, it's always, always been the case that when a new card comes out, you need to look at a slew of review scores from a number of sites. It's especially important when a card is being compared to it's competing product from another of the big manufacturers. Time and time again we've seen sites somehow give low scores, or high scores, or use the wrong drivers, or not enable AA/AF, or use particularly weird time demos, or any one of a number of other things that provide head-scratching results.

In this case, we've only seen one set of proper benchmark scores, and that's from Daily Tech. We know those numbers are strange, being similar to the R580, and we know the provenance of the hardware they have is suspect (we know it's not an officially sanctioned NDA sample).

So while this generates a lot of page hits and allows the conspiracy theorists to go bananas, it doesn't really tell us a whole lot with any kind of certainty until we see what everyone else gets in their reviews. Right now we are trying to look at the whole product through this little keyhole of one set of benchmarks, rumours and unsubstantiated guesswork.

Take a deep breath, let it out slowly, and wait for the range of reviews to come in that we need to have in order to assemble a complete picture of R600.

[seahawk]

It is very possible, yet even likely, that the 2900XT will beat the GTX in most cases when using the rc driver at feault settings and using a GTX at default driver settings

[Twinkie]

That's what I was gonna say!!:lol: :lol:

Heh, i didnt realise you guys take the words "super" so seriously.. enough to point out that extreme is better

:lol:

:roll:

[hoom]

Well, 'Extreme' has a certain connotation around here & I'm not talking Extreme Cameraman Wally Symons 'Extreme' here :lol:

[_xxx_]

Anyway, I will make this bet to anyone willing to take it:

Once Hexus publishes their review, anyone who accepts the bet will get $5 (USD) from me for every test XT beats GTX in (both clocked at default). Conversely, the bettor will owe me $5 for every test that GTX wins. The bets are to be settled within a week after review's publication, via PayPal. And no, Rys can't particiapte :wink:

Hey, stop stealing my ideas!!! I have a bet running about the price tag of $399 already!

That aside, those "my father's friend told me" is a sure BS-detector.

Noone doubts that R600 will be as fast or faster in most cases, just not by a huge margin. As I said some half a year ago, expect to see some ~15% more best case and that's about it.

[compres]

Hey, stop stealing my ideas!!! I have a bet running about the price tag of $399 already!

That aside, those "my father's friend told me" is a sure BS-detector.

Noone doubts that R600 will be as fast or faster in most cases, just not by a huge margin. As I said some half a year ago, expect to see some ~15% more best case and that's about it.

You sound like those are accepted or proven truths, which they are not...

[_xxx_]

What truths? Those are my personal predictions and nothing else. But "my father's friend told me" is a well-proven BS-detector for sure.

[jamis]

Hey, stop stealing my ideas!!! I have a bet running about the price tag of $399 already!

I thought you lost it.:lol:

[Jawed]

3. Also, the problem is that the ALUs on GPUs are deeply pipelined compared to CPUs, and you can't forward from the ALU until the result is complete, so even if it saves you a couple of cycles (forwarding only saves you from writing back to the register file and reading from it again), it's probably not that big a difference compared to a CPU where the ALUs are only like 2 cycle FP ADD and 4 cycle FP MUL, so saving a couple of cycles via forwarding cuts down latency a lot.
http://forum.beyond3d.com/showpost.php?p=881816&postcount=261

Jawed

[Jawed]

:shock: :shock: :shock:

HD2600 has a 3:1 ALU:TEX ratio, while HD2400 has a 2:1 ratio there.

Only 4 RBEs (ROPs) in HD2600! You thought 8600GTS had a hard time!

HD2400 really has been stripped back. The lack of Hierarchical Z is intriguing. I don't remember if any prior ATI value GPUs had no H-Z. Can't help wondering if this is because it's "designed for HyperMemory, where the memory system is so slow, who cares about the speed-up H-Z brings?" Lack of compression for colour, I suppose, implies this GPU really is never gonna see the HDR light except when being brutalised by reviewers.

Jawed

[Razor1]

WTF did ATi do, are these slides real?

[satein]

WTF did ATi do, are these slides real?

Honestly, I don't know... But who care...

It's a rumour mill, isn't it? :lol:

[satein]

:shock: :shock: :shock:

HD2600 has a 3:1 ALU:TEX ratio, while HD2400 has a 2:1 ratio there.

Only 4 RBEs (ROPs) in HD2600! You thought 8600GTS had a hard time!

HD2400 really has been stripped back. The lack of Hierarchical Z is intriguing. I don't remember if any prior ATI value GPUs had no H-Z. Can't help wondering if this is because it's "designed for HyperMemory, where the memory system is so slow, who cares about the speed-up H-Z brings?" Lack of compression for colour, I suppose, implies this GPU really is never gonna see the HDR light except when being brutalised by reviewers.

Jawed

That a 3:1 ALU:TEX ratio of HD2600, it is the same as R580, isn't it? Could this be a reason for early rumour that HD2600 perform as good as X19xx card?

[Jawed]

That a 3:1 ALU:TEX ratio of HD2600, it is the same as R580, isn't it? Could this be a reason for early rumour that HD2600 perform as good as X19xx card?
Ratio tells you nothing about absolute performance. I can't imagine anyone would have made that connection, thus generating such a rumour. It's nuts.

Jawed

[Razor1]

Honestly, I don't know... But who care...

It's a rumour mill, isn't it? :lol:


True enough :), but I thought AMD was going to go full storm in the midrange

[AnarchX]

The lack of Hierarchical Z is intriguing. I don't remember if any prior ATI value GPUs had no H-Z.

AFAIK RV3xx. It had something to do with that it was a one quad-design, R4xx was also affected(X800 VE).

[Jawed]

AFAIK RV3xx. It had something to do with that it was a one quad-design, R4xx was also affected(X800 VE).
Nicely spotted.

OK, I guess we can all go back to ignoring the value parts :lol:

Jawed

[fellix]

HD2400 -- a perfect blueprint for the next killish IGP outfit. :lol:

[INKster]

Fudo has seen the light regarding the 320 "stream processors" of R600, but he still insists on the "G80 has 160 sp's" story...
Why ?
And what's a Special Function Unit, i've never heard of this. Could he be mistaking them for the missing MUL ALU's ?

[ChrisRay]

Fudo has seen the light regarding the 320 "stream processors" of R600, but he still insists on the "G80 has 160 sp's" story...
Why ?
And what's a Special Function Unit, i've never heard of this. Could he be mistaking them for the missing MUL ALU's ?



If you counted the MUL itd be 256 SP units... dont read too much into anything FUDO says. :P

[Jawed]

I've just been given this, vaguely puzzling, information on texturing in R600:

R600 has 4 texturing units (or texturing engines):

8x texture addressing units
20x FP32 samplers (single-data read/clock)
4x FP32 filtering units (1 bilinear 64 bit sample per clock or one 128 bit sample every 2 clocks)
256KB L2 cache

8 TAs? single-clock fp16 filtered texels? Aha, so these are somewhat beefier than our old TMU friends. No specific clue as to whether it also does double-speed AF.

20x FP32s per clock? That must be for vertex fetch (point sampled, unfiltered, buffer fetch). Presumably good for 1D, 2D, or 3D? Optimised for 1D like in Xenos, though?

There was talk that R600 would "unify" vertex and texture fetches. Well this seems to be it, sort of. Except that there are prolly dedicated pipelines in the texturing engines for each type.

If you look closely at the functional overview diagram for R600 you can see 16 little boxes inside each TU. These might be the fp32 samplers. Well, 16 out of the 20, anyway. This would corresponding to fetching four 4xfp32 elements, half the maximum rate for D3D10 as I understand it (eight 4xfp32 streams are defined by D3D10). Obviously D3D10 doesn't define what the rate should be, but it does say that 8 streams should be supported.

---

Shader constants aren't handled in the way I was expecting (fetched on demand as a shader executes, cached as required in a similar way to textures). There's a cache for them, parallel to the instruction cache, which to me implies that constants are baked at the time instructions are issued.

I wonder what this means for the constant waterfalling problem and how R600 handles constant indices that are dynamic?...

Perhaps constants are split into two classes: easy constants with static (or no) index that are baked at instruction issue time, and bastards that are dynamically indexed requiring a fetch, potentially cached.

Jawed

[Jawed]

And what's a Special Function Unit, i've never heard of this.
You are kidding me?!! ARGH

Things like reciprocal, sine, log are passed over to the SFU. Normally the SFU can only produce a scalar result, so if you have a vector that you want to be wholly calculated, it'll take multiple calls to the SFU, once per vector component.

SFU instructions also tend to take longer than ADD or MAD. This very much depends on how the SFU is implemented. I know it's a cop-out, but it is, sorta, rocket science.

Jawed

[lopri]

I'm keeping on disclosing:

R600XT 3DMARK05

http://www.chiphell.com/attachments/month_0704/20070428_1c8110ce17d056f3d56a4qNhGWOKitQc.jpg



8800GTS 320 (albeit highly overclocked :grin:)

http://img264.imageshack.us/img264/2517/gts320de6.png (http://imageshack.us)

Just to present the usefulness of 3DMarks.

[lopri]

3DMARK06 XT E6600 2.4g

http://www.chiphell.com/attachments/month_0704/20070429_da39f7c4233fdd9cf0c0pecq13fv8Zhx.jpg


There is no way that E6600 @2.40GHz will score 3K in CPU tests. 3K is more like 3.50~3.60GHz.

[Sound_Card]

There is no way that E6600 @2.40GHz will score 3K in CPU tests. 3K is more like 3.50~3.60GHz.

pixie dust should work.

[INKster]

You are kidding me?!! ARGH

Things like reciprocal, sine, log are passed over to the SFU. Normally the SFU can only produce a scalar result, so if you have a vector that you want to be wholly calculated, it'll take multiple calls to the SFU, once per vector component.

SFU instructions also tend to take longer than ADD or MAD. This very much depends on how the SFU is implemented. I know it's a cop-out, but it is, sorta, rocket science.

Jawed

Thanks.
So, the "magic" Fudo talks about isn't magic at all. It's there being used right now.
Looks like he is regurgitating the same rumor he's read on some message board a while ago.

[Jawed]

Hey! I've just realised, that's a lot of cache. It seems R600 has 1MB of L2 texture cache in total. Damn, if that's true, that's a monumental change from previous GPUs.

:shock: :shock:

From the patent applications both vertex and texture fetch have an L1/L2 architecture. So, it seems likely to me that both types share the same honking big L2.

This is for vertex data:

Method and apparatus for single instruction multiple data caching (http://v3.espacenet.com/textdoc?DB=EPODOC&IDX=US2005251624&F=0)

Jawed

[trinibwoy]

I've just been given this, vaguely puzzling, information on texturing in R600:

Are those unit counts and cache per-engine?

[crux2005]

yes

[BlizzardOne]

There is no way that E6600 @2.40GHz will score 3K in CPU tests. 3K is more like 3.50~3.60GHz.

Yea, i believe he corrected himself in a later post, and it's OC@d to 3.6ish..:smile:

[Jawed]

Are those unit counts and cache per-engine?
The cache is potentially woolly (white space isn't always your friend :???: ), but the counts seem pretty definite.

Jawed

[Jawed]

Fudo has seen the light regarding the 320 "stream processors" of R600, but he still insists on the "G80 has 160 sp's" story...
Why ?
OK, well, having a look it seems that he's counting the 4 SF pipelines per cluster of 16 SPs as stream processors. Being honest, it's not unreasonable. That's what AMD's doing!

R600's ALUs appear to be configured as vec4+SF. Unlike in Xenos, where the SF can, I believe, do ADD, the SF in R600 can do MAD. Which is where we get the 5 MADs per clock number from.

I dare say it's looking likely that R600's ALUs are very much like Xenos's, but with added Integer gubbins and whatnot to make for D3D10 functionality. Fudo's article sorta gives the impression that that's the end of it, R600's ALUs have absolutely no "interesting" special sauce to improve utilisation. Right now, I'm kinda resigned to that being the case and am fairly disappointed...

I want dynamic thread packing and 4-fragment dynamic branching coherency :cry:

Jawed

[trinibwoy]

The cache is potentially woolly (white space isn't always your friend :???: ), but the counts seem pretty definite.

Jawed

Yeah I'd be surprised at 256KB cache per cluster. G80 has what, 128KB of L2 for the whole chip?

In terms of filtering - that 4xFP32 number - does that mean 4 FP32 bilerps per engine per cycle?

[Jawed]

Yeah I'd be surprised at 256KB cache per cluster. G80 has what, 128KB of L2 for the whole chip?
128KB doesn't divide by 6 (memory channels) very well, so I suspect it's some other quantity. I don't remember seeing an official specification for this. But I imagine it's somewhere in the 10s to 100s, simply because this cache covers all purposes, TMUs + vertex fetch + ROPs - though I am guessing, because I don't remember seeing anything official.

In terms of filtering - that 4xFP32 number - does that mean 4 FP32 bilerps per engine per cycle?
It says every 2 clocks.

If you look at G7x, it can do Int8 and FP16 filtering. But it can't do fp32 filtering. It can only fetch fp32 texels. So, historically, TMUs aren't necessarily fully orthogonal.

This seems to be stating that R600's TUs (like G80's) are fully orthogonal. The rates are the same for both, single-cycle fp16 or 2-cycle fp32.

It's kinda puzzling because of the presumption that fp32 filtering wouldn't be a feature of R600 (mostly my presumption, to be fair). This is the one option in D3D10 - but I'm having real doubts about that now LOL.

Jawed

[trinibwoy]

R600's ALUs appear to be configured as vec4+SF. Unlike in Xenos, where the SF can, I believe, do ADD, the SF in R600 can do MAD. Which is where we get the 5 MADs per clock number from.

Well at least we won't have to suffer the case of the missing MAD with R600! Anybody find G80's MUL yet? According to the B3D article it could be inline with the SF unit as well but that doesn't seem to be holding back R600 from exposing it.

[Rys]

Well at least we won't have to suffer the case of the missing MAD with R600! Anybody find G80's MUL yet? According to the B3D article it could be inline with the SF unit as well but that doesn't seem to be holding back R600 from exposing it.
They do SF a bit differently from each other. And you can use the MUL in G8x for general shading, just with certain drivers, seemingly certain hardware, the wind blowing in the right direction and an instruction stream written by the bastard offspring of Humus and Jawed.

[Kaotik]

Many of these have been probably said here already, but what the heck, what some people have said about the benches & R600 performance, cut/paste job from another forum


You want some truth?

People who have the cards, affiliated, and replying on those Daily Tech reports:


Kombatant: www.kombatant.com AMD/ATI employee, (ex) Beyond3D/Rage3D mod:

On a serious note, I still see some stuff out there that have no merit. Patience.

They christened the OEM version we've known and loved for quite a few months as an "XTX". That should tell you a lot about their credibility actually.

Ok, so before this goes totally out of hand, let me say this, and this will be my final say on the matter, until the NDA is lifted: AMD made certain decisions concerning this card. I took a hard look out there, to see what's being leaked, and it seems there are still some stuff that are totally made up - tbh, it smells like a FUD campaign to me, if I take into consideration certain emails that are flying around lately. Certainly there are some stuff out there that are true, and you will know which is which when the NDA lifts soonish. The journalists that were in Tunis certainly know, and are probably laughing at some of those at this minute.

So, to recap: What I've said in the past stands about the card (Sound_Card been on a mission to put all of my quotes in his sig so that we don't miss anything ). Unfortunately I can't reveal more at this point due to the NDA. And for those who are wondering, no, I am not a moderator/staff on Rage3D anymore, I stepped down two months ago, because it wouldn't be ethical imo with the new job I now have to continue do work here.

As I said, some info out there are accurate, some are not. Whatever rumours are out there certainly won't force AMD to reveal stuff sooner, that simply doesn't work, whether you're called Intel, AMD, nVIDIA or whatever.


Bum_JCRules @ THG (under NDA with cards):

Total Crap.. well almost:
While I am required to follow the NDA, the stuff up on Daily Tech today is almost worthless. Yes Anandtech was present in Tunisia (signing Non-disclosure agreements like the Inquirer), why they are posting this stuff is beyond me because their numbers are off. They must be only using the XP drivers and OS because the numbers in CF vs the GTX are very much different. So until I can officially comment on the architecture and the performance.. hold all of this as useless until the rest of the world writes about it.

I really would love to comment on this stuff...

I understand that DT and Anand are seperate but that is so childish. Derek was there and his cards got to his place of business before he returned home from Tunisia. That long board they have ... Not what Derek should have gotten in his delivery. That is all I will say before I go too far.


Kink (under NDA with cards):

Dailytechs benchmarks are inaccurate. Atleast in terms of 3dmark 06 (commenting on HD 2900 XT)


Metro.cl (under NDA with cards): www.chilehardware.com

laughable (whistling emoticon (wasn't me))


BenchZowner (under NDA and has card):

1) These benchies from DailyTech are quite off the reality, that's all I can say.

2)The 2900XTX & the 8800GTX are performing on par in Crysis at the moment ( direct info from a developer of the game )

3) Is it a deja-vu ? Remember the Mr Sanders case ? The...non invited to the press event at Ibiza editor of Hardware Analysis who wanted to "punish" ATi by publishing his...numbers of desire for the R580 ? He he, Deja-Vu

http://i12.tinypic.com/2py29hz.png

"Do I really have to express myself here ?
If these guys are so called my colleagues ( as of being reviewers like me ) then I should feel ashamed, really Evil or Very Mad

What are we looking at here ?

a) They used a better test system today, better drivers ( supposed to be ) and they managed to get the 2900XT to perform worse than their previews bench session with a worse test system & worse drivers ?
How come ?
On April 24 they got 84FPS on F.E.A.R. with the 2900XT with a QX6700 and pre-release drivers, and they got 79FPS today with a QX6800 and retail drivers ? Oh really ? Very Happy

b) Where's the result for the 2900XT in Company Of Heroes today ? Why N/A ?

c) On Half-Life 2:E1 today they got the 2900XT to score 1FPS more than the 2900XTX.
What could've caused this ? A typo ? Quite angelic.
Something else ? Using a CPU limited resolution which would cause both cards to behave like they're the same.And then there's the GTX surpassing the R600s by ~40 FPS. Quite the real leap over the Radeon X1950XTX at that game. Evil ? heh

d) Now, the best part...they scored ~48FPS in Oblivion on the 24th, and now they present us a 54FPS gain by a move from the QX6700 to the QX6800 and the small gain from running the RAM at 1T Command Rate ?

e) A reviewer in order to conduct comparable and a valuable & trustworthy review must use the same testbed, quite they opposite is what they did ( if they really went through a performance testing process )

f) And for what reason would somebody present unclear results in combination with unclear drivers & filters ( AF & AA ) settings ?
***** right boy Very Happy

My two cents ( oh wait, I have another one ) [ I'll save it for later ]

P.S. The stock core clock for the 2900XTX as of current is 800MHz and not 745MHz as they state.

P.S.2. That's pretty much all I can say at the moment.

P.S.3. Now I have to finish a memory roundup and then pack my stuff for a trip.


What did Crysis lead Developer (has the cards) say about the two (G80GTX and R600XT)?

And remember, Crysis is a DX9 game with an additional DirectX 10 codepath which will only help in getting better performance (gains) for the ones with DX10 cards.

28th April:
I'm gonna spit out a very last detail ( before they get me behind the bars for disclosing information )

The 2900XTX is 4fps in front of the 8800GTX ( in average ) in Crysis ( information directly from a developer of the game )

The 2900XTX won't be the G80 killer that everybody was expecting, but it'll be in front in most cases, with little to good margins depending on the game and settings.

nVIDIA did a great job with the G80 this time, and the performance leap over the previous generation is huge and only compared to the R300 launch in the past


That's STOCK clocks guys, it can OC high


BTW, from Fudzilla by "mouth": GeForce 8800 Ultra get 14000 scores in 3DMARK06 with no detail of test platform and no driver detail:
http://www.fudzilla.com/index....=view&id=678&Itemid=1

From DailyTech: Radeon HD 2900XT get 14005 scores in 3DMARK06 with full detail of test platform and driver detail:
http://www.dailytech.com/Overc...+R600/article7044.htm


AND NONE of them are with with release drivers, look at the GPU clocks very carefully aswell, look for the silicon version, look at the benchmarks used aswell and the detail in each setting (compare), look at the motherboard used too and that 8800GTX is no where to be found BUT by BFG at $950 as a watercooled card. PLUS those are EARLY pre-release sample cards, ES samples if you understand what that means and do you remember the abysmal performance of the faulty 8800GTX early ES samples with the wrong resistor values on release? What about the driver optimization that took place over months on end to get decent performance? Joe is on the JEDEC board BTW, and an AMD/ATi employee who helped produce GDDR4 for X1950XTX and made it a killer product so now you are telling me after 2 years spent on it like nVidia had spent 4 years on the G80, and millions of dollars, it simply doesn't beat the X1950 in benchmarks and no higher clock on the GDDR4? BTW, those benchmarks are not remotely correct either. We have 100's of GTX split around in our corporation and my own work system has 2 in SLi at 650/2100 with watercooling. Seriously, some of them benchmarks for let's say Oblivion, are 100% increased by h*ll knows what!

Hint #200: Memory clocks can go much higher and bandwidth (i.e. performance too) is much greater in upcoming titles and higher res/detail.

BTW the owner of Fudzilla, Fuad, is the ex-Graphics Editor of The Inquirer if you didn't already know. Take everything with a brick of salt.


I'll leave you with a little pre-release bang

[Kanyamagufa]

Nice sum-up Kaotik, breaks down all the leaks quite nicely. :wink:

[Sound_Card]

Many of these have been probably said here already, but what the heck, what some people have said about the benches & R600 performance, cut/paste job from another forum


Don't forget Gibbo's and CJ's comments as well.

Anyways, nice sum up and that put things a better light for sure. I have all of Komb quotes lol.

[Jawed]

They do SF a bit differently from each other. And you can use the MUL in G8x for general shading, just with certain drivers, seemingly certain hardware, the wind blowing in the right direction and an instruction stream written by the bastard offspring of Humus and Jawed.
I'd love to see that instruction stream...

Jawed

[trinibwoy]

Lol, hadn't seen these before. I guess DT isn't very popular with the guys under NDA huh? :smile:

3) Is it a deja-vu ? Remember the Mr Sanders case ? The...non invited to the press event at Ibiza editor of Hardware Analysis who wanted to "punish" ATi by publishing his...numbers of desire for the R580 ? He he, Deja-Vu

"Do I really have to express myself here ?
If these guys are so called my colleagues ( as of being reviewers like me ) then I should feel ashamed, really Evil or Very Mad

[AnarchX]

FEAR 1.08
1280*1024 4AA 8AF soft shadow OFF

HD 2900XT

http://img126.imageshack.us/img126/2613/fearhd2900xtkb7.jpg (http://imageshack.us)

source
(http://www.chiphell.com/viewthread.php?tid=3407&extra=page%3D1&page=15)

[spidy]

Sorry guys, I didn't read the whole thread. But just one small question: At this point, is the XTX gonna be released with the XT? A rumor indicates that the XTX has been pushed to the third quarter.

[tuanies]

Sorry guys, I didn't read the whole thread. But just one small question: At this point, is the XTX gonna be released with the XT? A rumor indicates that the XTX has been pushed to the third quarter.

From what I've learned, the XTX, we've come to know in its current form, 1GB GDDR4, has been canned. However, since ATI has yet to use the XTX suffix, it will probably return later on.

[tuanies]

3) Is it a deja-vu ? Remember the Mr Sanders case ? The...non invited to the press event at Ibiza editor of Hardware Analysis who wanted to "punish" ATi by publishing his...numbers of desire for the R580 ? He he, Deja-Vu

"Do I really have to express myself here ?
If these guys are so called my colleagues ( as of being reviewers like me ) then I should feel ashamed, really Evil or Very Mad

We don't attend product launch events because we never sign NDAs. We have no interest in the events.

DailyTech is not a review site. We are a news site and only that. We do not write reviews. We simply obtained the cards, ran benchmarks and presented our results. If it were a review we would've analyzed the results, drawn a conclusion and recommended a product. Instead it was a "heres what we know, here are our numbers" article.

And Ryszard hasn't threatened me with a gun yet so I'm pretty sure he still <3s me.

[trinibwoy]

FEAR 1.08
1280*1024 4AA 8AF soft shadow OFF

HD 2900XT - 84fps


That's exactly what DT got. But they claimed to have 4xAA enabled along with soft shadows.

[Razor1]

That's exactly what DT got. But they claimed to have 4xAA enabled along with soft shadows.


Well earlier in this thread there is a special procedure to turn on softshadows with AA on, so most likely DT didn't do that. I know I didn't when I ran those overclocking tests on the g80, didn't even realize the softshadows weren't working till someone pointed that out.

[Nebula]

Well earlier in this thread there is a special procedure to turn on softshadows with AA on, so most likely DT didn't do that. I know I didn't when I ran those overclocking tests on the g80, didn't even realize the softshadows weren't working till someone pointed that out.

Yes, it has to be done in the special order and everytime the application is restarted. One has to watch the benchmark to see that SS is truly on, otherwise it may be off (due incorrect order/not pressing OK twice for setting to stick).

Here is the trick if it's for use for anyone, to compare.
http://i176.photobucket.com/albums/w193/NebulasPhotoPocket/CheapExampleRetouched.jpg

[vertex_shader]

FEAR 1.08
1280*1024 4AA 8AF soft shadow OFF

HD 2900XT

http://img126.imageshack.us/img126/2613/fearhd2900xtkb7.jpg (http://imageshack.us)

source
(http://www.chiphell.com/viewthread.php?tid=3407&extra=page%3D1&page=15)

The game not have 1280x1024 resolution in the options menu.

[trinibwoy]

The game not have 1280x1024 resolution in the options menu.

What?

[Nebula]

The game not have 1280x1024 resolution in the options menu.

You can set resolution that goes above the menu options in the the config file.

Documents and Settings\All Users\Documents\Monolith Productions\FEAR\settings.cfg

[trinibwoy]

Not only that but we've had countless reviews with FEAR benchmarks at that resolution. Seems kind of silly to claim that the game isn't capable of that res at this point.

[vertex_shader]

You can set resolution that goes above the menu options in the the config file.

I know, but default its not available,the picture with the results are resized, so its grain of salt as everything else :wink:

[AnarchX]

The game not have 1280x1024 resolution in the options menu.

Yes not in game, but via config-tool:
http://www.chiphell.com/attachments/month_0704/20070429_31e394f9dc2f92d293cfXwLSBJ9Libxx.jpg
:wink:

[Kombatant]

Nice to be passed as an ex-Beyond3D mod btw... I must've worked and fought in the shade all this time :lol:

[Razor1]

Hmm FEAR at those settings with a GTX gets well over 100 fps, probably around 115 FPS actaully, just checked out FEAR extraction point, I'm getting a good 105ish

[Brian118]

Aren't those numbers kinda low for FEAR? I thought this guy said that the HD 2900xt "totally smash" the GTX?

[BrynS]

Slightly OT, but has there been any word/leaks on RD790 + Agena FX samples from Tunis and such samples being seeded to reviewers as a testing platform for R600 or at least RD790 for R6 Crossfire/(QuadFire?)/Stream Computing reviews?

[trinibwoy]

Aren't those numbers kinda low for FEAR? I thought this guy said that the HD 2900xt "totally smash" the GTX?

That was mao5 - different chinese source I think.

[Brian118]

But they are coming from the same forum. chiphell.com

[trinibwoy]

Oh right. Just realized mao is a "technical advisor" there. The smileys are funny as hell on that forum :lol:

[digitalwanderer]

We don't attend product launch events because we never sign NDAs. We have no interest in the events.

DailyTech is not a review site. We are a news site and only that. We do not write reviews. We simply obtained the cards, ran benchmarks and presented our results. If it were a review we would've analyzed the results, drawn a conclusion and recommended a product. Instead it was a "heres what we know, here are our numbers" article.

And Ryszard hasn't threatened me with a gun yet so I'm pretty sure he still <3s me.
DailyWho? :|
















;)

Me <3 DailyTech, Me <3 Tuan. :)

[nicolasb]

Okay, I have a question.

Does anyone know yet for sure if R600 is going to support the same digital-audio-movie-soundtrack-passthrough-via-HDMI feature that we're supposed to get on RV6xx? Initial rumours were that the R600 chip didn't have this feature at all. Later rumours suggested that it did. However, still-later rumours suggest that the board design for R600 cards includes only HDCP-enabled DVI outputs, not HDMI (with HDMI compatibility only via an external dongle). If that's correct then (I think) it means that R600 cards won't be able to output audio over HDMI at all. :(

Can anyone confirm or deny any of that?

[Geeforcer]

That's exactly what DT got. But they claimed to have 4xAA enabled along with soft shadows.

Well, not exactly. DT got 79 fps, but they had 16AF as oppoused to 8.

[trinibwoy]

That's from the XTX article. They got 84fps in their original XT article. And yeah they had 16xAF enabled.

[Arty]

Correct me if I'm wrong (which I usually am in most cases)

2600
120 SPUs (24 Vec5)
8 TMUs
4 ROPs

2400
40 SPUs (8 Vec5)
4 TMUs
4 ROPs

[Kaotik]

Correct me if I'm wrong (which I usually am in most cases)

2600
120 SPUs (24 Vec5)
8 TMUs
4 ROPs

2400
40 SPUs (8 Vec5)
4 TMUs
4 ROPs

I seriously hope 2600 will have 8 ROPs, how can they make the same mistake again? :???:

[Arty]

I seriously hope 2600 will have 8 ROPs, how can they make the same mistake again? :???:
There is something that we have no clue about (yet) which seems to be consuming significant space in all the parts. Looking at the size of RV630, in a simple world should be enough for 'half R600'. However if they can manage to squeeze only 120 SPUs (& 4 ROPs) then I wonder what size would end up as 'half R600' on 65nm.

P.S. I think UVD as the likely culprit was already discussed some where around here.

[Kaotik]

There is something that we have no clue about (yet) which seems to be consuming significant space in all the parts. Looking at the size of RV630, in a simple world should be enough for 'half R600'. However if they can manage to squeeze only 120 SPUs (& 4 ROPs) then I wonder what size would end up as 'half R600' on 65nm.

P.S. I think UVD as the likely culprit was already discussed some where around here.

Well UVD + Audio departments takes the same space from all 3 chips, does RV610 follow the same "it should have more" pattern?

[mao5]

I'm coming

[Jawed]

I've been musing over the organisation of R600's ALU pipeline.

The diagram shows a fat yellow unit sitting beside 4 skinnier yellow units. It's worth noting that those skinny units are disjoint, they're not a single ALU.

The implication here is that R600 is capable of issuing five separate scalar instructions per clock, per pixel (or vertex or primitive).

So, I propose that the fatboy slug is a combined special function and MAD ALU and its weedy cousins are boring old MAD/MUL/ADD ALUs.

So, it's time to roll out my super-nasty code fragment, for one last time (?) :

http://www.cupidity.f9.co.uk/b3d83.jpg

Note, that's the revised version with a DP3 at the end, not a RCP as I did originally.


http://www.cupidity.f9.co.uk/b3d95.gif


R600 doesn't get much chance to stretch its legs on this code, because the instruction dependencies are tighter than a slug's arse. But you should be able to see a bit of instruction-component jiggery-pokery going on in cyles 3 and 4.

Note I've clocked the pipeline at 800MHz on the basis that in theory there'll be such a thing (whose peak capability is 512GFLOPs)...

For comparison, R580:

http://www.cupidity.f9.co.uk/b3d86.gif

And Xenos:

http://www.cupidity.f9.co.uk/b3d88.gif

Even though the FLOPs utilisation of R600 is only marginally higher than R580 (and it's the same as Xenos), you can see that the per-component utilisation is much higher. Summarising utilisation on the basis of components:

R600 = 80%
R580 = 52.5%
Xenos = 70%
Jawed

[Geeforcer]

I'm coming

Wow, the I didn't know R600 numbers were THAT good!

:wink:

[Richteralan]

I'm coming

So you are going to ask that guy yourself,
or I can help you to ask in that forum:lol:

P.S. I can read&write both Simp. and Trad. Chinese:wink:

[mao5]

So you are going to ask that guy yourself,
or I can help you to ask in that forum:lol:

P.S. I can read&write both Simp. and Trad. Chinese:wink:

I help that guy test "TEST DRIVE UNLIMITED", so you can save my time.

[bigtabs]

:lol: I'd much rather you did another FEAR test. I was really only joking about TDU as I hate NFS so much, however TDU may be a good run for it aswell.

FEAR on my X1900XT + AMD64 @2.8 gets not far off those posted results, albeit at 10x7 stock max settings (AAx4, SS off).

Could you do a stock max settings test on FEAR with the R600? Basically just set both Computer and Graphics to Maximum and change nothing else. That would be dandy.

Thanks for your efforts btw. :smile:

[mao5]

8800GTS 320 (albeit highly overclocked :grin:)

http://img264.imageshack.us/img264/2517/gts320de6.png (http://imageshack.us)

Just to present the usefulness of 3DMarks.

What's you cpu speed and 320 speed?

[mao5]

8800GTS 320 (albeit highly overclocked :grin:)

http://img264.imageshack.us/img264/2517/gts320de6.png (http://imageshack.us)

Just to present the usefulness of 3DMarks.

you compare default XT with your extreme oc 320? you are so smart.

[mao5]

There is no way that E6600 @2.40GHz will score 3K in CPU tests. 3K is more like 3.50~3.60GHz.

Have u read my post several pages before?

[Kombatant]

Let me boot into Vista and check out how much a E6600 @ 3.61GHz gives on 3DMark06 so we can put this to rest once and for all.


And no, I won't do it with the R600, just a X1900XTX :p

[neliz]

all we have to do is add 30% to the 1900 score then...

[mao5]

That's exactly what DT got. But they claimed to have 4xAA enabled along with soft shadows.

I don't known how to enable soft shadows in the timedemo with AA on. Anyway, the XT tested FEAT timedemo with default E6600 cpu and default XT,:1280*1024 4AA 8AF soft shadow OFF. How about DT? http://69.93.88.162/forum/images/smiles/rofl.gif

[mao5]

you guys mean DT using super hardware for XT to get the same result with all default E6600 and default XT? hahaha http://69.93.88.162/forum/images/smiles/rofl.gif. I'll no more trust DT.

http://i12.tinypic.com/2py29hz.png

[Kombatant]

Old run, but should still give you an idea: http://service.futuremark.com/compare?3dm06=631266

[mao5]

That was mao5 - different chinese source I think.

not me, some yankees in AMD forum

[mao5]

Oh right. Just realized mao is a "technical advisor" there. The smileys are funny as hell on that forum :lol:

sorry, I'm misunderstanding your word, sorry.

[mao5]

Okay, I have a question.

Does anyone know yet for sure if R600 is going to support the same digital-audio-movie-soundtrack-passthrough-via-HDMI feature that we're supposed to get on RV6xx? Initial rumours were that the R600 chip didn't have this feature at all. Later rumours suggested that it did. However, still-later rumours suggest that the board design for R600 cards includes only HDCP-enabled DVI outputs, not HDMI (with HDMI compatibility only via an external dongle). If that's correct then (I think) it means that R600 cards won't be able to output audio over HDMI at all. :(

Can anyone confirm or deny any of that?

There will be a DVI-HDMI convert in retail box, R6XX DVI can send 5.1 ss out

[Geeforcer]

oh,man, I hope you were in twin tower at that time as hell.

:huh: :huh: :huh:

[dnavas]

So, I propose that the fatboy slug is a combined special function and MAD ALU and its weedy cousins are boring old MAD/MUL/ADD ALUs.

It's an interesting design difference from the G80, if so. While G80 struggles to surface the SFU's MUL, ATI spent trannies to implement an entire MAD.

Could the fat boy also be responsible for branching, or is that entirely within the dispatcher? Maybe I'm misremembering "dedicated branch processors" from somewhere....

[Arty]

Well UVD + Audio departments takes the same space from all 3 chips, does RV610 follow the same "it should have more" pattern?
Yes. Compared to the 2900:

2600
~1/3 SPUs (37.5% to be exact)
1/2 TMUs
1/4 ROPs
1/2 Die-size

2400
1/8 SPUs
1/4 TMUs
1/4 ROPs
1/4 Die-size

[mao5]

:lol: I'd much rather you did another FEAR test. I was really only joking about TDU as I hate NFS so much, however TDU may be a good run for it aswell.

FEAR on my X1900XT + AMD64 @2.8 gets not far off those posted results, albeit at 10x7 stock max settings (AAx4, SS off).

Could you do a stock max settings test on FEAR with the R600? Basically just set both Computer and Graphics to Maximum and change nothing else. That would be dandy.

Thanks for your efforts btw. :smile:

see that pic pages before, the 84 scores was done by default E6600 and default R600XT. The tester went to buy a convert power line for 8pin, but failed, so he couldn't overclock his XT using 8.361 drv.

http://www.chiphell.com/attachments/month_0704/20070429_8fdc614c5d2e607f7a81uZTo0cTHJYF3.jpg

[mao5]

I will tell that guy to test this on R600XT, all though miracle may be small.

http://www.chiphell.com/attachments/month_0704/20070430_46cc3aa980ec22a8c0d3ofrc1z0WwWOQ.jpg

[Nebula]

I will tell that guy to test this on R600XT, all though miracle may be small.

Interesting, but do you know anything about the CP (driver CP) settings. Are they default, high-quality or other setting?

And I wonder what differences there are between the Nvidia default CP (for 8800GTX) and default ATI CP (2900XT) settings.

[AnarchX]

2400
1/8 SPUs
1/4 TMUs
1/4 ROPs
1/4 Die-size

Not forget, that RV610 has no Hir-Z, color-compression and dedicated L1/L2-Texture Caches.

[mao5]

Interesting, but do you know anything about the CP (driver CP) settings. Are they default, high-quality or other setting?

And I wonder what differences there are between the Nvidia default CP (for 8800GTX) and default ATI CP (2900XT) settings.

I'll ask him when he wake up.

[fellix]

Any chance to take few screens with FSAA Viewer, rolling different AA modes?
Curious about the patterns. :smile:

[Russell]

Yes. Compared to the 2900:

2600
~1/3 SPUs (37.5% to be exact)
1/2 TMUs
1/4 ROPs
1/2 Die-size

2400
1/8 SPUs
1/4 TMUs
1/4 ROPs
1/4 Die-size


Just like with the 8600 lineup, the midrange of this generation are really getting shafted. Clearly nVidia and AMD want to shift more sales towards the high-end and are making them considerably better to do so.

For what it is, the 2400's specs are appropriate (good ratios for a HTPC, IMO, as an example) but the 2600 (if these specs are accurate, of course) strikes me as crap.

[vertex_shader]

I'll ask him when he wake up.

The guy sleep too much :lol:

Whats the point with that company of heroes picture?

I never play the game so can't figure it out, but i see in the picture some medium details.

That FEAR benchmark looks ugly, only 200% faster than with e4300 default and x1950pro default.

[Silent_Buddha]

Just like with the 8600 lineup, the midrange of this generation are really getting shafted. Clearly nVidia and AMD want to shift more sales towards the high-end and are making them considerably better to do so.

For what it is, the 2400's specs are appropriate (good ratios for a HTPC, IMO, as an example) but the 2600 (if these specs are accurate, of course) strikes me as crap.

Maybe, but are the RV6xx ROPS the same as RV5xx ROPS? If they've been beefed up somehow it might not be so bad.

As to shafting mid/low range chips to sell more high end chips, I find this doubtful. It's probably more of a case that there are VERY stringent transistor budgets for mid/low range to make sure that cost is kept as low as possible.

After that they then see what they can cram into that transistor budget. I'm sure there's also a transistor budget for the high end, but I'd imagine they have a LOT more leeway as the cards are expected (generally) to sell for quite a bit anyways.

This might just be a case (for both AMD and Nvidia) that they can't support more performance because they have to get DX10 compat and both seem to also be focusing a lot more on Video Decode. Although AMD has gone a couple steps further by supporting VC-1 as well as HDMI Audio out. Although I don't "think" the RV6xx handle the HDMI audio on chip. I would imagine it's probably another chip on the board that is handling that.

Regards,
SB

[mao5]

The guy sleep too much :lol:

Whats the point with that company of heroes picture?

I never play the game so can't figure it out, but i see in the picture some medium details.

That FEAR benchmark looks ugly, only 200% faster than with e4300 default and x1950pro default.

don't you know what time is in china right now? AM 5:05. It's time for sleeping. The pic of COH is there dx10 shader quality option, but it's grey on my X1900GT, I will ask him to test the option on R600XT.

[Russell]

Maybe, but are the RV6xx ROPS the same as RV5xx ROPS? If they've been beefed up somehow it might not be so bad.


I'm merely talking about RV6x0 vs R600 and the degree of cutdown between the two, not intergenerational performance increases. What you say about transistor budgets does make perfect sense, I guess I just always like to see more. As it stands now, I wouldn't consider buying a 2600XT if these specs are accurate (I've typically been a mid-range buyer but I see that changing).

[vertex_shader]

don't you know what time is in china right now? AM 5:05. It's time for sleeping. The pic of COH is there dx10 shader quality option, but it's grey on my X1900GT, I will ask him to test the option on R600XT.

Dx10 is the options is there, but not active with dx10 cards, dx10 patch not out yet.

[pjbliverpool]

Just like with the 8600 lineup, the midrange of this generation are really getting shafted. Clearly nVidia and AMD want to shift more sales towards the high-end and are making them considerably better to do so.

For what it is, the 2400's specs are appropriate (good ratios for a HTPC, IMO, as an example) but the 2600 (if these specs are accurate, of course) strikes me as crap.

I agree, both seem to be leaving a massive hole between the high end and midrange that hasn't existed in the past. In fact im amazed ATI is doing this aswell given the opportunity nvidia has presented to them.

Im still hoping for a 48 ALU version that can give an indication of Xenos's performance.... you know, just for kicks :grin:

[Russell]

I agree, both seem to be leaving a massive hole between the high end and midrange that hasn't existed in the past. In fact im amazed ATI is doing this aswell given the opportunity nvidia has presented to them.

A hole that admittently could be filled at a later date. But would it be? nVidia still hasn't filled the gap between the GTS and GTX, but they haven't had a reason (such as competition) to do so (yet). I kind of doubt that it'll happen though.

I think AMD is following suit simply because it is more profitable to them this way. After all, money speaks louder than happy enthusiasts.

[Ailuros]

I've just seen some leaked slides in another forum and some interesting details about that new CFAA thingy; so let me get this straight: according to that logic the good ole Quincunx was 5xAA after all? :roll:

[Kaotik]

I've just seen some leaked slides in another forum and some interesting details about that new CFAA thingy; so let me get this straight: according to that logic the good ole Quincunx was 5xAA after all? :roll:

I hope those aren't real slides or that you've understood something wrong :shock:

[w0mbat]

Here are the slides:

http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449565&postcount=351
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449613&postcount=364

[INKster]

Humm, just quoting some whispers (may have innacurate english due to auto-"googlish"-translation)...

- Percentage Closer Filtering.
- up to 67 Megatexels (8192 x 8192).
- 64bit FP HDR textures bilinear filtered at full speed; 128bit FP HDR textures filtered at half speed.
- Trilinear and Anisotropic filtering supported on all Z formats.
- RGBE 9:9:9:5 (32bit HDR shared exponent texture format).
- Improved Aniso HQ filtering.
- Something called "Full texturing capability available to both pixel, vertex and geometry shading".


Again, just quoting what i've heard. ;)


edit
Argh, someone beat me to it. :D

[tEd]

Here are the slides:

http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449565&postcount=351
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449613&postcount=364

Now to the experts , what is possible with improved resolve filter quality wise? I know reading something from SA here about it some time ago.

[Russell]

Here are the slides:

http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449565&postcount=351
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449613&postcount=364


Well, those slides certainly appear to be real. They confirm a lot of the current rumors, such as the specs of the entire launch lineup. The section about the texture units makes me feel better about the overall performance that we can expect. Correct me if I'm wrong, as my understanding is limited, but they do look beefier than R580's texture units.

[fellix]

Wow, tent filtering kernel for the AA. Nice catch, there. :lol:

[Arnold Beckenbauer]

Better image quality?
ATI Updates Vista Driver with New Features and Information (http://www.pcper.com/article.php?aid=367)
The new MSAA adaptive anti-aliasing is scheduled to be put into the May Catalyst release. This feature will bring their AA options on-par or beyond what NVIDIA currently offers on the G80 graphics cards. This feature timeline is a target, and not yet a commitment from AMD at this time.
:grin:

[epicstruggle]

Here are the slides:

http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449565&postcount=351
http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449613&postcount=364
Wow, looking at the flops chart, it looks like you could get ridiculously close to 1 teraflop.

998 Gigaflops. 2x XT + the high end cpu described in the slide.

[Ailuros]

I hope those aren't real slides or that you've understood something wrong :shock:

Well have a close look at the slides above and answer my question. I'm not debating the possible quality increase of a custom filter on top of Nx amount of MSAA samples one bit; in fact I'd say that we do need better filters (yet under presuppositions). I have a hard time digesting that the added filter samples get counted as AA sample densities.

Whether in the end the added filters deliver a tolerable grade of overfiltering for my taste or not, is something I have to decide when I see the method in real time.

My biggest objection with Quincunx and any related method is that the added filter isn't applied on just polygon edges/intersections, yet on the entire scene.

[nAo]

New 'advanced' custom AA resolve passes are nice and everything but it seems to me that nothing could stop competition to implement the same filters, in fact I'm willing to bet R6xx uses its shaders code to implement those filters.

[Jawed]

It's an interesting design difference from the G80, if so. While G80 struggles to surface the SFU's MUL, ATI spent trannies to implement an entire MAD.
Maybe their SFU is so close to being a full MAD that they just went the extra mile. It has 3 mini-MADs in it, actually:

Method and system for approximating sine and cosine functions (http://v3.espacenet.com/textdoc?DB=EPODOC&IDX=US2005071401&F=0)

the biggest is 20-bit (just the mantissa). Dunno...

Could the fat boy also be responsible for branching, or is that entirely within the dispatcher? Maybe I'm misremembering "dedicated branch processors" from somewhere....
R5xx introduced a dedicated branch pipeline. The diagram appears to show a little blue runty thing, that's prolly it.

Jawed

[Jawed]

Maybe, but are the RV6xx ROPS the same as RV5xx ROPS? If they've been beefed up somehow it might not be so bad.
They've been beefed up. We're just not sure how much yet. 4xAA per clock is looking very likely.

Jawed

[fellix]

in fact I'm willing to bet R6xx uses its shaders code to implement those filters.
Hmm, that way can we hope for Gaussian disperse kernel in future drivers? :grin:

Btw, what about that fully ring-bus topology -- how much it will differ from the current one in x1000 GPUs -- does that mean, that each ring-stop is now actually a 1/4 attached block of the whole design, a kind of on-chip distributed network?! :shock:

[R300King!]

http://www.forum-3dcenter.org/vbulletin/showthread.php?p=5449558#post5449558

What's this?

R600:
64ALUs
16TMUs

RV630:
24ALUs
8TMUs

RV610:
8ALUs
4TMUs

[Ailuros]

New 'advanced' custom AA resolve passes are nice and everything but it seems to me that nothing could stop competition to implement the same filters, in fact I'm willing to bet R6xx uses its shaders code to implement those filters.


Just about when I was happy that Quincunx was gone....heh

[nAo]

Just about when I was happy that Quincunx was gone....heh
It's going to be much better than quincunx, I experimented a few months ago wiith custom AA resolve filters and even a simple gaussian filter on a 4X multisampled image (just a bit wider than a single pixel) was given extremely good results + a slight blur..

[Jawed]

Blimey that AA is going to be fun to play with :shock: But only 2x, 4x and 8xMSAA? Hmm, I was hoping for 4x, 8x, 12x. I really want to see movies of these CFAA modes now...

I see I just about got there in time to post my speculation on the ALU operation :mrgreen:

LOL, I moaned about the asymmetry of the ring-bus, now it's symmetric. I even convinced myself it was inevitably asymmetric, hahahaha.

I've asked before, I'm gonna ask again, what use are fp32-per-component bilinearly filtered texels?

Jawed

[BRiT]

More ATI Slides, these showing performance comparison to Radeon 1950XTX, for Stencil Shadow, HDR, and Crossfire [ http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449740&postcount=414 ]

[dnavas]

New 'advanced' custom AA resolve passes are nice and everything but it seems to me that nothing could stop competition to implement the same filters, in fact I'm willing to bet R6xx uses its shaders code to implement those filters.

In theory, the R600 has slightly more math/shader power to throw at something like this, though.

R5xx introduced a dedicated branch pipeline. The diagram appears to show a little blue runty thing, that's prolly it.

Thanks, I missed that.

-Dave

[Bouncing Zabaglione Bros.]

Just about when I was happy that Quincunx was gone....heh

Hasn't there been general talk for quite a while of moving away from hardcoded AA methods applied via control panel to the whole screen, to AA applied at the discretion of the developer directly where it is needed using shaders as the mechanism?

[tEd]

It's going to be much better than quincunx, I experimented a few months ago wiith custom AA resolve filters and even a simple gaussian filter on a 4X multisampled image (just a bit wider than a single pixel) was given extremely good results + a slight blur..

Can you share the results? (Pictures) :smile:

[Rangers]

Well it's nice to see this thread got exciting again with at least rumors that R600 can beat GTX.

I still have a hard time believing it though, because why $400 if that's the case?

So glad they moved NDA up to May 2..

Someone back there mentioned supposedly Crysis running faster on R600..again that wouldn't surprise me if the TMU limits are true. Crysis is probably going to stress shaders like no game ever before by a factor of several.

[INKster]

More ATI Slides, these showing performance comparison to Radeon 1950XTX, for Stencil Shadow, HDR, and Crossfire [ http://www.forum-3dcenter.org/vbulletin/showpost.php?p=5449740&postcount=414 ]

That scale on the slides comparing it to the R580...
Shady, very shady.

[nAo]

In theory, the R600 has slightly more math/shader power to throw at something like this, though.
That's true but you need also to sample your sub samples..;)

[Bouncing Zabaglione Bros.]

That scale on the slides comparing it to the R580...
Shady, very shady.

I noticed that, but it is somewhat mitigated by the R580 baseline being clearly marked, and the percentage increase clearly up the side, as well as just the frames on the bars.

[BRiT]

That scale on the slides comparing it to the R580...
Shady, very shady.

And what would you have them compare it to? Or do you mean how they only show 1 block of "80 to 100%" and then leave 5 more blocks above that to go upto 200% ? Yeah, it's very marketting-like, but at least they provide you with raw numbers too.

[INKster]

I noticed that, but it is somewhat mitigated by the R580 baseline being clearly marked, and the percentage increase clearly up the side, as well as just the frames on the bars.

I have no problem with that.
It's just that, this way, whoever sees those graphs and doesn't know any better will see a R600 with a bar that looks as if it was several times faster than a R580, when in fact it's closer to 0.8x times faster, at best.

If the graphs started at zero with 10% intervals, the impact would be much less pronounced.
I don't like it when there's this marketing effort trying to obscure clarity. Not everyone notices these subtleties.
And frankly, AMD is starting to "usurp" ATI's clear way of doing marketing (they took over this department directly, didn't they ?).

[Ailuros]

It's going to be much better than quincunx, I experimented a few months ago wiith custom AA resolve filters and even a simple gaussian filter on a 4X multisampled image (just a bit wider than a single pixel) was given extremely good results + a slight blur..

No objection there; one objection is that filter samples get counted as AA sample densities and a close second that if we'd be talking about 16x or even 24xCFAA how "slight" will the blur be there?

[Ailuros]

Hasn't there been general talk for quite a while of moving away from hardcoded AA methods applied via control panel to the whole screen, to AA applied at the discretion of the developer directly where it is needed using shaders as the mechanism?

I don't think the custom AA filter thingy of R600 has anything to do with shader antialiasing.

[Jawed]

And what would you have them compare it to? Or do you mean how they only show 1 block of "80 to 100%" and then leave 5 more blocks above that to go upto 200% ? Yeah, it's very marketting-like, but at least they provide you with raw numbers too.
That's the problem with this HD era, 16:9 aspect ratio robs your bars of their height :eek:

Bring back 4:3 I say.

Jawed

[Bouncing Zabaglione Bros.]

I have no problem with that.
It's just that, this way, whoever looks at those graphs and doesn't know any better will see a R600 with a bar that looks as if it was several times faster than a R580, when in fact it's closer to 0.8x times faster, at best.

Reading a graph like that isn't rocket science. If you're the sort of person that is only going to see the pretty coloured bars and not look to see what the scale means (especially when it is so clearly marked), you've really no business looking at anything as complex as a bar graph! :lol:

[nAo]

No objection there; one objection is that filter samples get counted as AA sample densities and a close second that if we'd be talking about 16x or even 24xCFAA how "slight" will the blur be there?
I agree on the first objection, it's just marketing but we shouldn't get suprised by that, every company acts this way.
Dunno about the second one but I'd like to find that out as soon as possible, even though the fact that they have narrow and wide filters is already a hint for us, imho ;)

[INKster]

Reading a graph like that isn't rocket science. If you're the sort of person that is only going to see the pretty coloured bars and not look to see what the scale means (especially when it is so clearly marked), you've really no business looking at anything as complex as a bar graph! :lol:

Refer to the second part of my previous post.
Like i said, i have no problem with that, but it's still "unusual" in ATI's ways of doing marketing.
I would expect that from Nvidia or Intel, as their marketing departments have a very aggressive reputation.

[tEd]

Refer to the second part of my previous post.
Like i said, i have no problem with that, but it's still "unusual" in ATI's ways of doing marketing.
I would expect that from Nvidia or Intel, as their marketing departments have a very aggressive reputation.

It's not ati anymore.

[Rangers]

Refer to the second part of my previous post.
Like i said, i have no problem with that, but it's still "unusual" in ATI's ways of doing marketing.
I would expect that from Nvidia or Intel, as their marketing departments have a very aggressive reputation.

ATI did the same thing on R520 and people bitched then too.

It's not new even from ATI.

[Bouncing Zabaglione Bros.]

Refer to the second part of my previous post.
Like i said, i have no problem with that, but it's still "unusual" in ATI's ways of doing marketing.
I would expect that from Nvidia or Intel, as their marketing departments have a very aggressive reputation.

And it's been flagged for a long time as being an area where the former ATI was deficient compared to the marketing beast that is Nvidia. I don't think it would be a bad thing if AMD gets a bit more aggressive on the marketing front. It's not like these graphs are misleading if you actually take the trouble to read them. They are pretty mild compared to other graphs I've seen where you don't even get a scale on the side.

[INKster]

It's not ati anymore.

Semantics. ;)
R600 is still an ATI child.

[Kaotik]

Refer to the second part of my previous post.
Like i said, i have no problem with that, but it's still "unusual" in ATI's ways of doing marketing.
I would expect that from Nvidia or Intel, as their marketing departments have a very aggressive reputation.

Maybe they put old AMD's marketing team on the job? :lol: