The NEXT LAST R600 Rumours & Speculation Thread

[ERK]

Or 64 Vec4+1 ALUs with 5 MADDs and you get 475 GLOP/s (740 MHz).
R520's eight VS ALUs are "5D" - Vec4+1 and groupped in one "block" (SIMD).
The next idea:

320/17= error
320/(16*(4*Vec1 ALUs)+16*Vec1 ALUs) = 4
Crazy, isn't it?

you're referring to this? (http://ati.amd.com/developer/brighton/05%20Graphics%20Performance.pdf)

[nutball]

i wouldn't consider it a paper launch if they lift the NDA now and say they will be available on XXth of XXXXXX. as long as it's available when they say, what's the problem.. if they're nowhere to be seen, fine call it a paper launch.

as long as they stay true to their word though, not IMO a paper launch.

Personally I think that the whole paper/hard/soft/firm/hard-yet-late/soft-yet-hello launch issue is too wrapped up in "vendor preference enthusiasm" nonsense for it even to be a topic of discussion. The board partners and big OEMs have a case for caring, everybody else is a fan-boy.

[Pete]

So, on the one hand we've got DT showing the XT to be on par with the GTS. OTOH, we've got The Inq apparently (http://www.theinquirer.net/default.aspx?article=39176) leaking info from the Tunis press day saying (http://www.theinquirer.net/default.aspx?article=39145)

"But performance is far beyond we have imagined, and hearing that this is not limited to a certain version of DX10 only speaks of better days for PC gaming."
"AMD promised extremely competitive pricing, and in many cases the HD2900XT will beat the 8800GTX, yet alone the 8800GTS."

and finally (http://www.theinquirer.net/default.aspx?article=39176)

"Richard Huddy showed benchmarks to loads of hacks who met them with indifference. It was almost as if we heard Richard Huddy scream at the hacks: "GoDAMMIT, this is exciting stuff, clap, damn you.""

Now, that last quote could indicate less-than-inspiring performance, but I have no idea how journos react at these events (soberly or drunkenly--er, enthusiastically :razz: ). What seems clear is that Inq was there (in force) and they basically repeated (in inimitable Inq fashion) what they were presented with, so it seems like we're getting the straight dope.

(Still, you'd have to wonder why DT would present #s they know wouldn't be representative and so wouldn't hold up. But would they know? It seems they weren't at Tunis, given the comment that AT's review boards would've got back to NC before AT's staff, implying that DT has the keys and was there, waiting. But did ATI hand out review samples in Tunis, or did they ship them to each reviewer individually? I thought that in previous events samples were handed out. Either way, I'm inclined to trust that DT's #s are real, but still seeking clues (http://forum.beyond3d.com/showpost.php?p=976125&postcount=2822) as to their apparent weirdness.)

Then again, maybe the clues are in the quotes. "Competitive pricing" and "[let] alone the 8800GTS" speaks to the XT targeting the GTS in both price and performance. "Hearing that" and "promised" points to The Inq not knowing actual performance (or that they're skeptical of ATI-supplied #s). "Indifference" points to the journos seeing nothing exciting (or, again, jaded skepticism). Maybe those 16 texture units and ROPs really do explain the performance in the games DT benchmarked.

Custom Filter AA. Is that "Temporal AA" redux, the next step toward pseudo-stochastic bliss? 24xAA lines up well with 8xAA. Maybe I've guessed enough for one post.

Kaotik, I'm with BlizzardOne. It'd seem to be more accurate to call it a soft launch, given NV's recent (good) track record of launches and availability.

[INKster]

Don't forget Kyle's comments over at [H] forums a few days ago.



So, we basically have three different camps right now:

- The "believers", who still think ATI has some sort of smoke screen in place (a crippled driver, on purpose) ready to beat the 8800's, and back those theories with certain discrepancies in known tests.

- The "nay-sayers", who have either higher-than-average access to hardware and/or software vendors and IHV's, or actually have hardware on their hands and are doing some tests on them (DT, Kyle, etc).

- The little known value-minded semi-enthusiast who would like to see a price-war (à la AMD vs Intel) in order to have early access to high-grade GPU hardware on the cheap (ok, that's me :D).


Who's right ?

[jamis]

So, we basically have three different camps right now:

- The "believers", ...
- The "nay-sayers", ...
- The little known value-minded semi-enthusiast who would like to see a price-war (Ã* la AMD vs Intel) in order to have early access to high-grade GPU hardware on the cheap (ok, that's me :D).


Who's right ?
I'm definetly one of the last ones, but dunno if we are right because I don't know what we are saying..:lol:

[Sobek]

The "believers", who still think ATI has some sort of smoke screen in place (a crippled driver, on purpose) ready to beat the 8800's, and back those theories with certain discrepancies in known tests.

Definitely one of the second two... At present if it suddenly turned out that ATi was intentionally crippling the R600 drivers to date to serve up a big whammy come launch time, I (and I would bet a lot of others) would be very, very pissed off. They'd have to be smart enough to realise subpar pre-launch performance reviews are going to absolutely sh*t on thier new lineups reputation come launch time, no matter how good it is (well...if it were extremely good, like 200% faster :razz: ) people are going to have the bitter taste in their mouth.

I know a good deal of people (rough translation : a LOT) who get their information from idiots who read into every single rumor or leak they hear about, to reading in magazines like PC Powerplay or Atomic or somesuch other nonsense, and the moment they see "R600 barely holds a candle to Nvidias G80" (as we've seen a couple of places saying), they'll instantly up-and-buy an 8800 because Nvidia seems like a more trustworthy platform in that case.

I just find the idea of misleading driver shenanigans to be a stupid move, not everyone who buys highend cards is smart enough to wait things out and be 100% sure which is better or not.

[Razor1]

Don't forget Kyle's comments over at [H] forums a few days ago.



So, we basically have three different camps right now:

- The "believers", who still think ATI has some sort of smoke screen in place (a crippled driver, on purpose) ready to beat the 8800's, and back those theories with certain discrepancies in known tests.

- The "nay-sayers", who have either higher-than-average access to hardware and/or software vendors and IHV's, or actually have hardware on their hands and are doing some tests on them (DT, Kyle, etc).

- The little known value-minded semi-enthusiast who would like to see a price-war (Ã* la AMD vs Intel) in order to have early access to high-grade GPU hardware on the cheap (ok, that's me :D).


Who's right ?


There is more that points to not so good or comparative with more power usage, then OMFGWTFSAUCE, I"m sticking with marginally faster, with more power usage, I'm not thinking its 15-20% faster anymore, more like 5-10% at most. I hope.

[nutball]

So, we basically have three different camps right now:

I think you missed out the "fiddling while Rome burns" camp.

[Razor1]

How long would it take for AMD to go with 65nm, from the time of delay of the r600 intial Editor's day? I'm thinking 6 months at least? Just trying to see if they were confident to setup editor's day and they came across problems, they wouldn't be looking at making the r600 65nm, it would have to be the r650, time would just not be enough.

Trying to decipher Kombatants statement of the dailytech card not being the xtx that will be released but it is the xtx that was planned for before.

[INKster]

I think you missed out the "fiddling while Rome burns" camp.

Better Rome than me. :D

[INKster]

There is more that points to not so good or comparative with more power usage, then OMFGWTFSAUCE, I"m sticking with marginally faster, with more power usage, I'm not thinking its 15-20% faster anymore, more like 5-10% at most. I hope.

The "10% more to 10% less" scenario is enough to engage the third option (i hope).

[Rangers]

one slide of $399 Radeon HD 2900XT confirms 16 texture units.

http://r800.blogspot.com/2007/04/radeon-hd-2900xt.html


It's still absolutely ridiculous that AMD hasn't upped their texture counts in three generations.

At least 24 probably would have made it very competitive with GTX.

16 means it's going to be very shader heavy, like R580.

AMD can only hope Crysis shows off the architecture, then again by the time that game comes out everybody will have bought Nvidia and nobody will care, the fate of the R580 series.

I hope AMD really cleans up ATI's engineers over this. Meaning fire a lot of them.

I have a question, we saw that X2900XT seemed to do unduly well in 3Dmark versus the games. Does that point to texture limits?

I would expect AMD's marketing to point out that the architecture is very shader heavy. It should have a good deal more shader power than G80 if my theories are right, while of course being much slower in actual games.

[Razor1]

3dmark 06 pushes pixel and vertex shaders hard, I don't think it is texture bound at all. Futuremark tends to make their benchmark in the eyes of what they think next gen games will be like, and I"m pretty sure they would have focused on shaders more then pushing the TMU's, I don't they really care or look much into the ROP's, thats why the default res is what it is and default settings. What we need is a good variation of benchmarks at different reses that can be varified. Dailytech got some results but well hmmm, nothing can be said about them :s.

[INKster]

I hope AMD really cleans up ATI's engineers over this. Meaning fire a lot of them.

And make things even worse ?

You know what would happen to those "oh-poor-fired-devils", right ?
They would start hearing mermaid voices singing words like "Larrabee" and "G100" (in other words, they would flock over to work for the very adversaries AMD is - or will, in a near future- be trying to compete with).
ATI's engineers are some of the best they have at AMD right now. You can't be innovative without new IP and the people who produce it, that's where it all starts.

[Rangers]

And make things even worse ?

You know what would happen to those "oh-poor-fired-devils", right ?
They would start hearing mermaid voices singing words like "Larrabee" and "G100" (in other words, they would flock over to work for the very adversaries AMD is - or will, in a near future- be trying to compete with).

And put 16 TMU's in G90 to go with 512 scalar shaders? Let em go :lol:

[INKster]

And put 16 TMU's in G90 to go with 512 scalar shaders? Let em go :lol:

Unless they want to cut back on what's already there in G80, i don't think David Kirk would let them do that. ;)

[trinibwoy]

What I find curious is that we still haven't seen any indication of what ATI plans to do with all that bandwidth potential given the relatively low TMU/ROP counts. Is it really just a case of needing to step up from 256-bit buses and they decided to go for the whole hog in one shot?

[ChrisRay]

And make things even worse ?

You know what would happen to those "oh-poor-fired-devils", right ?
They would start hearing mermaid voices singing words like "Larrabee" and "G100" (in other words, they would flock over to work for the very adversaries AMD is - or will, in a near future- be trying to compete with).
ATI's engineers are some of the best they have at AMD right now. You can't be innovative without new IP and the people who produce it, that's where it all starts.

Yup. Talented engineers would easily find a place at Nvidia or other SIlicon manu. Just because a company makes a bad design choice. Making an assumption that this is even the case. Doesnt mean engineers are easy to come by. Especially in this line of work.

[mao5]

It's 5R, means the card is a NDA XT with 512MB Hynnix GDDR3 mem.

http://bbs.souyo.com/uploadimage/igdbol_2007428_1284.jpg

[Ailuros]

Pete,

Ignore for a second the DT results and other sites' newsblurb and pretend you haven't read any of it. We know a launch was planned some weeks ago and has been postponed literally last minute.

Shortly after that we read official AMD statements as to why the launch was cancelled in the first place. The reason given didn't sit very well with me, as I have a hard time to believe that any company would sit on a finished product and lose sales in the meantime.

There could be many minor reasons involved, amongst them things like the driver not have reached back then more efficiency and/or someone recognized that they need some additional AA mode beyond 8x Multisampling as an answer to NV's CSAA. Who remember's Kombatant's comment about the delay?

Your suspicion about a form of that former "temporal AA" thingy isn't without merit, but the term "custom filter" leads me to think of something else. Anything that adds higher AA quality with Nx MSAA samples could be highly competitive and a good answer to NV's coverage sampling without killing memory footprint and bandwidth requirements.

As I've said many times in the past - sample amounts or supposed EER values are just one part of the story; there are other factors which if you add them all up and see one specific mode in realtime, then you can only define what is really useful and what serves for nonsensical marketing stunts.

Take f.e. the 32xS mode you can enable on G80's since recently; it's a combination of 8xRGMS and 4xOGSS. AA analyzing utilities show a rather irritating sample placement and in a couple of fillrate limited tests I ran it through in 1600*1200, 4*4 OGSS (16xSSAA) ends up actually faster. Thus uncheck performance as a starter and if you now go into the quality comparison I haven't seen anything as of yet that makes the high performance penalty worth over what 16xS (4xRGMS + 4xOGSS) delivers.

And to clarify I was amongst those that wanted to have the hybrid MS/SS and pure SSAA modes back, since I can find good use for them for older games I still play. Heck even if you compare 8x Multisampling vs. 16xQ on a G80, the usefullness of the latter can be debatable.

Anyway whatever that CFAA thingy after all is I have the feeling that no more than 8x Transparency AA samples are possible; once we found out what it stands for I'm sure comparisons to CSAA will be quite interesting.

Last but not least yeah we all know how "important" sterile performance ratings really are. Those supposed "journos" and other industry folks mark products usually and sadly enough by a couple of 3dmark results; I guess folks that mark products by a price/performance/output quality ratio are rather rare. I never expected R600 to be a huge "winner" as I've never personally expected to be a complete "loser" either, rather a very competitive product. Just the fact of its rather large delay is unfortunate for AMD's sales and a eulogy for NV's sales on the other.

[Geeforcer]

It's 5R, means the card is a NDA XT with 512MB Hynnix GDDR3 mem.

http://bbs.souyo.com/uploadimage/igdbol_2007428_1284.jpg

Do I get a cookie? ;)

Seriously, thanks for the confirmantion.

[mao5]

Do I get a cookie? ;)

Seriously, thanks for the confirmantion.

This pic is a new one, but the card is same

[silent_guy]

I hope AMD really cleans up ATI's engineers over this. Meaning fire a lot of them.

Would you care to explain how exactly that would solve the problem and how you would solve the consequences?

Disappointment for technical failure turning into anger about unfair treatment? Obvious increase in job insecurity and fear of screwing up? Turf wars and instability due to managers being shuffled around? Even higher workloads (there's a hiring freeze after all) ? Inheriting and maintaining the ugly code of a guy that left?

There's also the inconvenient reality that those who are allowed to stay have little reason to do so, even if R600 is a technical marvel (!): no budgets for bonuses, a small impact on the bottom line (as part of a larger entity who's other products are suffering), stock options that are far under water, and, finally, a zillion booming companies within a 10 mile radius who are aggressively courting exactly those chief engineers who are supposed to come up with the next big thing.

Having been through the downturn after the bubble, I can tell you that it takes exceptional talent as a manager to keep your troops together when things are down, and that was in an environment where it was bad for other companies too, so the defection rate was low.

This is not the case now: the biggest problem of ATI will be to figure out how to keep the best from leaving, rather than to find out who to fire.

Jack Welch, you are not...

[Rangers]

Would you care to explain how exactly that would solve the problem and how you would solve the consequences?

Disappointment for technical failure turning into anger about unfair treatment? Obvious increase in job insecurity and fear from screwing up? Turf wars and instability due to managers being shuffled around? Even higher workloads (there's a hiring freeze after all) ? Inheriting and maintaining the ugly code of a guy that left?

There's also the inconvenient reality that those who are allowed to stay have little reason to do so, even if R600 is a technical marvel (!): no budgets for bonuses, a small impact on the bottom line (as part of a larger entity who's other products are suffering), stock options that are far under water, and, finally, a zillion booming companies within a 10 mile radius who are aggressively courting exactly those chief engineers who are supposed to come up with the next big thing.

Having been through the downturn after the bubble, I can tell you that it takes exceptional talent as a manager to keep your troops together when things are down, and that was in an environment where it was bad for other companies too, so the defection rate was low.

This is not the case now: the biggest problem of ATI will be to figure out how to keep the best from leaving, rather than to find out who to fire.

Jack Welch, you are not...

What do you think Nvidia did the FX 5800 engineers?

I'll wager that many of those engineers responsible for key strategic decisions for that product where not looked kindly on. Of course that's just rampant speculation on my part.

[silent_guy]

What do you think Nvidia did the FX 5800 engineers?

I'll wager that many of those engineers responsible for key strategic decisions for that product where not looked kindly on. Of course that's just rampant speculation on my part.

Pointless to argue a specific case since we don't know. But I do know that I have never seen anybody fired for technical failures, unless it was a case of gross negligence where clear rules were repeatedly and knowingly violated.

Key strategic technical decisions are made by key strategic technical people. And they get to that position because they have proved themselves at being very smart. It is rare to see people promoted into a higher technical position when they don't deserve it.

A complex chip is always a gamble: there's a long lead time, the market can change unexpectedly, some features were considered too important, other were ignored. Architecture modeling didn't uncover certain flaws. There's a ton of things that can go wrong. And let's not forget: the other team gambled right or simply outsmarted you. It happens. Firing your own smart people isn't going to help you one bit.

[mao5]

Radeon HD 2900XT official Tests Result

http://r800.blogspot.com/

[Xenus]

I have to say I'd call bullshit on that Direct X 10 slide 4 to 213 and 1 to 52 seems a litle too outlandish to me.

[Geeforcer]

Radeon HD 2900XT official Tests Result

http://r800.blogspot.com/

If that's what they showed in Tunis, no wonder the journalist contingent was rolling their collective eyes.

[silent_guy]

I hope AMD really cleans up ATI's engineers over this. Meaning fire a lot of them.

Let's turn it around for a minute: you are a valued staff engineer at ATI/AMD in the Valley. You have a decent salary, but the last few years, increases have been disappointing due to little or no profit. Bonuses were low too. The acquisition made the stock jump a little bit and some money was made, but all your current stock options are underwater, maybe for years to come. Morale is very low. Cost cutting measures have been announced. Your friends at Nvidia, Google, Cisco, Broadcom, Marvell, Apple, Intel, and a ton of exciting startups are doing great, are having fun, and make quite a bit more money too. Your lunch buddy just resigned. You have 2 recruiters calling you each week.

Would you stay? I wouldn't...

Edit: ... unless they're really nice to me. :wink:

[trinibwoy]

Radeon HD 2900XT official Tests Result

http://r800.blogspot.com/

If those are ATI's numbers then things really aren't looking good for R600.

[Galduta]

http://i4.photobucket.com/albums/y117/jonelo/002.jpg


http://i4.photobucket.com/albums/y117/jonelo/003.jpg
http://forum.beyond3d.com/%5BIMG%5Dhttp://i4.photobucket.com/albums/y117/jonelo/003.jpg%5B/IMG%5Dhttp://bp2.blogger.com/_hqH07Z7-z90/RjL6efShOtI/AAAAAAAAAE4/spAjGo_TXaw/s1600/003.jpg

[vertex_shader]

It's 5R, means the card is a NDA XT with 512MB Hynnix GDDR3 mem.

http://bbs.souyo.com/uploadimage/igdbol_2007428_1284.jpg

The numbers in the first row:
HY(band=hynix) 5R(type=gddr3) S(process/power supply=vdd 1.8v,vddq 18.v) 57(density/refresh=256M 4K refresh/32ms) 32(organization=x32) 2(bank=4banks) 5(interface=pod18) A(die generation=2nd generation)

I can't decode the second row using this .pdf (http://www.hynix.com/eng/02_products/01_dram/down/GDDR.pdf) because the numbers are weird.

I think thats number 1 left in the second row means 1.0ns memory.

I think this is the same memory:
http://www.hkgolden.com/ArticleBase/big/1703/6506.jpg

[SugarCoat]

I have to say I'd call bullshit on that Direct X 10 slide 4 to 213 and 1 to 52 seems a litle too outlandish to me.

not if theres a driver issue which isnt out of the realm of possability at all. Most of us have seen slides that exploit driver bugs of the other side before from both teams. Remember the R520 slides for the FEAR demo?

[Russell]

If those are ATI's numbers then things really aren't looking good for R600.

R600XT sits exactly where I would expect it to sit. Close to, though slightly above the 8800GTS in most benchmarks. I am, of course, ignoring their "DX10 HD Gaming" benchmarks, as the results reek of propoganda worse than Dailytech's "benchmarks".

[vertex_shader]

Radeon HD 2900XT official Tests Result

http://r800.blogspot.com/

HDR+AA tests single card hd2900xt results looks very bad with so huge BW.

CF looks much better, but thats already 750$.

Why the HDR and HDR+AA results are so ugly?

[Geeforcer]

R600XT sits exactly where I would expect it to sit. Close to, though slightly above the 8800GTS in most benchmarks. I am, of course, ignoring their "DX10 HD Gaming" benchmarks, as the results reek of propoganda worse than Dailytech's "benchmarks".

Call me insane and/or crazy, but I for some reason expected the R600 to do better then be competetive with Nvida's second teir card from half a year ago.

[trinibwoy]

R600XT sits exactly where I would expect it to sit. Close to, though slightly above the 8800GTS in most benchmarks.

You really expected the XT to be slightly above a GTS? Don't let ATi's pricing make you lose sight of the fact that this is a full R600 vs a cut down G80. I think most people fully expected R600 in its full configuration to be going up against the GTX and beating it too.

It will be interesting to see if reviewers use adaptive AA in testing though - that's one setting sorely lacking from mainstream reviews IMO. In any case, these results are probably the typical IHV FUD. Since when does any Nvidia card not scale with SLI in Quake? :)

[SugarCoat]

R600XT sits exactly where I would expect it to sit. Close to, though slightly above the 8800GTS in most benchmarks. I am, of course, ignoring their "DX10 HD Gaming" benchmarks, as the results reek of propoganda worse than Dailytech's "benchmarks".

then realize the XT is launching as a flagship model 6 months later and also notice that the GTS cards have been selling at about a ~$350 or less street price even without it around. Unless this is a very bad joke or a switch needs to be flipped this has the makings of one of the worst launches since the Geforce FX. I am not pleased one bit as i had some really high hopes too.

Before anyone says it, i should also point out that even months ago we've been under the impression that this would still be a DX9 powerhouse, even before DX10, and now its either proving to be the opposite or neither.

The time table screwed them, this would of been far more acceptable if there hadnt been a huge delay.

[trinibwoy]

Anyone else notice the difference in performance drop when 16xAF was turned on in Oblivion?

[Geeforcer]

I hope that this is the most elaborate hoax ever.

[Russell]

You really expected the XT to be slightly above a GTS? Don't let ATi's pricing make you lose sight of the fact that this is a full R600 vs a cut down G80. I think most people fully expected R600 in its full configuration to be going up against the GTX and beating it too.


I didn't 6 months ago, however for the last month or so this is more or less where the signs pointed to it sitting. I confess though I expected a bit more of a lead than we're seeing.



I hope that this is the most elaborate hoax ever.

I agree. I'd love to be surprised somehow by this.

[mao5]

I think ati learned lesson from previous failure. Radeon HD 2900XT is set to gain high end market not the damn shit performance crown with no value.

[mao5]

I'm keeping on disclosing:

R600XT 3DMARK05

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

mh,mh,mh,

http://www.chiphell.com/attachments/month_0704/20070428_2b1913d30fceba3a38955wU82mucxbN6.jpg

[trinibwoy]

mao5, why bother with the high-end at all? They should've just made the fastest mid-range card and called it a day.....

Edit: w00t so now the XT is beating up on the GTX in 3dmark05? This gets better by the minute...literally! :lol:

[Geeforcer]

Just wait to see the number it does on G80 in 3Dmark 99.

[mao5]

mao5, why bother with the high-end at all? They should've just made the fastest mid-range card and called it a day.....

Edit: w00t so now the XT is beating up on the GTX in 3dmark05? This gets better by the minute...literally! :lol:

As I'm saying, AMD ATI and Radeon HD 2900XT are in bushfighting, got the battle guidance from great chinese chairman MAO.

[mao5]

Just wait to see the number it does on G80 in 3Dmark 99.

you can do that with GTX and 2900XT, can't wait.

[Lockon]

this is a full R600 vs a cut down G80. I think most people fully expected R600 in its full configuration to be going up against the GTX and beating it too.

And you know this how?

[Anteru]

Somehow I don't trust those benchmarks any more, there's no consistency at all, the XT ranges from slightly-slower/on-par with the GTS to superior to the GTX and there is no real trend visible which results are more likely. Either this is a large FUD campaign, or people are making numbers up to get more page impressions.

I believe the XT is on-par/slightly better than the GTS, even if G80 did catch AMD totally by surprise - I mean, they must have had something in plan to compete against those GX2 boards, so they surely aimed for doubling the performance. Maybe it turned out to be more difficult than slapping twice the shader units on board, but they had their time.

What's the status of the XTX, reading some posts here, has the XTX been canceled or further delayed?

[trinibwoy]

And you know this how?

Ah, the old prove the unprovable trick. Nice. You're fully welcome to believe otherwise. Do you have any rumours/hints/info that points to the XT having less active units than the XTX? If you do, please share :)

[Russell]

What's the status of the XTX, reading some posts here, has the XTX been canceled or further delayed?

Nobody is sure. I'm not even sure if AMD knows. We know that 1GB of GDDR4 is very expensive for relatively little gain and unless they can crank up the core clock there's very little performance boost over the XT model (at least today with games that don't require 1GB of memory for textures). If they release an XTX, I think this is what they'll have to do. Otherwise,what's the point. AMD can hold the XTX card to see what nVidia does with their Ultra then decide if they want to release something or not. Perhaps they will save R650 to be the XTX. There's lots they can do, but nobody really knows what.

[silent_guy]

Ah, the old prove the unprovable trick. Nice. You're fully welcome to believe otherwise. Do you have any rumours/hints/info that points to the XT having less active units than the XTX? If you do, please share :)

The lad must have overlooked that 512 == 512 but 320 != 384. It happens...
I guess it's still possible the XT version cuts down the number of shaders, but I doubt it.

[mao5]

Ah, the old prove the unprovable trick. Nice. You're fully welcome to believe otherwise. Do you have any rumours/hints/info that points to the XT having less active units than the XTX? If you do, please share :)

Bullxxxx, I don't need to remind you of the specification of GTS with 640MB GDDR3 and 1200MHz shader clock.

[mao5]

Keep on doing 1280x1024:

http://www.chiphell.com/attachments/month_0704/20070428_ad2b8ea486c0bd5382d50MzBzMonS3vZ.jpg

[Geeforcer]

Mao, before you post a 15K 3dmark05 score, would you explain what they mean, exactly?

[Robin B]

Dont see your point mao5, after all the 600XT should be named the 600XL after all we have seen from this card, they wasted 700mill transistors on a card that afterall should be faster than the GTX since it is 6 months late to the game.

[Lockon]

The lad must have overlooked that 512 == 512 but 320 != 384. It happens...
I guess it's still possible the XT version cuts down the number of shaders, but I doubt it.
I guess thats why I asked :)

[Geeforcer]

Ah, the old prove the unprovable trick. Nice. You're fully welcome to believe otherwise. Do you have any rumours/hints/info that points to the XT having less active units than the XTX? If you do, please share :)

Apperenlty AMD has multiple 700M transitor R600s.

[mao5]

1280x1024 4aa 8af

http://www.chiphell.com/attachments/month_0704/20070428_8f56ff8ea85706f3ef87VwMoywzN1HmG.jpg

[IbaneZ]

How the performance hit with 8xAA on a GTS? I wonder if the XT pulls away more at 4xAA, or if the gap closes.



1280x1024 4aa 8af



Impressive. Any GTS/GTX users here? Run 3dmark-05 with those settings then. :)

[vertex_shader]

1280x1024 4aa 8af

http://www.chiphell.com/attachments/month_0704/20070428_8f56ff8ea85706f3ef87VwMoywzN1HmG.jpg

Looks like the free 4xaa AA what some rumors suggest not so free, or only in 640x480 :wink:

[mao5]

Looks like the free 4xaa AA what some rumors suggest not so free, or only in 640x480 :wink:

don't worry, be happy, those scores above with an old driver.

[epicstruggle]

From what I gather, it looks like AMD is asking for a lot from the XT.

Its competing against a GTS (and winning) and competing, in a CF setup, against both the GTX and the ULTRA.

It will be interesting to see how it all works out.

[mao5]

From what I gather, it looks like AMD is asking for a lot from the XT.

Its competing against a GTS (and winning) and competing, in a CF setup, against both the GTX and the ULTRA.

It will be interesting to see how it all works out.

agree, ultra' s core and mem speed is much less than people expected, pls see theinq's report, this time theinq is right totally.

[MrBelmontvedere]

what all is known about the AA modes in R600? I found this on inquirer:

Anisotropic filtering and AntiAliasing were vastly improved as well, and R600 marchitecture marks off the start of "NextGen AA". A new AA mode is called Custom Filter AA (CFAA), and it peaks at 24x. 8x MultiSample AA mode was completely changed, so expect some fireworks and comparisons between G80 and R600 hardware. Of course, all existing AA tech is supported.

is this CFAA another implementation like nvidias coverage AA? but they say 8x multisample is changed too, hmmm..... xD

[Ailuros]

Apperenlty AMD has multiple 700M transitor R600s.

Alas if those things wouldn't have all that GFLOP throughput; that fillrate looks more than just weak compared to the G80 bi-AF fillrate.

[Galduta]

http://i4.photobucket.com/albums/y117/jonelo/gpun2073.jpg

http://i4.photobucket.com/albums/y117/jonelo/gpun2077.jpg


http://i4.photobucket.com/albums/y117/jonelo/gpun2076.jpg


http://www.gpumania.com/foro/viewtopic.php?t=8419

Core clock is 757 mhz ...


"the noise in 2D is very low, inaudible, but in 3D a acute noise can be listened. The temperature in 2D is very high (68º), although it has not verified if the CC are measuring it well. We cannot offer more information "

[IbaneZ]

What is this, official leak day? :lol: More more more! Keep it coming.

In that pic they run it with two six pins connectors.

[bigtabs]

http://img205.imageshack.us/img205/6453/gpun2076pv9.jpg

:smile:

[ChrisRay]

How about some control panel shots of the AA settings? :p

[Ailuros]

How about some control panel shots of the AA settings? :p

Vivere pericolosamente :P

[R300King!]

Here are some Crossfire % numbers compared to a single XT

Slide 1.7 (http://bp2.blogger.com/_BabjUDZIqPw/RjLz3GGObFI/AAAAAAAAACU/e3GewiH5ayI/s1600-h/001.jpg) Max Image Quality: 16x12 (8xAA 16xAF) with Adaptive AA

Serious Sam 2 (Shield) - 78.4%
5Quake 4 (id_demo) - 3%
Prey (guru3d_5) - 68.6%
Half Life 2 LC (lc_demo2) - 88.8%
Half Life 2 EP1 (EP_C17) - 83.6%
Half Life 2 (prison05) - 74.2%
FEAR - 77.7%
Far Cry (research) - 100%
Doom 3 (demo1) - 65.1%
3DMark06 - 87.5%
3DMark05 - 73.9%

Slide 1.7 average
Games Crossfire = 76.6%
3DMark Crossfire = 80.7%


----------

Slide 1.8 (http://bp1.blogger.com/_BabjUDZIqPw/RjL0F2GObGI/AAAAAAAAACc/2kJL3vJyWGs/s1600-h/002.jpg) HDR Gaming

Serious Sam2 (25x16 16xAF) - 63.1%
Rainbow Six: Vegas (19x12 16xAF) - 0%
Ghost Recon: AW (25x16 16AF) - 86.6%
Far Cry HDR (25x16 16xAF) - 87%
ES IV: Oblivion (25x16) - 72.9%
ES IV: Oblivion (25x16 16xAF) - 79.3%
ES IV: Oblivion (19x12 16xAF) - 65%

Slide 1.8 average
Games Crossfire = 64.8%


----------

Slide 1.9 (http://bp2.blogger.com/_BabjUDZIqPw/RjL0VGGObHI/AAAAAAAAACk/43mUeMXDEdg/s1600-h/003.jpg) HDR+AA Gaming

Serious Sam2 (16x12 4xAA 8xAF) - 80.8%
Far Cry HDR (16x12 4xAA 8xAF) - 81.3%
ES IV: Oblivion (16x12 4xAA 8xAF) - 77.1%

Slide 1.9 average
Games Crossfire = 79.7%


Crossfire scaling looks good if these are correct. :)

[mao5]

hehe today is a leakday

[epicstruggle]

hehe today is a leakday
Better than just quoting fud from theInq/fuad. ;)

[mao5]

Better than just quoting fud from theInq/fuad. ;)

yes

[Subtlesnake]

Maybe they can still keep the overall performance crown with crossfire!

[Razor1]

Single card mode it seems like its not much faster then the GTS?

And what the hell they are comparing to a buggy SLi driver? Ya know if nV has fixed SLi by review time......? Whats going to happen.

[Twinkie]

Single card mode it seems like its not much faster then the GTS?

And what the hell they are comparing to a buggy SLi driver? Ya know if nV has fixed SLi by review time......? Whats going to happen.

That was what i was thinking as well. Does SLi even work properly? Im wondering the OS that they used since SLi is non existent in Vista. Some games might work, but others see all sorts of problems.

[ChrisRay]

That was what i was thinking as well. Does SLi even work properly? Im wondering the OS that they used since SLi is non existent in Vista. Some games might work, but others see all sorts of problems.

It's most a case of newer AFR profiles ((We saw this with XP too when SLI first came out)) not properly swapping frames. Alot of the XP compatibility tweaks for SLI are slowly making their way in. The 158.XX series fixed most of them.

Chris

[Sobek]

http://img205.imageshack.us/img205/6453/gpun2076pv9.jpg

:smile:

I was wondering about that too...

Though you could probably just chalk it down to botchy Photoshop work while he/she/it/whatever was touching up the other part. Still an odd thing to see, what would they be pasting over there anyway? :???:

[IbaneZ]

Is the 8800 GTS "crippled" vs GTX when is comes to TMU's? Or does it have 32 too?

[fellix]

GTS have 96 shader units, which means 6 clusters * 4 texture unts = 24 (48 bileps) textured pixels per cycle.

[jamis]

GTS have 96 shader units, which means 6 clusters * 4 texture unts = 24 (48 bileps) textured pixels per cycle.
Hmm..Maybe they (Ati) screwed up and targeted R600 against chip like this all along?
Maybe they were fooled and didn't realize it has separate IO chip and thought G80 was going to be smaller....:lol:

[IbaneZ]

GTS have 96 shader units, which means 6 clusters * 4 texture unts = 24 (48 bileps) textured pixels per cycle.

Thanks.

I guess I should have been able to figure that one out by myself, but my brain is crippled by too much Koskenkorva. :???:

[Razor1]

Hmm..Maybe they (Ati) screwed up and targeted R600 against chip like this all along?
Maybe they were fooled and didn't realize it has separate IO chip and thought G80 was going to be smaller....:lol:


I think very few people knew what the g80 was till the last few weeks. But the target always is about x2 the performance of last gen at least at very high settings. So I don't see how ATi would have miss calculated on that.

[satein]

Hmm..Maybe they (Ati) screwed up and targeted R600 against chip like this all along?
Maybe they were fooled and didn't realize it has separate IO chip and thought G80 was going to be smaller....:lol:

I believe they (ATi/AMD & NV) took different design proposal on their GPU. I didn't see any point on how to design your GPU base on the opponent's specification.

ATi/AMD might believe that the proposal they took will convey a better performance on what they will offer to us.

Regards,

[jamis]

I think very few people knew what the g80 was till the last few weeks. But the target always is about x2 the performance of last gen at least at very high settings. So I don't see how ATi would have miss calculated on that.
But they did know that G80 is on 90nm. And on 90nm it couldn't be what it is without the separate IO chip. So who knows..:grin:

[jamis]

...
Wel OK, just a thought...:lol:

[compres]

What is this, official leak day? :lol: More more more! Keep it coming.

In that pic they run it with two six pins connectors.

I don't car eif they are fake, at least I am entertained now lol.

[PP]

I was wondering about that too...

Though you could probably just chalk it down to botchy Photoshop work while he/she/it/whatever was touching up the other part. Still an odd thing to see, what would they be pasting over there anyway? :???:

It's an issue related to Vista and CCC translation to me, it happens in the 2 CCC shots published (+ another one not published).

PD.: I only received the pics, don't own the card unfortunately.

[Arty]

Global Illumination Demo on R600

http://www.hardspell.com/english/doc/showcont.asp?news_id=479&pageid=580

[satein]

Global Illumination Demo on R600

http://www.hardspell.com/english/doc/showcont.asp?news_id=479&pageid=580

Cool :cool:

Those images reminded me of Humus's work.

[neliz]

When I view the fud slide with dx10 juarez and the explicit mention of physics through geometry shading I suddenly remember the mentions about r600 being "a lot faster" to inq's "5x faster at geometry shading"

Maybe there's still some hope


(and yes.. I know.. I *HAVE* to wear pink glasses.. I ordered one..)

[tEd]

damit i wanna know more about the new AA. Why do those leaker always leak the wrong stuff :lol:

[Jawed]

Well, those AMD slides are pretty damning (presuming they're the real thing, of course). Look no further than the HDR+AA benchmarks. DX9 case closed.

The TMUs and ROPs may well be more efficient than R5xx, but they omitted to put in enough of them to use all that bandwidth :!:

Orton's "we didn't have the tools" excuse at the TFLOP in a box/RD690 press conference, when asked about heat/power sounded very ominous.

R600 clearly needs to be able to run at 1GHz core to be competitive with 8800GTX in DX9 games. And then it would still have ~23% more bandwidth than it knows what to do with. The core needs this bump because there was no more transistor budget left after all the D3D10 goodness...

HD2900XT v 8800GTS-640:

bandwidth 166%
fillrate 119%
AA rate (presuming 4x per loop) 119%
bilinear texture rate 99%
fragment rate 124%
MAD GFLOPs 206%I'm sorta mystified where all the performance disappears - the HDR+AA tests should be fillrate/bandwidth limited... Are we going to have to resort to excusing this with "drivers"?

Jawed

[Jawed]

Global Illumination Demo on R600

http://www.hardspell.com/english/doc/showcont.asp?news_id=479&pageid=580
Hey, at least that looks groovy. If G80 is really running that in the 1-2FPS range (drivers?) then, well, you've gotta laugh.

Jawed

[trinibwoy]

Are we going to have to resort to excusing this with "drivers"?

It's as good an excuse as any. It does remind me of ATI's insistence that the hardware was ready though. What if the delays were just due to big software engineering problems?

[Arty]

Fuad latest is about the final XT clocks.

2D: 507/514
3D: 757/828

[AnarchX]

[...]
fillrate 119%
[...]


If R600 has only 16TAUs/TMUs, the 2900XT has 49% of the fillrate of 8800GTS(20TAUs/40TMUs), if you are using AF.

[AnarchX]

If R600 has only 16TAUs/TMUs, the 2900XT has 49% of the fillrate of 8800GTS(20TAUs/40TMUs), if you are using AF.
correction:

8800GTS has 24TAUs/48TMUs.

[IbaneZ]

Fuad latest is about the final XT clocks.

2D: 507/514
3D: 757/828

Yeah, he's probably keeping an eye on this thread.

http://forum.beyond3d.com/showpost.php?p=977769&postcount=3317

[Jawed]

It's as good an excuse as any. It does remind me of ATI's insistence that the hardware was ready though. What if the delays were just due to big software engineering problems?
I'd feel much more comfortable with gameplay performance comparisons based upon R580+. There is a reason why B3D reviews used to focus on architectural scaling based on prior GPUs from the same IHV.

R5xx saw significant driver improvements in ROP/memory performance for months after its release (though remember that released drivers lag behind driver development by about 3 months, it seems). But right now, I just can't help feeling they've had more than long enough to sort this stuff out.

The 3DMk06 1600x1200 8xAA/16xAF with adaptive AA test should be home turf for R600, but it's a dead heat with 8800GTS.

RV630 is looking like it's destined for an early grave, $125 by September.

Jawed

[fellix]

Global Illumination Demo on R600

http://www.hardspell.com/english/doc/showcont.asp?news_id=479&pageid=580
The font, used in the app window, is looking exactly the same as Humus used in his demos. :grin:
Does he...

Speaking of GI, I wonder how R600 will bare the same effect enabled in S.T.A.L.K.E.R.

[Razor1]

why will the rv630 will die so quickly?

[Jawed]

correction:

8800GTS has 24TAUs/48TMUs.
Fillrate is what ROPs do. For texturing I listed bilinear texture rate, 99%.

Jawed

[Bjorn]

Well, those AMD slides are pretty damning (presuming they're the real thing, of course). Look no further than the HDR+AA benchmarks. DX9 case closed.


Why am i not surprised, it looks like it's a 400$ card that performs like a 400$ card.

[trinibwoy]

Why am i not surprised, it looks like it's a 400$ card that performs like a 400$ card.

Looks like a $600 card that just couldn't cut it to me.

[AnarchX]

Fillrate is what ROPs do. For texturing I listed bilinear texture rate, 99%.


Of course. I was a little bit distracted so I quoted the wrong, sorry.

But how i said, the bilinear texture rate is only 49%, if AF is used, what is usual for this videocard-class. (In case R600 has only 16 bilinear TMUs, how it looks at the moment.)

[mao5]

When I view the fud slide with dx10 juarez and the explicit mention of physics through geometry shading I suddenly remember the mentions about r600 being "a lot faster" to inq's "5x faster at geometry shading"

Maybe there's still some hope


(and yes.. I know.. I *HAVE* to wear pink glasses.. I ordered one..)

correct

[Jawed]

why will the rv630 will die so quickly?
Because its bilinear texture rate will be ~60% of a stock-G84's at around 800MHz and its bandwidth advantage will be non-existent.

Trilinear/AF maybe won't be so bad, if R6xx has some trilinear/AF tricks up it sleeve. Arguably, it does have some tricks there (actually that's worth a separate posting), because it keeps up with 8800GTS, which definitely has tricks in that department.

Jawed

[Jawed]

The font, used in the app window, is looking exactly the same as Humus used in his demos. :grin:
Does he...
...have an 8800 to check the functionality/performance of his demos? His demos tend not to work well on NVidia hardware, at least not initially.

Jawed

[Jawed]

Why am i not surprised, it looks like it's a 400$ card that performs like a 400$ card.
With a $600 card's memory system.

Jawed

[Jawed]

Of course. I was a little bit distracted so I quoted the wrong, sorry.

But how i said, the bilinear texture rate is only 49%, if AF is used, what is usual for this videocard-class. (In case R600 has only 16 bilinear TMUs, how it looks at the moment.)
Technically, even though G80 has 64 TFs, it can only do 32 bilinear textures per clock.

It appears that R600 does have some zest in trilinear/AF because it's not totally dying when they're turned on. But if we look at the Oblivion 2560x1600 numbers:

HD2900XT - 37 no-AF, 29 AF = -22%
8800GTS-640MB - 34 no-AF, 29 AF = -15%That's the only data point though... And I'm not convinced that G80 isn't a bit of a dud at that resolution, generally, anyway (which would invalidate the AF scaling comparison). 4MP seems to be over the cliff in a lot of game tests for G80.

Jawed

[Silent_Buddha]

Just being ornery here. And postulating a huge WHAT IF...

WHAT IF those DX10 Call of Juarez numbers are correct?

Would this mean that NV's DX10 drivers are completely borked? Or that G80's DX10 support is mostly there as a check box and they don't actually plan on competing in DX10 until the next product? Or just that G80's Geometry Shaders were skimped on and considerably slower than R600's Geometry Shaders?

If it does prove true it would be an interesting case where...

R600 = Poor/Decent DX9 perf yet excellent DX10 perf.
G80 = Excellent DX9 perf yet poor/decent DX10 perf.

So in that case R600 would be the FX5800 in DX9 games while G80 would be the FX5800 in DX10 games? :wink: :smile: :lol: :grin: :lol:

It could be possible that ATI budgeted a LOT more transistors to DX10 than NV did. But consequently skimped on DX9 compared to NV.

A lot of the things in the DX10 Call of Juarez slide mentioned heavy emphasis on the Geometry Shaders. One aspect that R600 is consistantly rumored to be quite a bit faster than G80.

I seriously can't wait until actual reviews can be posted so we can start weeding through what is real and what was just made up.

Regards,
SB

Would this mean that R600 is a 400 USD DX9 card but a 600 USD DX10 card? While the G80 is a 600 USD DX9 card, but a 400 USD DX10 card? :D

[Edit] - If true, one wonders if R600 is more about experiementing with a different setup (ala - R520) with a more competant part coming later (ala - R580).

[Jawed]

Or that G80's DX10 support is mostly there as a check box and they don't actually plan on competing in DX10 until the next product?
Do you think anyone will be surprised if this is the case?

Jawed

[neliz]

ah yes.. my point... and that would bring us back to rumors from a couple of years ago that G80 indeed is a dx9 card with dx10 bolted on.. ati just messed up on the dx9 setup.. so far we've seen a card that isn't that much faster than r580+

[INKster]

Did anyone see this yet ?

http://bbs.expreview.com/attachments/month_0704/20070428_deff9681f706dcab9a16HwlraKhteBqZ.jpg

source: http://bbs.expreview.com

[Bjorn]

Looks like a $600 card that just couldn't cut it to me.

True enough. I was just commenting on the talk about why they priced it at 400$ when it should perform like the 550$ GTX that was talked about here when the price was revealed. And it looks like the reason was the most obvious.

[AnarchX]

It appears that R600 does have some zest in trilinear/AF because it's not totally dying when they're turned on. But if we look at the Oblivion 2560x1600 numbers:

HD2900XT - 37 no-AF, 29 AF = -22%
8800GTS-640MB - 34 no-AF, 29 AF = -15%That's the only data point

The problem is here, that we do not now if the 2900XT was tested with angle-independent HQ-AF, which is default setting for the 8800GTS.
It could make a bigger difference.

[Silent_Buddha]

The problem is here, that we do not now if the 2900XT was tested with angle-independent HQ-AF, which is default setting for the 8800GTS.
It could make a bigger difference.

I'd be really surprised if they weren't tested that way as the R520 was already headed in the direction of angle independant AF. I wouldn't be surprised to see R600 not even offer the option for angle dependant AF. Then again I also wouldn't be surprised if there was still an option for it.

Regards,
SB

[Jawed]

[Edit] - If true, one wonders if R600 is more about experiementing with a different setup (ala - R520) with a more competant part coming later (ala - R580).
At the high end it seems 65nm will merely bring slightly better clocks and less power/heat. I don't think R6xx contains an R520 part.

I think that's what the coding of R600 means: R300, R420, R520. A 0 in that position means initial variant of a fundamental architectural concept... :razz:

Jawed

[CJ]

Radeon HD 2900XT official Tests Result

http://r800.blogspot.com/

Those slides aren't showing the proper scores. ;)

So I wouldn't call them "official" until you've seen the real scores. :cool:

[neliz]

Those slides aren't showing the proper scores. ;)

So I wouldn't call them "official" until you've seen the real scores. :cool:

If I wasn't a polite person I'd shout "cocktease" right about now!

[Silent_Buddha]

Those slides aren't showing the proper scores. ;)

So I wouldn't call them "official" until you've seen the real scores. :cool:

Damn you CJ. :wink: You are one HUGE tease if you've seen the real scores and aren't telling us. :)

Regards,
SB

[fellix]

Did anyone see this yet ?
source: http://bbs.expreview.com
So, looks like a dual Z/Depth rate there, and some hell of a weird (unclear) ALU stacking.
And that L1 texture cache -- btw, was it presented in R5x0 that way?

[Jawed]

Nice picture Inkster, pity about the resolution.

The L1, L2 texture caches are what I was expecting. As is the vertex cache.

I was actually expecting that the constant, z/stencil and colour caches would all be homogenised into one mother of a cache, the L2 (jointly with texture cache). My reasoning was that if R600 has an L2 aware virtual memory system, then a single L2 would satisfy all memory operations, with L1s localised to the specific pipes. I was going to write about this last night but got lazy.

The big news is the memory read/write cache. For streamout and buffer fetch (but not constants, textures or vertices), I guess. It may have big repercussions in GPGPU applications if it allows for full-bandwidth scatters/gathers. Sorta similar to G80's parallel data cache. But, erm, different...

There's a bright yellow programmable tessellator :?:

Also, if you look at the ALUs, there's one big yellow unit and 4 smaller yellow units :?:

Both the texture units and the render back-ends have 4-way blocks and 16-way blocks :?:

Jawed

[trinibwoy]

Programmable tesselator? Dedicated geometry processing hardware?

[IbaneZ]

If I wasn't a polite person I'd shout "cocktease" right about now!

:lol:

[Jawed]

The problem is here, that we do not now if the 2900XT was tested with angle-independent HQ-AF, which is default setting for the 8800GTS.
It could make a bigger difference.
There's a vague rumour that R600 has better AF, so maybe they're OK.

To be honest, I don't know how G80 AF scales depending on whether it's set to default- or highest-quality AF.

Jawed

[mao5]

PHK is so ugly, the insider who provide the ppt to him will be killed by AMD ATI.

[trinibwoy]

PHK is so ugly, the insider who provide the ppt to him will be killed by AMD ATI.

That's a bit harsh :lol:

[fellix]

Any guess, whether R600 will provide full HW [box] filtering across all texture formats (meaning FP)? And what about VTF/R2VB?

[Jawed]

Programmable tesselator? Dedicated geometry processing hardware?
In Xenos you have a Vertex Grouper and Tessellator, a single unit. In R600 it seems there's a Vertex Assembler and a Programmable Tessellator. I imagine they're very close, conceptually, but that "programmable" word is obviously intriguing. Prolly driven with some simple parameters, perhaps only accessible to the driver. In other words, prolly not exposed directly to the developer.

Jawed

[tertsi]

X2900XT@gpuinstr

MAD1 - MAD 1 512 256 100 144.9562 ms 46.2960 GInstr/sec
MAD3 - MAD 3 512 256 100 142.4918 ms 47.0966 GInstr/sec
MAD4 - MAD 4 512 256 100 142.5442 ms 47.0793 GInstr/sec

MUL1 - MUL 1 512 256 100 71.2619 ms 94.1722 GInstr/sec
MUL3 - MUL 3 512 256 100 71.8172 ms 93.4440 GInstr/sec
MUL4 - MUL 4 512 256 100 88.9417 ms 75.4526 GInstr/sec

ADD1 - ADD 1 512 256 100 36.4790 ms 183.9657 GInstr/sec
ADD3 - ADD 3 512 256 100 43.6579 ms 153.7152 GInstr/sec
ADD4 - ADD 4 512 256 100 57.4707 ms 116.7706 GInstr/sec

DP3 - DP3 1 512 256 100 142.6667 ms 47.0389 GInstr/sec
DP3 - DP3 3 512 256 100 141.9595 ms 47.2732 GInstr/sec
DP3 - DP3 4 512 256 100 143.7289 ms 46.6913 GInstr/sec

DP4 - DP4 1 512 256 100 143.0978 ms 46.8972 GInstr/sec
DP4 - DP4 3 512 256 100 142.7959 ms 46.9963 GInstr/sec
DP4 - DP4 4 512 256 100 141.9480 ms 47.2771 GInstr/sec

Maybe more later. Have fun :)

[Jawed]

And what about VTF/R2VB?
They're both required by D3D10 as far as I know.

Jawed

[fellix]

Both of them? Well, I guess that's good for the legacy R5x0 one. :lol:

[trinibwoy]

X2900XT@gpuinstr

Very strange stuff - R600 vec4 ADDS > vec4 MULS > vec4 MADDS ??!!

[Jawed]

Very strange stuff - R600 vec4 ADDS > vec4 MULS > vec4 MADDS ??!!
I suspect his shaders are too short to be meaningful. At the same time, I think the architecture is sufficiently funky that it may not be possible to reveal anything meaningful with a single-instruction shader, which is what I suspect he's using.

I don't understand why this "benchmark" has any credibility.

Jawed

[trinibwoy]

It seems to give expected results on G80 for these simple instruction tests. It even gives the expected MADD results for R600 - 47GInstr/s ties out well with 64 Vec4 MADDs at ~ 750Mhz. What's curious is the much higher vec4 MUL and ADD tests.

[trinibwoy]

MAD1 - MAD 1 512 256 100 144.9562 ms 46.2960 GInstr/sec
MAD3 - MAD 3 512 256 100 142.4918 ms 47.0966 GInstr/sec
MAD4 - MAD 4 512 256 100 142.5442 ms 47.0793 GInstr/sec

Shouldn't these be 2,6 and 8 flops ? Why do you have 1,3,4?

[mao5]

I have a dream, the dream's SPU arranged as 5-way SSP and my dream invade DX10.1

Anyone can understand my dream?

[trinibwoy]

Your dream was posted a couple months ago mao :razz:

[mao5]

Your dream was posted a couple months ago mao :razz:

give me a link

[mao5]

OMG, 3DMARK2001 SE http://www.chiphell.com/images/smilies/b.gif

http://www.chiphell.com/attachments/month_0704/20070428_1c1a353f4effff7f1ecdKWSQZibFPLOw.jpg

OMG, 3DMARK03

http://www.chiphell.com/attachments/month_0704/20070429_794341e05ece44261700h3X8YMnP0Nqf.jpg

[INKster]

OMG, 3DMARK2001 SE http://www.chiphell.com/images/smilies/b.gif

http://www.chiphell.com/attachments/month_0704/20070428_1c1a353f4effff7f1ecdKWSQZibFPLOw.jpg

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

[trinibwoy]

give me a link

lol, no I'm not going to search for it. But there has been talk about R600 going beyond DX10 requirements for a while now. And the talk about vec5 goes back even further given its Xenos heritage.

[Robin B]

Guess this is the fastest dx6 and dx7 card.:wink: Next is 3Dmark99 right ?

[trinibwoy]

OMG, 3DMARK2001 SE
OMG, 3DMARK03

Enough with the pics, what are the system specs?

[Jawed]

What's curious is the much higher vec4 MUL and ADD tests.
That's my point. The combination of a simplistic test and an unknown architecture makes for meaninglessness.

Plus I've already explained why other tests in the suite are knackered.

Maybe R600 has a dedicated vector MAD/DP/SF unit that can do single-cycle vector/scalar instructions that fall into those categories (scalar-only for SF?). In addition to that it has four scalar ADD/MUL units that can be ganged for vector operations or perform as independent scalars. Since a multiplier is much more complicated than an adder, it might be a single cycle adder but a 2-cycle multiplier.

Now, I might have accidentally hit upon the correct architecture there, but it is a guess and that diagram informs the guess as much as those "rates". My guess also requires a varying number of fragments to be issued to run side-by-side through the ALU pipeline. It's whacky...

Jawed

[Sound_Card]

Things just keep heating up. :razz:

[mao5]

lol, no I'm not going to search for it. But there has been talk about R600 going beyond DX10 requirements for a while now. And the talk about vec5 goes back even further given its Xenos heritage.

ok, but do you remember the detail of beyond DX10 REQ? pls tell me.http://www.chiphell.com/images/smilies/l.gif

[nutball]

Man, these 3DMark scores just keep going up and up :cool: Looking better and better for R600 every time. Can't wait for the 3DMarkMCMLXXXXVII scores.

[Jawed]

Shouldn't these be 2,6 and 8 flops ? Why do you have 1,3,4?
These are instruction rates, not flops.

Jawed

[mao5]

Things just keep heating up. :razz:

yeah,baby. http://www.chiphell.com/images/smilies/h.gif

[mao5]

Enough with the pics, what are the system specs?

what I know is the CPU :E6600 2.4G

[trinibwoy]

These are instruction rates, not flops.

Jawed

Yeah he confirmed that it was vector width. I asked because some other gpudip results show flops in that column.

[INKster]

Man, these 3DMark scores just keep going up and up :cool: Looking better and better for R600 every time. Can't wait for the 3DMarkMCMLXXXXVII scores.

MCMLXXXXVII ? I don't even know what that means...
1987 is MCMLXXXVII.
1988 is MCMLXXXVIII.

:D

Going back to topic..., DX10.1 ? I had no idea MS had that spec already set in stone.
Just give us Direct3D 10.0 software first. ;)

[trinibwoy]

ok, but do you remember the detail of beyond DX10 REQ? pls tell me.

Nope, all I remember is people claiming R600 is DX10.1 and other people saying that there's no such thing.

[mao5]

Nope, all I remember is people claiming R600 is DX10.1 and other people saying that there's no such thing.

I guess MSAA+MRT, custom AA and Tesseletor are beyond DX10. MSAA+MRT surely is one of the spec of dx10.1, you can see dx10.1 ppt from MS.

[nutball]

MCMLXXXXVII ? I don't even know what that means...

1997. There are two representations of 90 in roman numerals, LXXXX and XC.

Anyway, yes, back on topic I'm still finding this speculation about ATI engineering in performant features that won't get used in the generation rather problematic. I find it *really* tough to accept that that it is good idea.

[Sound_Card]

Funny, I remember seeing old slides of R600 from back in Novmeber I believe. Intresting enough, some DX 10.1 slides were leaked almost right afterwards in that time frame. Sigh, I'm trying to dig those slides.

[Sobek]

OMG, 3DMARK2001 SE http://www.chiphell.com/images/smilies/b.gif

http://www.chiphell.com/attachments/month_0704/20070428_1c1a353f4effff7f1ecdKWSQZibFPLOw.jpg

OMG, 3DMARK03

http://www.chiphell.com/attachments/month_0704/20070429_794341e05ece44261700h3X8YMnP0Nqf.jpg

Any chance we could get some working links? Those are just red X's :sad:

[INKster]

1997. There are two representations of 90 in roman numerals, LXXXX and XC.

Anyway, yes, back on topic I'm still finding this speculation about ATI engineering in performant features that won't get used in the generation rather problematic. I find it *really* tough to accept that that it is good idea.

Here we don't use the quad-X representation. There's probably no language closer to the original latin than our own (except maybe for Italian).

-----------------------

And to complement your line of thought, ATI would be breaking somewhat of a tradition to simply stick to the spec and enhance performance based solely on it (R300, R520, etc).

[trinibwoy]

I guess MSAA+MRT, custom AA and Tesseletor are beyond DX10. MSAA+MRT surely is one of the spec of dx10.1, you can see dx10.1 ppt from MS.

MSAA on MRT's will be a coup for sure, but where's the API support?

[mao5]

MSAA on MRT's will be a coup for sure, but where's the API support?

it doesn't exist in DX10, but I think it will work through some kind of IHV game patch.

[mao5]

some vendor in china is ready for this:

http://www.chiphell.com/attachments/month_0704/20070429_e6e9c8d4265913b99ea9EFruDQGlb4b8.jpg

[russo121]

MCMLXXXXVII ? I don't even know what that means...
1987 is MCMLXXXVII.
1988 is MCMLXXXVIII.

:D

Going back to topic..., DX10.1 ? I had no idea MS had that spec already set in stone.
Just give us Direct3D 10.0 software first. ;)

OT: It is possible to write both ways: http://labienus.home.texas.net/RomanOutpost/Antiquities/romanmath.txt

[mao5]

3DMARK06 XT E6600 2.4g

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

3DMARK06 GTX E660 2.8g

http://www.chiphell.com/attachments/month_0704/20070429_b6479c89355e3dea6764HWJwJMDpVk3X.jpg

3DMARK06 XT from DT X6800: 14005

[Kaotik]

Any chance we could get some working links? Those are just red X's :sad:

In case you still see just red x's:
3DMark2001SE - 61138 3DMarks
3DMark03 - 37990 3DMarks

[jamis]

Nope, all I remember is people claiming R600 is DX10.1 and other people saying that there's no such thing.
It does exist, allthough R600 supporting it seems unlikely. On the other hand DX10.1 requires things like 32-bit floating-point filtering and higher precision so it could explain why the damn thing is so huge being only 64-16-16 chip.

[mao5]

It does exist, allthough R600 supporting it seems unlikely. On the other hand DX10.1 requires things like 32-bit floating-point filtering and higher precision so it could explain why the damn thing is so huge being only 64-16-16 chip.

not unlikely, I say surely

[Sound_Card]

Ok, I'm still looking for the 10.1 slides but I did run across something very very intresting from the past.

http://i21.photobucket.com/albums/b299/Genocide737/zplaa92921teqi4kd2.jpg

1080p, H.264, VC-1, HDDVD/blu-ray support
Intergrated HDCP, HDMI and HD audio
over 80gb of bandwith

Pretty correct right? Back then(Novemeber), people were very doubtful of those feaures. But here is something that is a little more intresting...

"Massive geometry performance":razz:


Let me bring this slide out as well.

http://i21.photobucket.com/albums/b299/Genocide737/u3t9pvpwxd10fa9am7.jpg

Everything looks normal and dandy right? Untill you get to the last bullet...

"new texture cache disign":shock:

Given the latest leak of R600 inards, We see L1 and L2 cache right on the TMU pipelines, something that was not in ATi hardware before. (Danm good catch Jawed).

Also, "major z/stencil arc improvement" seems rather intresting as well. I take that has a hint towards R600's ROP's having much more beef than R5xx. Double pumped?

[nutball]

so it could explain why the damn thing is so huge being only 64-16-16 chip.

Seems to me that the yellow box on the left of the SPUs in this piccie (http://forum.beyond3d.com/showpost.php?p=977879&postcount=3367) could also explain why it's so big. Where it says "memory read/write cache", substitute "stuffing great memory read/write cache". If you wanted to rule the world at GPGPU seems to me that more stuffing and more great in this area might do it for you.

I'd hazard a guess that ATI would earn more dollars from R600 being a king of GPGPU than it being a king of DX10.1 at this point in time.

[Cuthalu]

3DMARK06 XT E6600 2.4g

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

3DMARK06 GTX E660 2.8g

http://www.chiphell.com/attachments/month_0704/20070429_b6479c89355e3dea6764HWJwJMDpVk3X.jpg

3DMARK06 XT from DT X6800: 14005

What's with the lower CPU-score?

[Geeforcer]

3DMARK06 XT E6600 2.4g

CPU score: 3092

3DMARK06 GTX E660 2.8g

CPU score: 2440


:roll:

[NocturnDragon]

Here we don't use the quad-X representation. There's probably no language closer to the original latin than our own (except maybe for Italian).



OT: i'm italian and we don't use the quad X representation either.

IT: After reading pages and pages I didn't get it. Will the NDA end on the 2nd or the 14th?

[AnarchX]

The detailed results of 3DMark01 would be more interessting, espacially mt-fillrate and high poly count. :wink:

@D3D10.1: Afaik this needs WDDM2.1, so I doubt that R600 will support it.

[mao5]

:roll:

no idea

[mao5]

The testing xt card is stable oc at 890MHz, core speed. 20% up

[Nebula]

3DMARK06 XT E6600 2.4g

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

3DMARK06 GTX E660 2.8g

http://www.chiphell.com/attachments/month_0704/20070429_b6479c89355e3dea6764HWJwJMDpVk3X.jpg

3DMARK06 XT from DT X6800: 14005

Umm how can the '3DMARK06 XT E6600 2.4g' get higher CPU score than the '3DMARK06 GTX E660 2.8g'. The E6600 2.4Ghz scores 3092 but the E6600 2.8Ghz scores 2440?

Did I miss something?

[Fox5]

1997. There are two representations of 90 in roman numerals, LXXXX and XC.

Anyway, yes, back on topic I'm still finding this speculation about ATI engineering in performant features that won't get used in the generation rather problematic. I find it *really* tough to accept that that it is good idea.

ATI did it in every generation except R300 and R420 gens.

Model 3 (not ati, but the design team went to ati, right?) hardware was overdesigned.
Radeon had useless anisotropic filtering and the 1-3 pixel-texel ratio.
Radeon 8500 has ps1.4.
R520/580 had good dynamic branching performance.

And they lost in sales every time, so probably not a good idea when nvidia gets by supporting features at low performance or not at all and does just fine.

[mao5]

Umm how can the '3DMARK06 XT E6600 2.4g' get higher CPU score than the '3DMARK06 GTX E660 2.8g'. The E6600 2.4Ghz scores 3092 but the E6600 2.8Ghz scores 2440?

Did I miss something?

I'm investigating it. pls be patient

[Skinner]

I'm investigating it. pls be patient


Maybe you mixed the scores up?
edit : no, that's not it sm2/3 scores are up.

BTw, any chance to fire up Stalker?

[Cuthalu]

Mao, what about power consumption, temperatures and noise?

[Jawed]

Ok, I'm still looking for the 10.1 slides but I did run across something very very intresting from the past.
At the time those appeared I said they had a fair likelihood of being real...


"new texture cache disign":shock:

Given the latest leak of R600 inards, We see L1 and L2 cache right on the TMU pipelines, something that was not in ATi hardware before. (Danm good catch Jawed).
Well, I try...

Also, "major z/stencil arc improvement" seems rather intresting as well. I take that has a hint towards R600's ROP's having much more beef than R5xx. Double pumped?
This function point was something that I specifically identified as making these slides likely to be real - there was a patent application I linked to back this up. Stuff that would require a faker to be ***really*** paying attention.

This last point actually indicates a whole suite of patents on the subject. Basically faster early-Z culling and faster rasterisation all combined. In addition, R600's ROPs (hey, maybe we can start calling them RBEs) have improved compression algorithms to deal with floating point render targets.

Jawed

[BlizzardOne]

Umm how can the '3DMARK06 XT E6600 2.4g' get higher CPU score than the '3DMARK06 GTX E660 2.8g'. The E6600 2.4Ghz scores 3092 but the E6600 2.8Ghz scores 2440?

Did I miss something?

maybe wrong way round...??maybe the GTX is paired with the 2.4ghz rather than 2.8 :razz:

[Robin B]

Is it just me, or does the XT need a lot of Viagra to get it up and beyond .....:lol:

[Jawed]

Seems to me that the yellow box on the left of the SPUs in this piccie (http://forum.beyond3d.com/showpost.php?p=977879&postcount=3367) could also explain why it's so big. Where it says "memory read/write cache", substitute "stuffing great memory read/write cache". If you wanted to rule the world at GPGPU seems to me that more stuffing and more great in this area might do it for you.
The question is, is it big enough to make a real difference?

After all, it might be relatively small, something in the region of the caches used for RBEs. Still, it's fun to ogle this stuff...

Jawed

[Sound_Card]

This function point was something that I specifically identified as making these slides likely to be real - there was a patent application I linked to back this up. Stuff that would require a faker to be ***really*** paying attention.

This last point actually indicates a whole suite of patents on the subject. Basically faster early-Z culling and faster rasterisation all combined. In addition, R600's ROPs (hey, maybe we can start calling them RBEs) have improved compression algorithms to deal with floating point render targets.

Jawed

What is a RBE?

[Jawed]

What is a RBE?
Render Back-end.

Jawed

[bigtabs]

Render Back End.

..and no.. I don't know what that means.

[BlizzardOne]

ok, that e6600 @ 2.4 is definately not at 2.4ghz.

my X6800 @ 2.93 (i.e stock) comes in at around 2600 on the CPU score...

[Jawed]

Render Back End.

..and no.. I don't know what that means.
http://forum.beyond3d.com/showpost.php?p=977879&postcount=3367

http://bbs.expreview.com/attachments/month_0704/20070428_deff9681f706dcab9a16HwlraKhteBqZ.jpg

Just another name for ROPs.

Jawed

[Sound_Card]

Render Back-end.

Jawed

Ahh thanks....

[neliz]

Finally.. remember Geo's "6800GT wears red?" signature?
it all makes perfect sense now.. I only wonder if ATI wil get the same respect for the XT as nV got for the GT.
So ATI didn't postpone the launch for that reason... we can make up some new theories now...

[fellix]

Render Back End.

..and no.. I don't know what that means.
The part of the GPU pipeline, where your screen pixels are being baked of and ready for your viewing pleasure. :lol:

[trinibwoy]

I only wonder if ATI wil get the same respect for the XT as nV got for the GT.

Probably not - it's all about the timing. You don't get props for coming in below expectations.

[Geeforcer]

Finally.. remember Geo's "6800GT wears red?" signature?
it all makes perfect sense now.. I only wonder if ATI wil get the same respect for the XT as nV got for the GT.
So ATI didn't postpone the launch for that reason... we can make up some new theories now...

IDK.. would 6800 GT gotten the same level of respect if it was the fasterst 6800 card nvidia had?

[INKster]

IDK.. would 6800 GT gotten the same level of respect if it was the fasterst 6800 card nvidia had?

Probably.
The actual performance difference between a 6800 GT and a 6800 Ultra is really very small.
However, the 8800 GTS is significantly slower than the 8800 GTX. The gap is widening.

[bigtabs]

IDK.. would 6800 GT gotten the same level of respect if it was the fasterst 6800 card nvidia had?

Isn't the more pertinent question whether the 6800GT would've had the same level of respect if ATI had a widely available card out 7 months prior that demolished it; even if it was priced higher?

[mao5]

sorry let you guys waiting so long, the answer is:


11603, R600XT 745/828, C2D E6600 @ 3.6GHz

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

10855, 8800GTX 575/900, C2D E6600 @ 3.1GHz

http://www.chiphell.com/attachments/month_0704/20070429_4cfd5037ac13924be2d3Z31QAnNY5MRx.jpg

XT AND GTX ARR ALL AT THEIRE DEFAULT CLOCK

[mao5]

NFS CARON 1280x1024, MAX Detail

http://www.chiphell.com/attachments/month_0704/20070429_8931abca6d850246e14fFFnaxxaGdX9T.jpg

[bigtabs]

God damn that game sucks. :twisted:

What's it like in Test Drive Unlimited?

:lol:

[Brian118]

WTH!? Now the XT is doing better than a GTX?!

[mao5]

God damn that game sucks. :twisted:

What's it like in Test Drive Unlimited?

:lol:

default clock,man, R600XT 745/828, anyone can provide default clock GTX NFS carbon score at 1280x1024 MAX detail?

[BlizzardOne]

WTH!? Now the XT is doing better than a GTX?!

well, after such a long wait, i would bloody well hope so :p

[mao5]

WTH!? Now the XT is doing better than a GTX?!

default xt is doing better than default GTX in 01,03,05,06, but I don't know extreme oc result.

[Plano]

WTH!? Now the XT is doing better than a GTX?!

Well in that test 2900XT is running on faster CPU.

[Brian118]

But I was getting the impression that it was barely keeping up with the GTS...now this? Someone is going to be wrong and I hope it is not the chinese :)

[mao5]

Well in that test 2900XT is running on faster CPU.

yes, but would you pls compared SM2.0 and SM3.0 scores? thank you.

[mao5]

But I was getting the impression that it was barely keeping up with the GTS...now this? Someone is going to be wrong and I hope it is not the chinese :)

All you guys here put dozen pages of suspicion on dailytech, oh, chinese must be the boss of Dailytech.http://69.93.88.162/forum/images/smiles/rofl.gifhttp://69.93.88.162/forum/images/smiles/rofl.gifhttp://69.93.88.162/forum/images/smiles/rofl.gif

[BlizzardOne]

yes, but would you pls compared SM2.0 and SM3.0 scores? thank you.

assuming CPU speed has very little effect (i.e none)on the SM 2.0 and 3.0 scores;
the R600 XT scores very similarly as CF'd XTX's:

3DMark Score 11129 3DMarks

SM 2.0 Score 4785 Marks (approx. -50 vs R600)

SM 3.0 Score 5445 Marks (approx. -100 vs R600)

CPU Score 2566 Marks

that's with a X6800 @ stock

[Razor1]

default clock,man, R600XT 745/828, anyone can provide default clock GTX NFS carbon score at 1280x1024 MAX detail?


75-80, I would like to know what drivers they used for those 3dmark scores, cause I'm getting around that GTX score right now and my CPU isn't overclocked.

[INKster]

WTH!? Now the XT is doing better than a GTX?!

Look at the CPU speed.
The GTX is only using a Core 2 at 3.1GHz, versus 3.6GHz for the R600 XT.
Not exactly fair game...

[mao5]

assuming CPU speed has very little effect (i.e none)on the SM 2.0 and 3.0 scores;
the R600 XT scores very similarly as CF'd XTX's:

3DMark Score 11129 3DMarks

SM 2.0 Score 4785 Marks (approx. -50 vs R600)

SM 3.0 Score 5445 Marks (approx. -100 vs R600)

CPU Score 2566 Marks

that's with a X6800 @ stock

so dault GTX scores also similaryly as CF'd XTX?

[BlizzardOne]

so dault GTX scores also similaryly as CF'd XTX?

SM2.0, yes.. SM3.0, about 400-500 points behind..

[mao5]

75-80, I would like to know what drivers they used for those 3dmark scores, cause I'm getting around that GTX score right now and my CPU isn't overclocked.

sorry the tester is in sleep now, I'' ll ask him 8 hours later.

[mao5]

SM2.0, yes.. SM3.0, about 400-500 points behind..

CF XTX? what's XTX? R580 XTX? R600 XTX? Where did you get the scores? Weird.

[BlizzardOne]

CF XTX? what's XTX? R580 XTX? R600 XTX? Where did you get the scores? Weird.

CF'd X1950 XTX's + C2D X6800, it's my ststem i'm on atm :) all stock http://service.futuremark.com/compare?3dm06=1697955

[mao5]

CF'd X1950 XTX's + C2D X6800, it's my ststem i'm on atm :) all stock http://service.futuremark.com/compare?3dm06=1697955

so this means you don't need to upgrade, unless you want a dx10 patch of Call of Jauze.

[BlizzardOne]

so this means you don't need to upgrade, unless you want a dx10 patch of Call of Jauze.

pretty much, power consumption, if rumors of ~225-230w are correct, put it on par with CF'd X1950's.. for roughly the same performance. so either way, in DX 9, no real difference.. unless AA/AF hit is significantly reduced, at significantly improved quality!

i'm actally rather curious as to whether replacing my dual X1950's with an R600 XT (X?) is worth it... only a few more days i guess :)

[mao5]

pretty much, power consumption, if rumors of ~225-230w are correct, put it on par with CF'd X1950's.. for roughly the same performance. so either way, in DX 9, no real difference.. unless AA/AF hit is significantly reduced, at significantly improved quality!

i'm actally rather curious as to whether replacing my dual X1950's with an R600 XT (X?) is worth it... only a few more days i guess :)

worth it

[Unknown Soldier]

Now mao5, how will my QX6600 perform?? ;)

US

[mao5]

Now mao5, how will my QX6600 perform?? ;)

US

QX6800 in DT got 14005, What do you think?

[BlizzardOne]

worth it

:cry:

my bank account is going to curse me for an eternity

[neliz]

:cry:

my bank account is going to curse me for an eternity

I have to chose between a months rent and a XT.. trustworthy people tell me... XT...

[BlizzardOne]

I have to chose between a months rent and a XT.. trustworthy people tell me... XT...

thank god i don't have rent to pay :p

[Nebula]

sorry let you guys waiting so long, the answer is:


11603, R600XT 745/828, C2D E6600 @ 3.6GHz

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

10855, 8800GTX 575/900, C2D E6600 @ 3.1GHz

http://www.chiphell.com/attachments/month_0704/20070429_4cfd5037ac13924be2d3Z31QAnNY5MRx.jpg

XT AND GTX ARR ALL AT THEIRE DEFAULT CLOCK

WTF?
You posted GTX scores on the previous page. The perfomance is now lower on the GTX (SM2.0/SM3.0)?

http://www.chiphell.com/attachments/month_0704/20070429_b6479c89355e3dea6764HWJwJMDpVk3X.jpg

[bigtabs]

Aren't these all from different users of a forum though?

[Silent_Buddha]

WTF?
You posted GTX scores on the previous page. The perfomance is now lower on the GTX (SM2.0/SM3.0)?

http://www.chiphell.com/attachments/month_0704/20070429_b6479c89355e3dea6764HWJwJMDpVk3X.jpg

Look at the SM 2.0 and SM 3.0 scores. The original screenshot was of an overclocked GTX.

The second set used a stock GTX to compare to a stock HD 2900 XT.

In other words, at Stock. HD 2900 XT has similar SM 2.0 perf in 3dm06. But has superior SM 3.0 perf.

While a stock 2900 XT vs a O/C'd GTX has lower SM 2.0 perf. And slightly lower SM 3.0 perf.

Mao mentions the guy can easily O/C the card to 890 core. I'd like to see a comparison of that to a O/C'd GTX.

Granted 3dm06 doesn't tell us a whole lot about real game play, but it at least gives us something to look at. :D

Regards,
SB

[Nebula]

Look at the SM 2.0 and SM 3.0 scores. The original screenshot was of an overclocked GTX.

The second set used a stock GTX to compare to a stock HD 2900 XT.

Regards,
SB

Hmm ,OK but he didn't state that it was overclocked in the previous post where he posted both benchmark results.

[nAo]

X2900XT@gpuinstr

MAD1 - MAD 1 512 256 100 144.9562 ms 46.2960 GInstr/sec
MAD3 - MAD 3 512 256 100 142.4918 ms 47.0966 GInstr/sec
MAD4 - MAD 4 512 256 100 142.5442 ms 47.0793 GInstr/sec

MUL1 - MUL 1 512 256 100 71.2619 ms 94.1722 GInstr/sec
MUL3 - MUL 3 512 256 100 71.8172 ms 93.4440 GInstr/sec
MUL4 - MUL 4 512 256 100 88.9417 ms 75.4526 GInstr/sec

ADD1 - ADD 1 512 256 100 36.4790 ms 183.9657 GInstr/sec
ADD3 - ADD 3 512 256 100 43.6579 ms 153.7152 GInstr/sec
ADD4 - ADD 4 512 256 100 57.4707 ms 116.7706 GInstr/sec

DP3 - DP3 1 512 256 100 142.6667 ms 47.0389 GInstr/sec
DP3 - DP3 3 512 256 100 141.9595 ms 47.2732 GInstr/sec
DP3 - DP3 4 512 256 100 143.7289 ms 46.6913 GInstr/sec

DP4 - DP4 1 512 256 100 143.0978 ms 46.8972 GInstr/sec
DP4 - DP4 3 512 256 100 142.7959 ms 46.9963 GInstr/sec
DP4 - DP4 4 512 256 100 141.9480 ms 47.2771 GInstr/sec

Maybe more later. Have fun :)
I would expect those numbers from a non scalar architecture

[Silent_Buddha]

Hmm ,OK but he didn't state that it was overclocked in the previous post where he posted both benchmark results.

Just went back a couple pages to check, and no mention was made of the clocks of either graphics card. Only the supposed clocks of the CPUs.

Regards,
SB

[mao5]

WTF?
You posted GTX scores on the previous page. The perfomance is now lower on the GTX (SM2.0/SM3.0)?

http://www.chiphell.com/attachments/month_0704/20070429_b6479c89355e3dea6764HWJwJMDpVk3X.jpg

That's a extreme oc GTX provided by someone in our forum, he is not the guy who test R600XT. That guy won't mention the extreme oc speed of GTX

[dnavas]

Nice picture Inkster, pity about the resolution.

I kept staring at "lite potatoes" trying to figure out what they were talking about.
Finally figured out it meant interpolaters :)
That slide will be far more interesting with a description of all the boxes.

There's a bright yellow programmable tessellator :?:

I wonder what it can be called on to do?

Also, if you look at the ALUs, there's one big yellow unit and 4 smaller yellow units :?:

Well, I'm thinking special functions.... That's definitely not symmetric 320 shaders, is it?

Both the texture units and the render back-ends have 4-way blocks and 16-way blocks :?:

Are they breaking out color channels?

I think the most interesting thing, though, is still the balance between math and texture. It's a bit of a surprise, really, as I thought 3:1 was pretty math heavy.... What in heaven's name are they using their bandwidth for?

-Dave

[mao5]

I kept staring at "lite potatoes" trying to figure out what they were talking about.
Finally figured out it meant interpolaters :)
That slide will be far more interesting with a description of all the boxes.



I wonder what it can be called on to do?



Well, I'm thinking special functions.... That's definitely not symmetric 320 shaders, is it?



Are they breaking out color channels?

I think the most interesting thing, though, is still the balance between math and texture. It's a bit of a surprise, really, as I thought 3:1 was pretty math heavy.... What in heaven's name are they using their bandwidth for?

-Dave

All I can tell you the pic is true, no revise only to make word blur, that's all

[Galduta]

default clock,man, R600XT 745/828, anyone can provide default clock GTX NFS carbon score at 1280x1024 MAX detail?

http://elblogdelosjuegosmalditos.blogspot.com/2007/01/need-for-speed-carbon-demo.html

In this GPU or in 8800 gtx , NFS carbon have a CPU limitation - this score is with one C2 3200 MHZ -

[Galduta]

Edit

[trinibwoy]

sorry the tester is in sleep now, I'' ll ask him 8 hours later.

Ask him what driver he used too.

[neliz]

Ask him what driver he used too.

8.37 is still worked on as a "review/benchmark" driver as far as I know, it certainly won't be the driver that ships with the boards though

[Fox5]

Seems to me that the yellow box on the left of the SPUs in this piccie (http://forum.beyond3d.com/showpost.php?p=977879&postcount=3367) could also explain why it's so big. Where it says "memory read/write cache", substitute "stuffing great memory read/write cache". If you wanted to rule the world at GPGPU seems to me that more stuffing and more great in this area might do it for you.

I'd hazard a guess that ATI would earn more dollars from R600 being a king of GPGPU than it being a king of DX10.1 at this point in time.

What would it take to do that? Full IEEE 32 bit compliancy?

[trinibwoy]

8.37 is still worked on as a "review/benchmark" driver as far as I know, it certainly won't be the driver that ships with the boards though

Ok, it seems DT used 8.631 - so maybe we still don't know anything about R600 performance :) But I'm still skeptical that ATI is going to sell a world beater for $400.

[Kaotik]

QX6800 in DT got 14005, What do you think?

The QX6800 couldn't have been at it's default clocks, GTX does not score 14k+ with that, if you look at ORB, all the C2Q's getting 14k+ numbers are OC'd to 3.6GHz+

[CJ]

Ok, it seems DT used 8.631 - so maybe we still don't know anything about R600 performance :)

I wouldn't pay too much attention to the scores on DT... all they've done is create a lot of confusion and panic.

[Razor1]

The QX6800 couldn't have been at it's default clocks, GTX does not score 14k+ with that, if you look at ORB, all the C2Q's getting 14k+ numbers are OC'd to 3.6GHz+

its possible there is one Orb result with 14+k with Quad c2d don't know if they are SLi, but looking at the memory size of the graphics card 512 mb in the details, I think its a single card.

http://service.futuremark.com/compare?3dm06=1709802

And this is with a 2.67 ghz C2D quad, but the graphics card has to be overclocked

[Kaotik]

its possible there is one Orb result with 14+k with Quad c2d don't know if they are SLi, but looking at the memory size of the graphics card 512 mb in the details, I think its a single card.

http://service.futuremark.com/compare?3dm06=1709802

And this is with a 2.67 ghz C2D quad, but the graphics card has to be overclocked

Yeah, and has to be OC'd a lot, I mean if you could get it with C2Q @ 2.67GHz by normal means (and with that old driver set, too), ORB would be filled with those results for sure

[Razor1]

CJ we got dailytech's numbers, supposed leaked AMD slides numbers, a guy on Tom's Hardware talking about dual card AMD numbers vs. a single GTS (weird that this leaked slide happened to be this :wink: ), faud tell the world AMD stating the r600 won't compete with the GTX,

And then you saying everyone is wrong....

Whats going on :lol: :runaway: :?: :?: :?:

[Mize]

Okay, time for a new "LAST R600 rumors" thread with a nice synopsis of all the leaks and FUD at the top for those of us who haven't had time to refresh this thread every five minutes! :) :)

[Twinkie]

CJ we got dailytech's numbers, supposed leaked AMD slides numbers, a guy on Tom's Hardware talking about dual card AMD numbers vs. a single GTS (weird that this leaked slide happened to be this :wink: ), faud tell the world AMD stating the r600 won't compete with the GTX,

And then you saying everyone is wrong....

Whats going on :lol: :runaway: :?: :?: :?:

QFT. Kombatant, gibbo, etc etc are saying that some of the rumours/leaked benches are not true..

But is that in a good way or a bad way?

This is very confusing indeed.

[Razor1]

QFT. Kombatant, gibbo, etc etc are saying that some of the rumours/leaked benches are not true..

But is that in a good way or a bad way?

This is very confusing indeed.


I hope its in a good way!