View Full Version : Is Nvidia making a strategic mistake with NV30?
siconik
13-May-2002, 14:02
In the other thread where we were discussing Nvidia's apparent unwillingness to implement 256bit bus as well as their focus on "better", rather then "faster" pixels. IMO such strategy is very risky, considering that consumer is very benchmark-oriented. Take a look at this latest review (http://www.anandtech.com/video/showdoc.html?i=1619) on Anandtech.
Many people regard IQ above everything else, yet despite fact that performance as well as quality of FSAA and advanced texture filtering is one of the key selling points of today’s cards the review doesn't even motion them, let alone provide appropriate charts and screenshots. Just like 2 years ago, today major websites like Anandtech focus on straight benchmarking while completely ignoring IQ, and there is no indication that this trend will change in the near future. In light of this, could Nvidia be making the same mistake that 3dfx made before them by putting most of their eggs in the wrong (IQ) basket?
Foodman
13-May-2002, 14:19
I don't think they are.
FSAA is the new emphasis, and cards are often benchmarked with that. Anyone with a GeForce 3 class part probably runs a good number of games with 2x on.
3dfx simply didn't have the fillrate to pull it off. Its there in oodles now, and if the new cores don't do rendering without AA very well, I don't think its such a big deal.
I mean, honestly, who gives a shit about 16 bit rendering anymore? With your theory, we'd be screaming bloody murder because the 16 bit performance isn't awesome with the GeForce 4.
siconik
13-May-2002, 14:45
Funny you should mention it: some people were screaming bloody murder when Geforce 3 failed to outperform Geforce 2 Ultra in some of the low-resolution/16bit tests. However, that's not the point. Since 1998, almost every review included 32-bit tests. However, despite the fact that usable FSAA has been around in consumer cards since 2000, big sites like Anand do not bother to include benchmarks and screenshots with FSAA on; anisotropic filtering is not faring any better. Let's face it, as much as I'd like to think that most people base their decisions on B3D-caliber articles, I am afraid that a single Q3 chart on Tom's or Anand’s sways more people then Reverand's painstaking work.
Ascended Saiyan
13-May-2002, 15:24
I don't think they are.I have to agree there,what I do think is that it all depends on who is doing the benchmarking & if they'are doing it properly.For some reason Nvidia reminds me of 3dfx when they say that that sort of feature isn't needed.
Joe DeFuria
13-May-2002, 15:25
Whether or not nVidia is making a mistke depdens, IMO, on what nVidia means by "higher quality" pixels.
If nvidia means something like "By using DX9 programming techniques, one can create a better quality pixel with the same performance as a DX8 technique getting similar results", then I think nVidia is making a mistake. They would have to rely on tech demos, rather than mainstream benchmarks, to demonstrate their point.
If, on the other hand, nVidia implements some AA and/or filtertering tehnique that has a new level of performance with good quality, then I think nVidia is OK.
Yes, I think you are right. Nvidia is walking on thin ice by pushing quality over performance. On one hand it's good that they are pushing the boundaries of real time graphic capabilities. On the other hand, realistically we have all been burned by getting cards with wiz-bang new features that aren't widely supported for years to come. (example: when do you think direct x 9 games will really be out there in abundance?) Even when the new features do get supported in games, it's likely that that card will be considered slow by that time. You just can't win.
I have to disagree. I think a company like nvidia "drives" the market. I think the focus on performance now is because nvidia heavily invested marketing efforts in T&L which was after all a feature to increase performance drastically.
They know more than anyone else that image quality is underperforming these days, and as it gets more and more difficult to make huge performance increases it seems also quite logical that nvidia and other graphics companies invest their time and money in enhancing image quality!
Jurgen
Livecoma
13-May-2002, 15:54
Hmm I'll wait for NV30 to show up before I say if It was a mistake or not... Taking a look at NVIDIA's sales and growth I think It's safe to say they have maintained a trend of delivering what the majority wants, and I don't see anything now that would make me think otherwise.
I'm definatly sure they are not ignorant to what's cooking at Matrox, 3dlabs, ATi, etc...
Joe DeFuria
13-May-2002, 16:10
Of course, we'll all wait for it, but that won't stop us from speculating. ;)
and I don't see anything now that would make me think otherwise.
I do, lack of a 256 bit bus. ;) (We don't know this for sure of course, but it certainly appears that way.)
I'm definatly sure they are not ignorant to what's cooking at Matrox, 3dlabs, ATi, etc...
Don't know what difference that makes...I could say I'm definitely sure that Matrox, 3dlabs, and ATI knows what's cooking at nVidia. In fact, I would say that Matrox and 3D Labs have the advantage in the secrecy department, because they do not sell chips to 3rd pary IHVs. Less opportunity for leaks....
A couple of interesting and relevant news snippets today:
http://www.digitimes.com/NewsShow/Article.asp?datePublish=2002/05/13&pages=06&seq=38
According to the companies’ roadmaps, both Nvidia and ATI said that they will showcase NV18 and RV250-based chips at Computex Taipei early next month, and the chips will replace the current GeForce4 MX440 and Radeon 7500 as the companies’ mainstream products. Their top-end products, the DirectX 9-supporting NV30 and R300 graphics cores, will not be launched until the end of the third quarter at the earliest.
http://www.anandtech.com/news/shownews.html?i=3414&t=sn
I must say that right now the video card sector is by far the most interesting of them all in this industry. Matrox has a pretty big announcement tomorrow that I'll be covering, 3DLabs just recently announced their powerful VPU, NV30 is amazing, and Matthew and I will be up to see ATI later this week.
It's looking more and more like the 0.13 micron parts (both from ATI and nVidia) will end up shipping later rahter than sooner (at least in any reasonable quantity.) Matrox's and 3D Labs decision to base their first gen on 0.15 microns could end up paying off significantly in terms of time-to market. That will not only help with initial sales, but with driver development.
JF_Aidan_Pryde
13-May-2002, 16:14
I think NVIDIA understands this more than you can imagine.
They know speed always comes before quality in real time 3D.
It is reasonable to assume they will have speed and quality for the NV30.
Livecoma
13-May-2002, 16:38
I do, lack of a 256 bit bus. (We don't know this for sure of course, but it certainly appears that way.)
I assumed you knew this already, but there is quite a bit more to a graphics card then a 256bit bus. :wink:
Hellbinder
13-May-2002, 16:41
1c4i4,
I disagree. 3dFX is the company that strarted the internet benchmark war. they only lost when they decided to go for Quality instead of quantity.
How is that scinario ANY different from what Nvidia is trying now? the idea the Nvidia drives consumer perception is just false. They simply follow the rules that 3dfx created long ago. You break the rules.. you go down in flames. Its that simple. Anyone who thinks otherwise is out of touch with reality.
R300 is all about speed, plain and simple. If Nvidia launches a product with supposed "better pixels" and the R300 kicks their ass in the benchmarks... What do you think will happen? What if the GF5 is not even as fast, or equal to the Parhelia-512? then what...
there simply is no brand loyalt that extreme... 3DFX OWNED the planet, It was theirs to lose.... and they lost it...
In the end its all about speed and you know it.
Livecoma
13-May-2002, 16:45
]1c4i4,
I disagree. 3dFX is the company that strarted the internet benchmark war. they only lost when they decided to go for Quality instead of quantity.
How is that scinario ANY different from what Nvidia is trying now? the idea the Nvidia drives consumer perception is just false. They simply follow the rules that 3dfx created long ago. You break the rules.. you go down in flames. Its that simple. Anyone who thinks otherwise is out of touch with reality.
R300 is all about speed, plain and simple. If Nvidia launches a product with supposed "better pixels" and the R300 kicks their ass in the benchmarks... What do you think will happen? What if the GF5 is not even as fast, or equal to the Parhelia-512? then what...
there simply is no brand loyalt that extreme... 3DFX OWNED the planet, It was theirs to lose.... and they lost it...
In the end its all about speed and you know it.
Dude how do you know any of this stuff? It seriously sounds like your talking out of your buttocks. Performance and Image Quality go hand and hand.
Hellbinder
13-May-2002, 16:49
dude,
Do you live on the same planet as me? I wonder... I am a Radeon 8500 owner.. Trust me.. its not about IQ. not to the general public, Not at all. Otherwise the 8500 would be crowned king everywhere.
I simply cant believe that you Nvidians can turn around and be such hipocrites so fast.
Give me a break.
Furthermore, How do you suppose that the Gf5 could deliver a "better pixel" than the Pahelia?
Joe DeFuria
13-May-2002, 16:55
Livecoma,
Yes, I know there is more to a graphics card than a 256 bus. ;) However, you stated that you saw "no reason" to expect that nVidia isn't going to deliver what "everyone wants." Everyone wants speed. And "all else being equal", 256 bits is going to be much faster than 128 bits. So, that is "my reason" to think that nVidia might not deliver what everyone wants.
Please, no one bother to respond and say "but all else isn't equal...nVidia has some demonstrated some effectiveness of their bandwidth saving techniques and memory controller, etc. I know this. ;)
This talk about quality is market smoke.
Matrox is not only aparentlly offering speed with 256bits, it is offering quality with:
- High quality planar response video filters
- high quality 10bits R, G and B
- nice filtering
- nice 16x AA
- all with great memory bandwith
Good to see Matrox coming back and I hope to see some 256bits DDR .13 micron Parhelia in Q1/2003 (maybe full DX9).
Remenber sustained high quality needs lots of bandwith. Is not one in the detriment of the other.
DX9 is as important now as was T&L when GF1 was released (1999?):
- Will be good for the game community in the long term creating a better common denominator.
- Will not be good for the individual gamer waiting 2 or 3 years to see some nice effect with DX9. Maybe you will be able to see some nice DX9 3DMark2003.
The question is do you want to pay now to create the future common denominator or want to pay to play a nice game now?
Ichneumon
13-May-2002, 18:21
There is one important thing about having a DX9 card out first, or one of the first...
That is that whoever does that, and does it best, will be the card that the next generation of games will primarily be developed and tested on.
Cards always run games better when they're in that environment... the Dx api might make a game able to run on other comparable hardware when it hasn't been tested much by the developer, but those cards will never thrive.
That is a big incentive for companies to get Dx9 cards out as soon as they can once it is released.
Ascended Saiyan
13-May-2002, 19:16
]1c4i4,
I disagree. 3dFX is the company that strarted the internet benchmark war. they only lost when they decided to go for Quality instead of quantity.
How is that scinario ANY different from what Nvidia is trying now? the idea the Nvidia drives consumer perception is just false. They simply follow the rules that 3dfx created long ago. You break the rules.. you go down in flames. Its that simple. Anyone who thinks otherwise is out of touch with reality.
R300 is all about speed, plain and simple. If Nvidia launches a product with supposed "better pixels" and the R300 kicks their ass in the benchmarks... What do you think will happen? What if the GF5 is not even as fast, or equal to the Parhelia-512? then what...
there simply is no brand loyalt that extreme... 3DFX OWNED the planet, It was theirs to lose.... and they lost it...
In the end its all about speed and you know it.
Dude how do you know any of this stuff? It seriously sounds like your talking out of your buttocks. Performance and Image Quality go hand and hand.I agree with you statement but I think it could have been phrased a little better.Hellbinder,I think what Livecoma is getting at is that in this day & age both are considered to be equal.Image quality cannot do well in games unless there is enough FPS for it to be maintained.FPS is pointless if the game looks horrible with smudgy textures or what have you.So it would make perfect sense that you have both because neither can do without the other.
Foodman
13-May-2002, 19:22
Funny you should mention it: some people were screaming bloody murder when Geforce 3 failed to outperform Geforce 2 Ultra in some of the low-resolution/16bit tests. However, that's not the point. Since 1998, almost every review included 32-bit tests.
And since the tail end of the GeForce 2 era, most everyone includes FSAA benchmarks. Who still uses 16 bit?
See, you missed my point. nVidia knows that people are benchmarking FSAA, and they've de-emphasised non FSAA/Low res scores. No one tests at 640x480 anymore.
They've decided people are paying enough attention to it (FSAA) now that they figure they can get away with it. They might be banking on this one generation too soon. Perhaps they should have done this at NV35.
Livecoma,
And "all else being equal", 256 bits is going to be much faster than 128 bits. Please, no one bother to respond and say "but all else isn't equal...nVidia has some demonstrated some effectiveness of their bandwidth saving techniques and memory controller, etc. I know this. ;)
I can't help it, but i have to respond :-)
Isnt that the problem then, the all else being equal.
If the NV30 or RV300 or whatever it might be with a 128 bit bus is faster then a parhelia with a 256 bit bus and also cheaper (all of this pure speculations of course), then who cares if it doesn't have a 256 bit bus or not ?
What i'm a little bit confused by though, is that none of the usual "anti brute force" people hasn't voiced their concerns/complaints yet.
Since afa looking at the picture of the Parhelia chip design (fake or not) goes, it doesn't seem to have any bandwidth saving functionality and no crossbar mem controller.
Doesn't the Parhelia have the potential to be a very ineffective chip/card ?
Well, ok, i guess we'll know more tomorrow so it doesn't do us that much good to take this much further at this point.
But as you said Joe, how much fun would we have if we could'nt speculate about it ? :-)
Since afa looking at the picture of the Parhelia chip design (fake or not) goes, it doesn't seem to have any bandwidth saving functionality and no crossbar mem controller.
Doesn't the Parhelia have the potential to be a very ineffective chip/card ?
umm.. take a another look. check Depth Cache and Depth Acceleration parts. No one has been able to explain what those do, but most likely they have something to do Z-Buffer (perhaps Early Z-reject and/or Z-buffer compression.)
but we have only "[22:11] -> 18hrs 48mins 19secs'" time left, so tomorrow we will hear what they have in it. Before that I am willing to doubt other sources. (in good as also in bad.)
Joe DeFuria
13-May-2002, 20:11
Lol Bjorn! I figured someone would respond despite my plea! ;)
If the NV30 or RV300 or whatever it might be with a 128 bit bus is faster then a parhelia with a 256 bit bus and also cheaper...
Yes, it's possible that both could be the case. However, I have a hard time believing that any card (save a deferred rendering card) could overcome a 2X deficit in memory bandwidth. (Unless the Parhelia is horrendously ineffective with bandwidth use.) Certainly, "cheaper" sounds like a much greater possibility, but even that isn't really clear, because Parhelia wouldn't need to use the most exotic, "at a high premium" memory and still maintain a sizable bandwidth advantage over 128 bit counterparts. There's a trade-off for the price of the complex 256 bus, vs. the cost of the most exotic memory available.
So, to be clear, what I do care about is indeed the bottom line. (Actual price / performance). "256 biut bus" is certainly not some checklist feature that I require my next card to have. It's just that clock-for-clock, I'm not sure I see any "traditional" 128 bit card would be able to compete in raw performance.
Joe:
So, to be clear, what I do care about is indeed the bottom line. (Actual price / performance). "256 biut bus" is certainly not some checklist feature that I require my next card to have. It's just that clock-for-clock, I'm not sure I see any "traditional" 128 bit card would be able to compete in raw performance. Well said, specially with trully next generation games.
Since afa looking at the picture of the Parhelia chip design (fake or not) goes, it doesn't seem to have any bandwidth saving functionality and no crossbar mem controller.
Doesn't the Parhelia have the potential to be a very ineffective chip/card ?
umm.. take a another look. check Depth Cache and Depth Acceleration parts. No one has been able to explain what those do, but most likely they have something to do Z-Buffer (perhaps Early Z-reject and/or Z-buffer compression.)
but we have only "[22:11] -> 18hrs 48mins 19secs'" time left, so tomorrow we will hear what they have in it. Before that I am willing to doubt other sources. (in good as also in bad.)
Well, i would be surprised if they didn't implement any bandwidth saving functionality at all.
Still nothing about a crossbar controller though :-)
Hellbinder
13-May-2002, 20:58
im sorry but...
What the &%*$ are you guys talking about *crossbar* memory controler...
Its a 256bit memory controler!!!! Good God people... Do you really think the world revolves around Nvidia design philosophy??? The damn thing has 20GB bandwidth... Can you read? 20gb bandwidth! That is 2x a GF4. You actually are trying to compare this to a 128bit Crossbar controler?? as if its somehow superior???
You have already been told that the P-512 was 30% faster than a GF4 Ti 4600 in ALPHA. Yet you are going to make these kind of comparisons anyway? (meaning comparissons to 128bit technology AKA Nvidias memory architecture)
This stuff is Hilarious!
Geeforcer
13-May-2002, 21:02
]Its a 256bit memory controler!!!! Good God people... Do you really think the world revolves around Nvidia design philosophy??? The damn thing has 20GB bandwidth... Can you read? 20gb bandwidth!
Effective bandwidth use is always important, whether you have 2, 20 or 200GB/s.
My guess it is a 2-way memory controller (2x256bits).
It means not only very high bandwith with comodity RAM but also potentially low latency.
When Nvidia GF3 need to read a 256bits texture and some framebuffer it will need at least 2 cycles. Maybe with this 512 bits you can do it in 1 cycle.
Well... what do people think about Parhelia vs. p10? At equal clockrates p10 should have slightly more geometry power (not sure about displacement mapping however). It's a little hard to see how well it will do with texturing: it has less TMUs, and anistropic requires programming. However, perhaps each TMU in the p10 is capable of trilinear filtering?
The pixel math ops can be done with 16 arithmetic units per pipe (versus 5 combiners for the parhelia), so the pixel compute power seems approximately equivalent. No idea as to which will be more flexible.
The fact that p10 also handles multiple graphics threads and has a virtual memory manager is a big plus.
That said, Matrox may have similar features, we'll see tomorrow...
Regards,
Serge
Joe DeFuria
13-May-2002, 21:23
You have already been told that the P-512 was 30% faster than a GF4 Ti 4600 in ALPHA. Yet you are going to make these kind of comparisons anyway? (meaning comparissons to 128bit technology AKA Nvidias memory architecture)
Well, to be clear, we're really speculating on P-512 performance next to nVidia's "Next-gen" chip, not the GF4. I fully expect Parhelia to pretty much hand the GeForce4 Ti its hat in performance....especially with AA turned on.
But who knows what nVidia is doing with their NV30 tech: we are assuming still 128 bits DDR bus...but what additional bandwidth efficiency measures beyond what's already in the GeForce4 ti? No one knows. Might be significant...might be "not much."
Again, I don't see a 128 bit NV30 beating a similarly clocked 256 bit Parhelia, in fillrate / bandwidth limiting situations. But depending on exactly how much improved nVidia's NV30 pipelines and bandwidth efficiency is, and how good/bad Parhelia's bandwidth efficiency is, it may end up being competitive.
i'd got for the dualbus ..
they've always had that since , the g200 ..
i don't think matrox wants to battle it out with the nv30 and r300 using the P-512.. when they come out , after a few months matrox will release.. a new and improved P-512 ..
( let's hope it's not a g450 )
Ascended Saiyan
13-May-2002, 21:56
[quote]Again, I don't see a 128 bit NV30 beating a similarly clocked 256 bit Parhelia, in fillrate / bandwidth limiting situations. But depending on exactly how much improved nVidia's NV30 pipelines and bandwidth efficiency is, and how good/bad Parhelia's bandwidth efficiency is, it may end up being competitive.I think it would be quite impossible especially with anti-aliasing,an enhanced color depth & @ 1280 x 1024 for a Nv-30 with 900 Mhz 128 bit memory to compete with Parhelia even with all the bandwith-saving techniques unless some devine entity showed them how,14.4 GB might not be enough.
BTW,Isn't the P10 supposed to be marketed at the highend desktop & when is it suppose to make some sort of debut ?
Geeforcer
13-May-2002, 22:05
i don't think matrox wants to battle it out with the nv30 and r300 using the P-512.. when they come out , after a few months matrox will release.. a new and improved P-512 ..
( let's hope it's not a g450 )
I very much doubt that in light of this (http://forum.matrox.com/mgaforum/Forum8/HTML/001120.html):
All I can say is that you'll have to save up more than the GF4 4400.
Allow me to be in sales mode for this Remember that we do not depreciate the value of our own products every six months just becuase of more frame rates. Hence, your value with our card will go alot longer.
]im sorry but...
What the &%*$ are you guys talking about *crossbar* memory controler...
Its a 256bit memory controler!!!! Good God people... Do you really think the world revolves around Nvidia design philosophy??? The damn thing has 20GB bandwidth... Can you read? 20gb bandwidth! That is 2x a GF4. You actually are trying to compare this to a 128bit Crossbar controler?? as if its somehow superior???
You have already been told that the P-512 was 30% faster than a GF4 Ti 4600 in ALPHA. Yet you are going to make these kind of comparisons anyway? (meaning comparissons to 128bit technology AKA Nvidias memory architecture)
This stuff is Hilarious!
2, 0 G, b . b, a, n, d, w, i, d, t , h. hmm, 20 gb.. band.. width.
yes, 20 gb of bandwidth.
Took a while since i just learned how to read :-)
Anyway, who has said that a 128 crossbar controler is superior ?
Also, afaik, the wider the bus, the more inefficient it becomes without a crossbar controler. And since a 256 bit bus is more expensive, well, then you can use more exotic RAM with a 128 bit bus.
Now, as Joe said, which will be better afa price/performance goes ?
That still remains to be seen but it sure will be interesting to follow the "battle".
And regarding the "already 30% faster then the Ti4600.
I think we're talking about the Parhelia vs NV30,R300 and so forth and who knows what kind of bandwidth saving techniques they have ?
Edit: i forgot to update the thread. I see that Joe already answered Hellbinder.
Hellbinder
13-May-2002, 22:26
right,
However I dont expect that the GF5 will be any more than about 40% faster than a Gf4ti. Thats why I brought the point up.
Of course thats just a feeling i have, and based on some stuff I read about the NV30 pushing about 200mpollies peak. I dont think its going to be that big a jump in real world performance. I think Parahelia-512 is going to be the same speed or even slightly faster than a GF5.
If might even be considerably faster, if rumors i have seen hold true... Time will tell.....
Geeforcer
13-May-2002, 22:41
BTW, quite frankly I don't understand where all this talk about Matrox releasing a die-shrunk, DX9 compatible version "in several month". Consider their history: Their last gaming card was released in 1999; since then they have released 2 business-oriented derivatives that were actually slower then the original G400. Taking this alongside with Haig's comments, I find the notion that Matrox is suddenly going to adopt an aggressive 6-month cycle quite absurd.
GeeForcer: I agree partially, but there is a another point of view...
Parhelia is new core from scratch. something like this last they have done with G100.
everyone remember G450 and G550, but no one remembers the jump between G -series cores before that.
this how they planned it:
G100 -> G200 (->G250) -> G400 (->G450) -> G800
and this is how it turned out:
G100 -> G200 (->G250) -> G400 (->G450) -> G550
what happened to G800 then?? I am too tired to explain it about 50th time, so look for it using search. ;)
so I am suspecting that we see Parhelia core based refresh in H1/2003. Matrox has normally released one chip in the year so it would fit nicely. Also SiS is introducing their DX9 core in H1/2003, so it would be perfect time frame for Matrox too. I don't except big changes to general design, but die shrink, higher clocks and DX9 compliancy perhaps.
Matrox has now pretty good base to build up next three years of parhelia based cores.
McElvis
13-May-2002, 22:52
About possible NV30 speed -
When the GeForce 1 was first released, was it not only slightly faster than the TNT2 Ultra?
When the GeForce 3 was first released, was it not slightly slower than the GeForce 2 Ultra?
What makes anyone think that the NV30, when it's first released, will be significantly faster than the GeForce 4 Ultra (sorry GF4 4600...)?
McElvis: what makes situation so interesting this time? well three little letters: ATI. Because ATI is most likely to be releasing their rival at the same time. not 6-7 months later like in previous times...
I am pretty sure that both have been "spying" each others plans for DX9 HW, but have they (ATI and nVidia) had time for following Matrox and 3DLabs?
when the battle is duel, you only need to win competitor mariginally. There is no need taking huge leaps forward, just winning rival is enough. If 3rd competitor with different product cycles comes in, planning "The Best chip / card" gets in to new dimensions.
aargh.. too much typos... I am going to see TrueIdle3D with perfect S.L.E.E.P AA. ;)
McElvis
13-May-2002, 23:13
McElvis: what makes situation so interesting this time? well three little letters: ATI. Because ATI is most likely to be releasing their rival at the same time. not 6-7 months later like in previous times...
I am pretty sure that both have been "spying" each others plans for DX9 HW, but have they (ATI and nVidia) had time for following Matrox and 3DLabs?
when the battle is duel, you only need to win competitor mariginally. There is no need taking huge leaps forward, just winning rival is enough. If 3rd competitor with different product cycles comes in, planning "The Best chip / card" gets in to new dimensions.
aargh.. too much typos... I am going to see TrueIdle3D with perfect S.L.E.E.P AA. ;)
True, true.
BTW, I like Anand's little comment on his latest news post. Wonder what he ment when he said that "NV30 is amazing" :wink:
Well, without going into something that I can't , perhaps, just perhaps this fall is going to be exciting for the first time in years in 3d graphics...
Entropy
13-May-2002, 23:30
Two observations:
First re: What do the people want?
Just look at this forum. What generated the buzz, the elegant, enabling and intriguing new chip from 3DLabs, or the new Matrox? Why has Parhelia gotten so much more attention? Becasue it promises to be a "Geforce4 killer".
Second re: NV30 will produce higher quality pixels (than current Geforces)
Well, it had better, because it seems quite certain that the Parhelia will be able to produce higher quality pixels than the current Geforces as well. It's safe to say that most any new design that comes to market will have to provide either better speed or features to be able to compete with what is already available at the same price point on the market. Frankly that comment from Kirk was a pure (and rather thin) attempt at damage control. I don't see how anyone could believe otherwise. Obviously, being a factor of two behind in bandwidth is a rather serious handicap. Yes you can use BW-saving techniques, more sophisticated cacheing and prefetching, push the clocks higher, muck around to find various optimums, bla bla bla, but unless the competition is incompetent you are not likely to do more than narrow the gap. Fundamentally, fast data paths are enabling technology. No matter who you are, it will always be an advantage to have more bandwidth, rather than less. With Matrox and Creative both releasing higher bandwidth solutions, it would be stupid of nVidia and ATI to balk at adding some pins and traces. It requires no out-of-house IP, royalties or circuitry. Even though bandwidth is not the only thing influencing performance it is still _very_ important, so why give your competition an advantage like that? They'll only do it if their next part is so far gone that it can't be re-engineered, which is likely to be the case for the R300 and NV30. But it would surprise me a lot if we saw any card with high-end ambitions introduced after these which doesn't use 256-bit memory paths or similar (, or better still). It wouldn't make sense.
Entropy
"elegant, enabling and intriguing"? Thats a matter of perspective I guess, anyway the most important difference is that this looks to be a _product_ launch. All the 3dlabs whitepapers in the world arent going to make as much impact as an honest to god product ...
Entropy
13-May-2002, 23:46
All the 3dlabs whitepapers in the world arent going to make as much impact as an honest to god product ...
:) True, though at this point, we actually have more on their products to come than on Matrox'. And note that 95% of the discussion is about Parhelia performance, not about its' quality enhancing/assuring tech.
Speed comes first. When we have sufficient speed, improved IQ becomes important. The focus of the forum is reasonable.
Entropy
dksuiko
13-May-2002, 23:51
Ben,
From your recent posts:
Quote 1: "Well, without going into something that I can't..."
<then don't go into it>
Quote 2: "I am hearing from a undisclosed source (sorry I'm not going to even hint as to who it is.)"
<nobody asked for your source>
Quote 3" "Actually, now I can't say at all. No , I'm not under any kind of NDA if that's what you're thinking..."
<then don't say anything>
Quote 4: "Hrm without directly answering the question..."
<then don't answer>
Quote 5: "Finally a solid date. Beyond that , no comment..."
<nobody asked, don't comment>
Okay, really.. Why post that you can't post info over not posting at all? Just only somewhat annoyed...
-dksuiko
Dave Baumann
13-May-2002, 23:52
a matter of perspective I guess, anyway the most important difference is that this looks to be a _product_ launch.
Not if the product isn't due until late summer. I'd suggest that products based on P10 may well be available before Matrox's part, and there were extenuating circumstances tha prevented 3Dlabs in some areas.
Laa-Yosh
14-May-2002, 00:00
Well, alpha cards are already out for a while now. Some game devs here have 175MHz boards...
It's already the 14th of May here, and within 5 minutes it'll be 12:00 AM GMT as well, so NDA's are lifted as far as I know :)
point taken, dksuilko. As to the specific posts , I could comment on it before, as all I had were guesses, but now that situation perhaps has changed. As I haven't signed an NDA on the matter , nor has the source actually asked me not to discuss it, it's a gray area. Out of respect to the companies involved, I decided not to and provide a suitably ambiguous answer. Does that make any sense? If it doesn't sorry to have offended your sensibilities.
It's .. at 3pm EST
EST
not your time, their time ..
Can someone refresh my memory as to these questions:
When did development on NV30 start (in all likelyhood)?
When was 3dfx acquired? (Did this acquisition delay NV30, seems like NVidia missed a "completely new product" cycle?)
When did development on R300 start (in all likelyhood)?
When was ArtX acquired by ATI?
As for ATI and NVidia competing with 3dlabs/Matrox, I think they
will at least in features/programmability, as both seem to be aiming
for DX9...
Another question: DX9 and DX9.1... Are we going to see a repeat of the GF3 vs. R8500? Either ATI or NVidia having features which require an update to DX to utilize fully?
Laa-Yosh
14-May-2002, 00:08
It's .. at 3pm EST
EST
not your time, their time ..
So are those Japanese guys in violation of their NDA? Se the other thread here...
Ailuros
14-May-2002, 00:15
A couple of notes:
Frankly I can´t understand how people can get that exited with just numbers or even come as far to draw conclusions. I rather wait for Parhelia or any P10 derivative to be extensively tested/analyzed before I can even draw any conclusion.
It becomes an even more moot point to predict how the forementioned will stack up with upcoming R300 or NV30, when not even specs are widely known to the public.
yeah , they are .. but eh i don't think they'll make a big deal out of it
it's not like , we're 2 months away
There are many solutions to improving memory bandwidth. Different 3d vendors will be using different techniques in the DX9 timeframe.
When did development on NV30 start (in all likelyhood)?
When was 3dfx acquired? (Did this acquisition delay NV30, seems like NVidia missed a "completely new product" cycle?)
Advanced development probobly started in 2000. 3dfx was officially aquired in Dec. 2000, but there was nVidia technical staff looking at 3dfx's roadmap and technology for a little bit with an eye for integration.
IMHO, NV30 has undergone an overhaul, especially in the texturing department. Nvidia has the pool of resource and talent to make this kind of redesign in short time... something a Matrox, or perhaps even ATI, don't have.
Another question: DX9 and DX9.1... Are we going to see a repeat of the GF3 vs. R8500? Either ATI or NVidia having features which require an update to DX to utilize fully?
Good question, ATI seems to have the advantage with MS using them as the reference platforms for DX9 development. Anyone know what ever became of nVidia and MS's little falling-out?
Typedef Enum
14-May-2002, 04:09
Anyone know what ever became of nVidia and MS's little falling-out?
Yeah...A lawsuit! Time for a rant...
Microsoft...of all freakin' people...charge that nVidia is actually asking _too much_ for the NV2A chips.
Of course, I don't seem to recall Microsoft ever complaining all the years that they've been ripping people off with some of the utter rediculous pricing of their own products.
Doomtrooper
14-May-2002, 05:32
Ben,
From your recent posts:
Quote 1: "Well, without going into something that I can't..."
<then don't go into it>
Quote 2: "I am hearing from a undisclosed source (sorry I'm not going to even hint as to who it is.)"
<nobody asked for your source>
Quote 3" "Actually, now I can't say at all. No , I'm not under any kind of NDA if that's what you're thinking..."
<then don't say anything>
Quote 4: "Hrm without directly answering the question..."
<then don't answer>
Quote 5: "Finally a solid date. Beyond that , no comment..."
<nobody asked, don't comment>
Okay, really.. Why post that you can't post info over not posting at all? Just only somewhat annoyed...
-dksuiko
You get used to it after a while, a post about nothing that says nothing then when there is a post with something its so far out to left field.. Like some of classic Geforce 3 posts that were made on the old forums.
There are some people that feel if they know a buddy, that is married to his cousin who has a beer with a buddy that knows somebody from Brand X Video company posts that as reliable info :wink:
Haven't we all learned that there is so much more to a card's performance than just bandwidth alone?
Cache architecture is huge, and rarely talked about by the fans.
Regardless, I think Nv30 is going to stay with 128bits, and instead opt for silicon in the efficiency department.
(hierarchical z buffers, better texture compression, possible framebuffer compression, more z check units, lower lvl of crossbar, smart adaptive AA techniques, etc etc)
All of these can go a long way in evaporating a brute force approach.
Its all about price/performance tradeoffs and how the engineers best feel they can achieve it (and Nvidia is no stranger to bruteforce techniques)
for the god's sake, look what you are talking about...
no one even knows pratical speed of parhelia, without even mentioning that card models, their specs and pricing isn't officially announced!
so I would wait for that before talking about Price/performance ratio.
this is almost ridiculous. :D
Ailuros
14-May-2002, 07:37
Of course, I don't seem to recall Microsoft ever complaining all the years that they've been ripping people off with some of the utter rediculous pricing of their own products.
Since the major volume is software, I'd say that they're fully aware that there are hundreds of thousands of illegal copies floating around.
You can't exactly "hack" your old Dreamcast and turn it into a Xbox can you?
There are some people that feel if they know a buddy, that is married to his cousin who has a beer with a buddy that knows somebody from Brand X Video company posts that as reliable info.
To the best of my knowledge it's not the case with Ben, since the original comment was addressed to him; but even if it would be the case it's pretty harmless.
Ben has written more than just a couple of good reviews and he's quite objective with any piece of silicon I've seen him touch so far; a "talent" that's rather rare.
Entropy
14-May-2002, 11:44
Haven't we all learned that there is so much more to a card's performance than just bandwidth alone?
Cache architecture is huge, and rarely talked about by the fans.
Regardless, I think Nv30 is going to stay with 128bits, and instead opt for silicon in the efficiency department.
(hierarchical z buffers, better texture compression, possible framebuffer compression, more z check units, lower lvl of crossbar, smart adaptive AA techniques, etc etc)
All of these can go a long way in evaporating a brute force approach.
Its all about price/performance tradeoffs and how the engineers best feel they can achieve it (and Nvidia is no stranger to bruteforce techniques)
I'm not saying that you are wrong per se, but why are you implicitly assuming that the other players are unaware of bandwidth efficiency enhancing techniques? Crossbar memory controllers for instance are touted as an amazing thing in these circles, but they've been around for decades. (Look at the last thirty years of supercomputing, and you are likely to find a host of tricks, some of which are yet to be utilized where they potentially could be profitable.) We have no reason to assume that other manufacturers are bumbling incompetents in these regards. After all, if they hadn't been concerned with bandwidth they wouldn't have gone to wider datapaths in the first place. I think it is very safe to assume that they will have all the obvious bases covered, and some advantages of their own.
If the NV30 still uses a 128-bit DDR memory interface, it will obviously be at a disadvantage bandwidthwise vs. its competition. If that's the case, then it is due to nVidia somewhat underestimating the roadmaps of the competition. nVidea are likely to minimize the impact by letting their driver team burn the midnight oil, push clockspeeds of GPU and memory, lower their profit margins to improve their price/performance ratio and hype the hell out of whatever new features the NV30 brings to the table. They will probably do a good job at these things, at least sufficiently to tide them over until their next part, gradually sliding into what ATI is doing with the 8500 now ie position the NV30 as a medium-priced part, offering good price/performance.
But to believe that nVidia can easily compensate for a factor of two bandwidth disadvantage requires the implicit assumption of incompetence on the part of their competitors.
The leaks show that Matrox will bring lots of good technology to desktop graphics - hardware displacement mapping, fragment based high level AA, depth adaptive tesselation, 10-bits per colour channel (yes!) plus their multi-display tech and their high video bandwidth - not signs of incompetence, in my book.
Entropy
You assume that its free to implement occlusion detection hardware. It's not, it costs silicon, die space, money and power consumption.
The 256 memory bus is expensive to implement (the extra pins), so there is a certain amount of leeway a straight 128 bit chip has (all things being equal).
This is fairly reminiscent of the arguments 3dfxers gave during the V5 vs Gf2 era (Nvidia is doomed, there next card is going to suck b/c bandwidth will limit them just like GF2, with 3dfx the problem is solved). Yet sure enough the next chip did improve the generational divide. Its almost like clockwork with Nvidia, in there 40% increase in speed + additional performance in certain scenes/driver updates.
Anyway, lets wait till we actually see the cards perform before predicting gloom and doom.
Dave Baumann
14-May-2002, 22:51
Just look at this forum. What generated the buzz, the elegant, enabling and intriguing new chip from 3DLabs, or the new Matrox? Why has Parhelia gotten so much more attention? Becasue it promises to be a "Geforce4 killer".
Entropy, I actually think that this is where 3Dlabs’ biggest problem will lie – marketing. The nature of the 3Dlabs architecture is one of flexability and as such its devoid many annoying marketing buzzwords ‘nFiniteFX’, ‘SMARTSHADER’, ‘LightSpeed Memory Architecture’, ‘HyperZ’, etc., etc. – all of these are actually just names for fixed function areas of the pipeline, something that P10 just doesn’t have.
Its going to be very difficult to market that, as the noise here demonstrates, and people can easily overlook what a cunning architecture it actually appears to be.
Hellbinder
15-May-2002, 01:30
BTW, I like Anand's little comment on his latest news post. Wonder what he ment when he said that "NV30 is amazing"
McElvis...
Well, My personal feeling is that Anand is an Nvidia @$$ kissing freak. With no sense of ethics at all. But thats my personal opinion of that worthless bumb ;)
At any rate, he says crap like that as if the R300 does not even exsist. WEll see what is *amazing* when the time comes. The R300 was designed by a whole new team at ATi, Engineers that built the Flipper. I know at least a part of what it brings to the tabe from information I have personally seen, and it was legit.. you could just tell. Unless things have changed dramatically, R300 has 8 pixel Pipelines and 4 Geometry pipes. Who knows what else.. Nv30 Was origionally leaked to have 6 pixel pipes and 3 Geometry. i dont know about the NV30, But the R300 I am very confident that what i stated is in fact what they have at a minimum.
My point being Anand and his Nvidia Fanboy, blatant take the money from daddy Articles and *reviews* are sickening, and IMO dishonest.
More reasons for more discussion on Parehelia:
It's easier to grasp the Matrox info since it's an enhanced traditional architecture, while P10 is more of an revolution. There are question marks around Parhelia, but they are "small enough" that we can make guesses on how it could be implemented. P10 has got so "big" questionmarks that it's difficult to do any guesses. We don't know anything about how the programming is done, instruction set, registers, input variables, memory accesses, flexibility in framebuffers, ... , everything is open.
Parhelia is directed towards highend gamers (at home or hidden behind a professional disguise :)), while P10 at least now initially leans more to the professional (by pricing).
I get the impression that Parhelia is closer to production, at least for cards that a gamer possibly could buy.
(But if I were to get one of Parhelia and P10, I wouldn't hesitate a second. Programmability all the way, I'd want a P10 even if it got whoped in most games.)
[Edit]
Gaahh, why do I use swedish spelling.
Basic,
Definitely agree. Hopefully we will not have to wait too long for p10 and its programming manual.
I really hope a similar level of programmability will be available in ATI and NVidia's next parts. One NVidia pdf i have read seems to indicate that full programmability is the "future". But then right afterwards they say this will take some time to happen...
Serge
Oompa Loompa
15-May-2002, 02:15
]
Well, My personal feeling is that Anand is an Nvidia @$$ kissing freak. With no sense of ethics at all. But thats my personal opinion of that worthless bumb ;)
Well, OK, but I invite you to consider an alternative hypothesis: Anand is an objective, well-intentioned writer who occasionally makes mistakes but is still far above average in the internet "press", whereas you have lost your objectivity.
Reading into the OpenGL 2.0 whitepapers by 3dlabs may give some insights about P10 VPU.
For an example, I was puzzled by the "multi-threading" capability provided by P10. When I read the "Async. OpenGL" whitepaper again, I suddenly realize that its "multi-threading" capability is designed for the async. functions in the OpenGL 2.0 whitepaper.
It definitely will be a very interesting product :)
Now if they were to say bolt on the Imagine Stream chip as a vertex shader front end Id be first to agree with yall on programmability, but I have to say Im slightly underwhelmed by what even the P10 has to offer in that respect ... its only slightly more usefull as VS 2.0 IMO, whereas DX9 in its turn has some T&L features for which the P10 will almost certainly not even be programmable enough to emulate (namely displacement mapping). So much for programmability ...
<lazy>
Where do I find OpenGL2.0 specs?
Didn't find anything when giving a quick look at www.opengl.org
</lazy>
Dave Baumann
15-May-2002, 03:04
http://www.3dlabs.com/support/developer/ogl2/index.htm
Now if they were to say bolt on the Imagine Stream chip as a vertex shader front end Id be first to agree with yall on programmability, but I have to say Im slightly underwhelmed by what even the P10 has to offer in that respect ... its only slightly more usefull as VS 2.0 IMO, whereas DX9 in its turn has some T&L features for which the P10 will almost certainly not even be programmable enough to emulate (namely displacement mapping). So much for programmability ...
I don't think there is a clean and good method for hardware displacement mapping. Since DX9 has not gone beta yet, I don't know what method they decided to go. However, it all about an extra constant register storing a "displacement value" in the vertex shader, for each vertex. How the displacement value is computed is definitely not programmable, hence we have two methods. Therefore, I don't consider the displacement mapping is about programmability. It is just another feature.
That was hardly my point, if it were programmable enough to efficiently add displacement mapping through programming Id be more impressed. But really "its only slightly more usefull as VS 2.0 IMO".
Well, then we have to wait to see :)
Although I'm not sure about the capability of P10 VPU, but if it resembles the OpenGL 2.0 vertex shader functions, I'd say it is much better than VS 2.0.
IIRC VS 2.0 only provides more registers, longer instruction length, and some limited flow control instructions (only operates on constants). The OpenGL 2.0 vertex shader provides flow control instructions not limited to constants. This alone is much better than VS 2.0 IMHO.
"
At any rate, he says crap like that as if the R300 does not even exsist.
Don't be too quick to judge from the "nv30 is amazing" quote. In the next sentence Anand said he was going to see Ati in a week. He obviously can't call R300 amazing if he hasn't seen it yet.
pcchen, DX9 will include Matrox's displacement mapping method. I wasn't too excited about this feature until I watched the video's on Matrox's site. Now I'm interested to see how/if developers will use it.
http://www.matrox.com/mga/products/parhelia512/technology/disp_map.cfm
I especially liked the Character Rendering Process and Depth Adaptive Tesselation videos.
Maybe they'll send Humus a card and he can tell us how easy it is to use. :)
MfA,
You do have a point. Imagine has 20GFlops peak performance with 20million transistors. For 3d I'd say remove most of their integer execution units and replace with fp would be a good idea, so I figure it could probably be even higher than that.
Looking at the design it looks like an 8 way vertex shader (different ISA) with operations generalized to arbitrary "stream" elements (which is basically what vertices with their arbitrary attributes are beginning to look like). So it looks great for vertex shaders (and even pixel shaders IMO), except where the stream programming model breaks down (texture access, displacement map access).
i don't see that this programming model/architecture is well suited to something like adaptive subdivision surfaces, tesselation of HOSes, displacement mapping - basically higher-level processing of primitives -
but it's definitely a step up from any vertex shader i've seen so far.
Also, I seem to recall 3dlabs saying that their vertex unit could probably handle VS2.0 (but that they weren't certain since DX9 hasn't been released yet), which would seem to indicate that displacement mapping might be possible?
It seems to me that in 3d you need support not just for streams of records (pixels and vertices), but also N-D arrays of records (textures, displacement maps). For things like subdivision surfaces you probably need access to a set of records at time...
i'll go pick it up myself , and i'll cook something up .
eh
i did a nice "hello world" program a few years ago
westwood did the mountain demo , i've read
pcchen, DX9 will include Matrox's displacement mapping method. I wasn't too excited about this feature until I watched the video's on Matrox's site. Now I'm interested to see how/if developers will use it.
I know that already. IIRC another method is from NVIDIA and is a bit more cumbersome.
However I don't know how exactly Matrox is doing. Using displacement mapping on terrain is straightforward and simple. However, to use it on a complex model is another story. As you can see, the amount of displacement is much larger on terrains than on complex character models. Of course there are good reasons for this.
To use displacement mapping on terrains should be quite easy. Note that the amount of displacement is represented as a constant in the vertex shader. Therefore, you can do many things about it easily. On the other hand, to use displacement mapping on a model is a bit harder, especially on the authoring side. Without proper tool, it is not easy to build a displacement map to create the effects shown on the Matrox's site. There are also some possible problems with displacement map filtering affecting the results.
The situation is not different from the bump mapping, or per pixel lighting. Bump mapping is not hard to use, and bump maps are not hard to create. However, there are still only few games supporting bump mapping or per pixel lighting on everything.
Dave Baumann
15-May-2002, 08:55
Also, I seem to recall 3dlabs saying that their vertex unit could probably handle VS2.0 (but that they weren't certain since DX9 hasn't been released yet), which would seem to indicate that displacement mapping might be possible?
Yes its possible.
Each of the 16 scalar processors used as the vertex engine in P10 are evidently just mini DSP's, so these can be programmed to do just about anything, including Displacement mapping.
Gunhead
15-May-2002, 11:19
I hope Matrox explains those Depth-related pieces!
Hellbinder, a crossbar memory controller isn't a Nvidia invention... SGI et al. has been using them for a long time. And of course we can and should compare different memory architectures, nothing wrong with that. Matrox could very well have a xbar controller in a future Parhelia. Although I would agree with you that people here have been unduly downplaying the power of Matrox' "crowbar" brute force controller :wink:
Hey folks, now all Matrox has to do in near future is to implement those same bandwidth-saving techniques (including the crossbar :P ) and they stay competitive (possibly even ahead, maybe), right?
[Edit: LOL, reading only page 1 before posting makes funny things to your context...]
Lessard
15-May-2002, 12:51
]
Its a 256bit memory controler!!!! Good God people... Do you really think the world revolves around Nvidia design philosophy??? The damn thing has 20GB bandwidth... Can you read? 20gb bandwidth! That is 2x a GF4. You actually are trying to compare this to a 128bit Crossbar controler?? as if its somehow superior???
I'm glad you found this stuff hilarious...
How do you explain Matrox engineers expect only a 20-50% perf increase in "raw" benchs vs the 4600 even if they have twice the memory bandwidth ? Anyway we were talking about the NV30 and there are a lot of chances that it will be faster than the Parhelia 512 in a majority of benchs like it or not ...
The best feature of the Parhelia is their clever AA (it it's not flawed) not their 256 bit bus. If Nvidia decided not to implement a 256 bit memory bus for the NV30 and go with 8 pipelines, they should have some reasons. Why jump to 8 pipelines if you're already seriously bandwidth limited with 4 pipelines ?
Guillaume
Joe DeFuria
15-May-2002, 14:13
How do you explain Matrox engineers expect only a 20-50% perf increase in "raw" benchs vs the 4600 even if they have twice the memory bandwidth
Well, there's a logical explanation for that: Because they expect the "raw" benchmarks to no longer be bandwidth limited on the Parhelia. Most "raw" benchmraks (that only us dual texturing) may be bandwidth limted on the 4600, but dual texturing benchmarks with 20 gb/sec on the Parhelia will end up being fill-rate limited. Thus, you don't get a 100% increase in performance....you only get an increase to the degree that the 4600 is bandwidth limited.
However, even without bringing AA into the equation: move to Quad texturing games, or turn on anisotropic filtering, and then even the Parhelia will probably end up pushing the limits of it's bandwidth, and that's when (theoretically) the performance delta will really be large. The GeForce 4600 will drop significantly in performacne, where the Parhelia should take a minimal hit.
Maybe Parhelia will be able to do Doom3 in a single pass and in a single cycle !!!
Parhelia is for new bandwith demanding games not old games. IIRC (from MURC) id already has Doom3 running with Parhelia but they cant comment about it.
I would like too se a Doom3 benchmark with 1024x768x32 with 64tap aniso and FSAA. Then we can talk about performance.
Kudos to Matrox and 3DLabs to use 256bits memory bus.
Ailuros
15-May-2002, 15:40
I would like too se a Doom3 benchmark with 1024x768x32 with 64tap aniso and FSAA. Then we can talk about performance.
If you mean Parhelia I'd rather skip the S out of FSAA, but that's just me.
Ok, lets skip the S, and?
To really make the the FAA enjoyable, you'll *have* to be able to use aniso, or texture quality will be inferior to FSAA.
Also, from what I seem to remember the Parhelia doesn't do 64tap aniso but only 16tap. 16tap is not the answer to my prayers but at least the performance hit when enabling it will be minimal.
Are you sure it doesnt do 64tap????
pretty sure, it takes 64 samples, but that doesn't mean its 64tap, its only 16tap if I didn't take the wrong pills today... :)
Again, it can not do 64 tap with multiple passes?
16tap it does with a single pass, right?
There is no mention at www.matrox.com about filters with aniostropy higer than 4 (what you call 16-tap), and you'll only get that with single textured "bilinear anisotropic". Use the trilinear version or multitextureing, and the anisotropy goes down. Maybe they can do some multi-cycle sampling, but don't you think its strange that they don't mention that?
If they haven't just forgot to put some info there, it seems as Parhelia will have the lowest level of anisotropy of the top dogs. Let's hope that it's just a PR-bug.
Each of the 16 scalar processors used as the vertex engine in P10 are evidently just mini DSP's, so these can be programmed to do just about anything, including Displacement mapping.
All we know is that they implement OpenGL 2.0. If they can efficiently access external memory is not known for sure by either of us I think, but I think you have the greater chance of being wrong.
MfA
In your opinion should the programmeable gfx chips of the future allow external memory access as CPUs do? Or would it be enough to access fixed length records in a stream/array/? Does it make sense to have memory access instructions only for the data structures common in 3d, but not for general memory access? Or are you leaning towards a CELL type approach where each computing unit is pretty much completely independent?
As far as p10 goes, i also doubt it can efficiently access external memory - i don't see any AGU/load units or caches on the vertex processor diagram.
Regards,
Serge
I think the stream approach would be flexible enough for 3D, Ill have to disagree with you on how suited it is to such things as adaptive tesselation etc. The problem to adapt to the different tesselation depths is no different from dealing with different size polygons (because rasterization also generates an adaptive number of samples) which they seemed to manage quite well. External memory access is possible, even if its done in a more structured way than with a general purpose processor. For displacement mapping for instance the tesselator would generate both a vertex stream, and a stream of indices which would generate the data stream (consisting of texels) needed for the later displacement.
Dave Baumann
15-May-2002, 21:37
Mfa,
All we know is that they implement OpenGL 2.0. If they can efficiently access external memory is not known for sure by either of us I think, but I think you have the greater chance of being wrong.
I spent two hours on the phone with Neil Trevett, 3Dlabs Senior Vice President Market Development, talking about the P10 architecture on Monday - I know for a fact that Displacement mapping can be enabled.
The converstation I had was going over the press breifing they have already given others, but obviously I knew a little more and could get more of a dialogue doing. I'll be bringing a tech preview of P10 similar to those already seen, however hopefully it should give a little more detail and answer the questions that have already been asked on this forum. If you have any more then I'll be able to mail 3Dlabs and I'm sure they'll be happy to answer before I do the article.
Imagine achieves 5.1 Gops from a peak of 20GFlops/40Gops, 14.9 million vertices/s (16.8 million pixels/s). Out of their test applications it's the least efficient in terms of processor usage. Granted they are doing the whole 3d pipe...
It would be interesting to see how well Imagine would do if it were just doing displacement mapping or tesselation + some vertex program...
What kind of vertex rate do they think they will manage with displacement mapped N patches ala the Matrox method? (including LOD) Will there be any overhead for this in the drivers? Will they expose the functionality needed to implement this? (because OpenGL 2.0 sure aint it, I dont think loading a texture as state parameters is realistic.)
Gunhead
15-May-2002, 22:05
Lessard,
One of the reviews around quoted a Matrox rep who said that -- IIRC -- they have tested about 20 games and FAA produced artifacts in about five; and that they are working on a compatibility list. Also that Parhelia sports a MS AA for games where FAA breaks. So there you go... While maybe not actually flawed, not universally applicable either (if there is such a thing).
vBulletin® v3.8.6, Copyright ©2000-2013, Jelsoft Enterprises Ltd.