The official GT300/Fermi Thread

[largon]

Are we talking about metal or silicon respins? Dunno why nV has only one letter and one number in the spin code - while ATi lists both silicon and metal spins. Anyways, if ATi's A0 was first revision then R600 would have gotten unrealistic total of 4 four respins, as early samples were A11 (1st rev silicon, 1st rev metal) when retail chips were A13. More likely, ATi's first rev is A11 meaning R600 had two metal respins (A11 -> A12 -> A13).

And, I haven't seen any nV, nor ATi, chips marked A0...

[Farinorco]

explain to me how 512 shaders is not over double 240 shaders. the bandwidth increased by 50% too. the theoretical numbers are not that impressive but you completely missed a lot of factors and posted wrong information. nvidia also said 1.5ghz is a conservative estimate for clockspeed.

512 shaders is over double 240 (x2.13 to be exact). But 48 ROPs is not over double 32 (x1.5 to be exact). And 230 MB/s is not over double 141 (x1.63 to be exact). So overall, it's not over double the specs of the previous one. I don't think it's so hard to get what I've said there, and I don't get where I've said anything about CPs not being double (I think I have mentioned +113%). I would also like to know what are all those lot of factors that I've missed and what wrong information I've posted, based on what we know at the moment.

And regarding clock speed, I would take it like talking about the shaders clock. I wouldn't expect much higher clocks than GTX285, if at all.

[Manicdan]

I wouldn't make conclusions yet, we haven't seen any GPGPU results for larrabee, or pricing.

however even if larrabee is slightly less powerful, I could still see businesses opting for it due to lower power usage and it will probably be cheaper

i believe intel said LRB was about 1Tflop at double precision, and GF100 i think is 3TFlop single and half that double. but this is all from memory and i could be wrong.

[ajaidev]

pic for ya




http://rs648.rapidshare.com/files/28...Key_Visual.jpg ( res. 6316 x 3240)

Fermi 1.4-1.6 x of GTX 295.

( 1.6-1.8x of GTX 285 is not too much)

rapidshare is not letting me download says :-"This file is neither allocated to a Premium Account, or a Collector's Account, and can therefore only be downloaded 10 times."

This card got torn apart on BTUK yesterday. Its been so crudely done (the PCB that is,) its insulting.

Any links cant find them, did Google noting came up...

[zalbard]

nvidia also said 1.5ghz is a conservative estimate for clockspeed.

Do you honestly believe in 1.5Ghz GT300 on stock settings?

[zerazax]

People were also talking about G200 being similar to G92 in clocks and look what happened at the first gen...

Fact is, when I hear Nvidia's own engineers claim that the design is delayed because it's incredibly hard, i'm not holding my breath on getting incredible clocks especially since they've had their own struggles going to 40nm

[003]

Do you honestly believe in 1.5Ghz GT300 on stock settings?

For the shaders, not the core. Obviously the core won't be 1.5GHz.

[Chumbucket843]

512 shaders is over double 240 (x2.13 to be exact). But 48 ROPs is not over double 32 (x1.5 to be exact). And 230 MB/s is not over double 141 (x1.63 to be exact). So overall, it's not over double the specs of the previous one. I don't think it's so hard to get what I've said there, and I don't get where I've said anything about CPs not being double (I think I have mentioned +113%). I would also like to know what are all those lot of factors that I've missed and what wrong information I've posted, based on what we know at the moment.

And regarding clock speed, I would take it like talking about the shaders clock. I wouldn't expect much higher clocks than GTX285, if at all.

games are bound by shaders in the majority of cases. you can see that clearly in the 5870. they are running games at ridiculously high resolutions on a single card and still its bandwidth that really bottlenecks pixel fillrates. the factors you missed were new memory hierarchy, better scheduling logic, predication, and instruction set improvements.

i would trust nvidia more than i trust you for the clockspeed.

[RaZz!]

--> http://www.nvidia.com/object/gpu_tec...onference.html

watch the opening keynote with Jen-Hsun Huang. He says it runs on Fermi. A lot better quality than youtube too.

And it's really interesting.

yep, especially the physx part is very interesting and impressive. it starts at about 1/4 or 1/5 of the video (no minutes or smth are shown :/).

[DilTech]

Here ya go guys...
http://www.hardocp.com/news/2009/10/...es_eyefinity63

I still don't care about multi-monitor for gaming until they make multi panel monitors into one frame, but good for those who do care.

[Manicdan]

Here ya go guys...
http://www.hardocp.com/news/2009/10/...es_eyefinity63

I still don't care about multi-monitor for gaming until they make multi panel monitors into one frame, but good for those who do care.

id expect both companies to have been able to do this for a while, just never cared to develop drivers to do it. if a x1800 can do 1920x1200, then i doubt they were hardware limited.

[Farinorco]

games are bound by shaders in the majority of cases. you can see that clearly in the 5870. they are running games at ridiculously high resolutions on a single card and still its bandwidth that really bottlenecks pixel fillrates. the factors you missed were new memory hierarchy, better scheduling logic, predication, and instruction set improvements.

I didn't miss that factors. They simply don't take any part in anything that I've said. And when it take it, I have mentioned them and considered them. Take the "trouble" of reading my posts and trying to understand them before quoting me, please, to not put things in my mouth.

And I don't know how to use HD5870 to know how games are shader bottlenecked since the proportion in which they have improved shader processing power it's the same than texture processing power, rasterizing operations processing power, and so.

There are more things involved in the 3D rendering process apart from shaders and memory bandwidth.

i would trust nvidia more than i trust you for the clockspeed.

Yeah, no doubt. But I think you have misunderstood them when you have the idea that they are talking about a clock of 1500MHz for the GPU core.

[Cybercat]

512 shaders is over double 240 (x2.13 to be exact). But 48 ROPs is not over double 32 (x1.5 to be exact). And 230 MB/s is not over double 141 (x1.63 to be exact). So overall, it's not over double the specs of the previous one. I don't think it's so hard to get what I've said there, and I don't get where I've said anything about CPs not being double (I think I have mentioned +113%). I would also like to know what are all those lot of factors that I've missed and what wrong information I've posted, based on what we know at the moment.

And regarding clock speed, I would take it like talking about the shaders clock. I wouldn't expect much higher clocks than GTX285, if at all.

Um, you don't have to double EVERYTHING to get doubled performance. More than anything this depends on the particular application you're running, and where the bottlenecks lie within it.

If you look at a past example where performance WAS doubled, like the 8800GTX, let's compare that to the previous gen flagship, the 7900GTX. The 8800GTX had almost exactly twice the GFLOPs of the 7900GTX, even taking into account the nearly useless MUL op. The 8800GTX had 69% more memory bandwidth, and get this, only 33% more pixel fillrate, and 18% more bilinear texture fillrate.

The GF100 is more of an improvement in raw specs over the GTX 285 than the 8800GTX was over the 7900GTX. So doubling performance is more than possible.

[003]

Um, you don't have to double EVERYTHING to get doubled performance. More than anything this depends on the particular application you're running, and where the bottlenecks lie within it.

Not always, for example the shaders are more efficient being MIMD/FMA.

Also you have to keep in mind, while the ROPs and TMUs were doubled on RV870, ask yourself, doubled to what? 32/80 respectively. GT200 already has 32/80.

[DilTech]

Are we talking about metal or silicon respins? Dunno why nV has only one letter and one number in the spin code - while ATi lists both silicon and metal spins. Anyways, if ATi's A0 was first revision then R600 would have gotten unrealistic total of 4 four respins, as early samples were A11 (1st rev silicon, 1st rev metal) when retail chips were A13. More likely, ATi's first rev is A11 meaning R600 had two metal respins (A11 -> A12 -> A13).

And, I haven't seen any nV, nor ATi, chips marked A0...

Of course you haven't, A0 is usually in-house only. Only chip I can think of that released as an A0 from NVidia is the NV15. Usually it takes a few revisions before they can release.

Also, the R600 DID take several respins before it could release if you remember. You're talking about a card that was 6 months+ late.

Now again, all that could have changed since then, but I've never heard or read anything to tell me that. I'd ask the reps, but that's likely information they aren't willing to let out.

[Farinorco]

Um, you don't have to double EVERYTHING to get doubled performance. More than anything this depends on the particular application you're running, and where the bottlenecks lie within it.

If you look at a past example where performance WAS doubled, like the 8800GTX, let's compare that to the previous gen flagship, the 7900GTX. The 8800GTX had almost exactly twice the GFLOPs of the 7900GTX, even taking into account the nearly useless MUL op. The 8800GTX had 69% more memory bandwidth, and get this, only 33% more pixel fillrate, and 18% more bilinear texture fillrate.

The GF100 is more of an improvement in raw specs over the GTX 285 than the 8800GTX was over the 7900GTX. So doubling performance is more than possible.

Where in the post you're quoting I say that you have to double everything to double performance? I'm aswering a specific question.

And you can't compare G80 with previous generation, as it's a completely different architecture. Starting by the unified shader processors (instead of units that only could calculate vertex or pixel shaders), with a completely different architecture, and the same for TMUs and ROPs.

Again, I've never said that doubling is not possible (why is everybody putting that words in my mouth? It's at least the 3rd person who says that, and I'm starting to be tired or repeating it). You can read it yourself in my post quoted by Chumbucket843 (that I should add it's taken from a conversation including more posts before and after).

I have only said that there is not a single evidence which grants that the GT300 is going to be more than twice the performance of GT200.

But oh, well. If all of you are getting hurt by hearing it, I'll correct myself and let's finish with this: "GT300 is going to be obligatory at least 2x the performance of GTX285, and probably more". ¿Happy there?

EDIT: I have edited the former paragraphs to give a much more accurate response.

[Manicdan]

Of course you haven't, A0 is usually in-house only. Only chip I can think of that released as an A0 from NVidia is the NV15. Usually it takes a few revisions before they can release.

Also, the R600 DID take several respins before it could release if you remember. You're talking about a card that was 6 months+ late.

the R600 was bigger and hotter than my epeen, biggest think ati ever made, and was on the wrong process i think.

[Calmatory]

I wouldn't make conclusions yet, we haven't seen any GPGPU results for larrabee, or pricing.

however even if larrabee is slightly less powerful, I could still see businesses opting for it due to lower power usage and it will probably be cheaper

Wasn't it reported to be 300W 600mm² thing like a year ago? Yeah, rumours based on guesstimates based on rumours, I know.

[Cybercat]

And you can't compare G80 with previous generation, as it's a completely different architecture. Starting by the unified shader processors (instead of units that only could calculate vertex or pixel shaders), with a completely different architecture, and the same for TMUs and ROPs.

Of course it was a completely different architecture, as is the GF100. It may not be to the same extent as the G80 enjoyed, but it makes up for that by increasing raw specs more than the G80 did.

It really doesn't bother me when people say the GF100 won't perform as well as such-and-such or whatever, because no one knows, and everyone's entitled to their opinion. The main thing that bothers me is how much importance you place in ROPs, TMUs and bandwidth, when those are insignificant factors in games that are GPU-limited. Granted, there aren't that many of those anymore, thanks to consoles and the perceived threat of piracy.

[Chumbucket843]

I didn't miss that factors. They simply don't take any part in anything that I've said. And when it take it, I have mentioned them and considered them. Take the "trouble" of reading my posts and trying to understand them before quoting me, please, to not put things in my mouth.

And I don't know how to use HD5870 to know how games are shader bottlenecked since the proportion in which they have improved shader processing power it's the same than texture processing power, rasterizing operations processing power, and so.

There are more things involved in the 3D rendering process apart from shaders and memory bandwidth.



Yeah, no doubt. But I think you have misunderstood them when you have the idea that they are talking about a clock of 1500MHz for the GPU core.

my reference to the 5870 was to show that rops are where they should be. too much and youre just wasting die space. they are running games at 7680x3200. the rop's were added to help texture filtering quality which wont double the performance in either gpu. if you dont believe me look at the ratio of shaders to rops over the past 5 years.

this is the statement i was referring to:

Consider that HD5870 is exactly double the HD4890 (+100% everything at the same clocks) except bandwidth (aprox. +30%) and it's far from double the real world performance (that's one of the most recent proves that doubling everything doesn't mean doubling real world performance), and NVIDIA is not even doubling processing units.

i responded to this part of your statement about shader clocks and you somehow got the idea i was talking about core?

And regarding clock speed, I would take it like talking about the shaders clock. I wouldn't expect much higher clocks than GTX285, if at all.

[Farinorco]

It really doesn't bother me when people say the GF100 won't perform as well as such-and-such or whatever, because no one knows, and everyone's entitled to their opinion. The main thing that bothers me is how much importance you place in ROPs, TMUs and bandwidth, when those are insignificant factors in games that are GPU-limited. Granted, there aren't that many of those anymore, thanks to consoles and the perceived threat of piracy.

That's exactly what I was trying to say.

I have never said "GF100 won't perform as well as papapa".

Exactly that's what I'm talking about.

Somebody said "GF100 is going to perform at least twice as well as" and I asked him "Why? What's the reason why do you think that? What info which we have now lead you to take that for granted?".

And then some of you started quoting me puting words in my mouth.

But oh, you know what? The guilt is all mine:

I should have started to reply "I know it. I didn't say otherwise. Read it again" since the very beginning.

[Farinorco]

my reference to the 5870 was to show that rops are where they should be. too much and youre just wasting die space. they are running games at 7680x3200. the rop's were added to help texture filtering quality which wont double the performance in either gpu. if you dont believe me look at the ratio of shaders to rops over the past 5 years.

this is the statement i was referring to:



i responded to this part of your statement about shader clocks and you somehow got the idea i was talking about core?

Processing units. CUDA cores are processing units. Texture units are processing units. Raster Operation Processors are processing units. So no, they are not "doubling processing units".

And regarding the clocks, obviously. If you understand it like shaders clock I don't know how it's an argument to say that they have doubled processing units power.

EDIT: And from my part, discussion about what I've said or left to say is over. My (at the present time favourable, and I think not unrealistic) opinions about GT300 are pretty clear at posts in previous pages (some of them quoted on this one), even when some people is absolutely determined to misunderstand them.

[Chumbucket843]

Wasn't it reported to be 300W 600mm² thing like a year ago? Yeah, rumours based on guesstimates based on rumours, I know.

this is from anandtech. they will hit a power wall if it clocks high enough before they have a huge die.

At 143 mm^2, Intel could fit 10 Larrabee-like cores so let's double that. Now we're at 286mm^2 (still smaller than GT200 and about the size of AMD's RV770) and 20-cores. Double that once more and we've got 40-cores and have a 572mm^2 die, virtually the same size as NVIDIA's GT200 but on a 65nm process.

The move to 45nm could scale as well as 50%, but chances are we'll see something closer to 60 - 70% of the die size simply by moving to 45nm (which is the node that Larrabee will be built on). Our 40-core Larrabee is now at ~370mm^2 on 45nm.

[LordEC911]

the whitepapers said there will be future versions with less double precision for gaming. that probably wont happen this gen though. no one is expecting 3x performance in games. 2x faster could be possible.

Please quote or tell me what page that is on. I have read through the whitepaper 3 times and haven't seen ANY mention of that.

[Chumbucket843]

Please quote or tell me what page that is on. I have read through the whitepaper 3 times and haven't seen ANY mention of that.

there are 4 papers on nvidia's website.
pg 19 of "fermi looking beyond graphics"

Fermi’s only vulnerability may be its attempt to combine world-class graphics performance with generalpurpose
compute performance in one chip. With three billion transistors, a Fermi GPU will be more than
twice as complex as NVIDIA’s existing GT200 chips. At some point, if not now, features intended to
boost compute performance may compromise the chip’s competitive position as an affordable graphics
processor. At that juncture, the architectures may have to diverge — especially if the professional
market grows larger than the consumer market.