GT300 News

[Tamlin]

According to Fudzilla, it will be demonstrated today. :)

http://www.fudzilla.com/content/view/15741/1/

[Paradigm Shifter]

Talked about, maybe. Huang will talk about how awesome nVidia are and how bad ATi are. Actual silicon demonstrated? Possibly, but I'll believe it when I see it.

(In four hours, I'll happily eat those words if he does actually have working silicon on demo...)

[Frag Maniac]

"Fermi has a lot of cache and supports instructions that use to be common only for CPU and many people believe that this is a hybrid between a graphics card and a CPU. This might be the direction that Nvidia will be heading."

Makes sense given that DX11 will treat a GPU more like a CPU with multi threading and physics. I get the feeling ATI is going to wish they'd gone this direction with their 5000 series and they may soon. Perhaps the 6000 series.

[Tamlin]

So it was a cGPU. :) Interesting for someone that loves new tech like me.

AMD have worked for years with their Bulldozer and Intel with their Larrabee, so the concept is old, but at least someone got out something concrete.

It won't do anything for DX11 I think. DX11 can runs some tasks on shaders via Directx compute and a cGPU won't make any difference. All computers have both cpu and gpu, so its more a matter of reducing overhead in programs that can handle it. It looks like Nvidias path will be to introduce cuda plugins for common program, so they will accelerate faster. Question is how much extra it will give if programs will be made more for DX11 and opencl. I find it interesting, since cGPU can reduce a lot of overhead and streamline GPGPU much more. It won't give much extra for gaming though, I believe. But gaming isn't all we do on computer.

This is only the beginning. We've seen little of what Fermi can do yet, so I wouldn't jump to any conclusions as for now. This conference had mostly to do with GPGPU and not so much for gaming in general. I bet Nvidia will focus more about what it can do for gamers later.

Looks like [H] was right though, that it comes in Q1 2010. By then, we might get some news about Larrabee and Bulldozer also. I think that Fermi might push the releases a bit forward if everything goes well.

2010 might be a fun year for GPGPU with opencl and dx11 when cards come that brings GPU and CPU closer together. :D

[Paradigm Shifter]

Most important bit of that keynote?

Real-time ray-tracing.

Within a year.

Win.

Other thoughts:

That said, this looks awesome for computational purposes... we thought the GT200 cards folded well...

...still no idea about gaming performance, though. I guess that'll depend on how much of that core is taken up with transistors to deal with all the DirectCompute/OpenCL/C++/C/Fortran/Java/Python things it's also supposed to be able to do.

These all sound amazing, but they're gonna need programs to take advantage of it.

No real detail on the graphics, though, although it's likely to be a monster just from the specs revealed. Nothing that obviously competes against EyeFinity, either. I guess nVidia are working on the idea of Matrox TripleHead working better with their cards. Who knows?

Pics of card. However, forgive me for not believing that it'd work if you plugged it in. It's also Tesla, rather than a proper GPU.

OK, the Radeon 5000 series can't do as much 'DirectCompute' stuff (on paper) as this is supposed to, but ultimately ATi is the graphics division of AMD. nVidia only has vid cards, so they need to do "everything" while AMD have Bulldozer and Intel have Larrabee...

...

Still, exciting times. Exciting times... :)

[Paradigm Shifter]

Update:

I've found the nVidia whitepaper about Fermi's architecture. It's linked to in this article: http://vr-zone.com/articles/nvidia-fermi--arriving-in-q1-2010/7786.html?doc=7786

The 48 ROPs thing has me worried, a bit. Sure, 512 shaders is more than double that of GT200, and four times that of G80, but the ROPs have only increased by 1.5x...

I'm currently thinking that Fermi will be a GPGPU monster, but not what it could be for games. Maybe 1.5-2x a GTX285. Going by ROPs we're looking at 1.5x, but the shader boost will probably mean it'll have excess shader power for normal GPU load... resulting in it probably be about 1.8x a GTX285. Of course, that also depends on clock speeds.

The ECC and on-chip DRAM are interesting as well - they're only really useful for cGPU/GPGPU, not that useful for graphics... or not as useful for graphics. Seems nVidia is betting a lot on GPGPU this time around. If the market picks it up, they gain a lot of ground over ATi, as their GPGPU implementation is easily a generation ahead of ATi. If the market is cautious and doesn't jump in with both feet, nVidia could find itself in trouble.

...

Given that Fermi supports DirectCompute and OpenCL as well as CUDA... I'm hopeful that the market will pick the open standard for GPGPU rather than risk potential fragmentation of the market. Of course, it depends how well Fermi does each of them. The interesting thing about the shader count is that it's probably got more than enough to do high-level PhysX calculations in-shader while still having more than enough grunt for the graphics side...

...

Also a bit worried about that line "DirectX 11 is of course supported, though Tesselation appears to be software driven through the CUDA cores."... that smacks of software T&L back in the DX7/8 transition days. :( Is this a sign that Fermi isn't going to comply 100% with the DX11 specification?

[Frag Maniac]

@Tamlin,
Irregardless of the acronyms used and technical definitions, I always thought of Larrabee and Bulldozer as being CPUs with minimal graphics capability onboard, vs a monster GPU that has CPU tendencies. To me that is not the same at all. The main thing that's kept Larrabee and Bulldozer from being talked about a lot by the gaming community is they know there's no gaming potential in the current projects for both.

I also think such a GPU should make well use of DX11 since DX11 will have much more going on inside the GPU. Of course it depends on how well devs make use of DX11 in games.

PS may have made the most telling comment regarding a possible imbalance in ROPs to shaders. However this architecture combined with DX11 may require less ROPs, though I must say it has me worried too. Also, from what I've read the memory bandwidth has not vastly increased as have the transistor and shader count.

Nvidia have on occasion made GPUs that don't make much sense for gaming, like the 5700 that had way higher clock speed than the Ti 4200, but couldn't perform any better than it due to very minimal memory bandwidth. I sincerely hope they haven't gone down that road again with the GT300 series because there's no way I'm sticking with ATI.

[Tamlin]

@Tamlin,
Irregardless of the acronyms used and technical definitions, I always thought of Larrabee and Bulldozer as being CPUs with minimal graphics capability onboard, vs a monster GPU that has CPU tendencies. To me that is not the same at all. The main thing that's kept Larrabee and Bulldozer from being talked about a lot by the gaming community is they know there's no gaming potential in the current projects for both.

Its been back and forth about how much CPU and how much GPU (the same with launch dates). Newest might be ATI Llano:

CPU + GPU is the ultimate fusion. If word on internet is anything to go by then we’ll see the first AMD Fusion chip by mid 2011. Dubbed as Llano, the combination of AMD processor and ATI graphics will be a native chip with either a dual or quad core, allowing users to derive tremendous computing power from the fusion.

http://techlime.com/graphics-card-memory-power-supply-cooling-devices/amd-llano-fusion-cpu-gpu-coming-2011

I also think such a GPU should make well use of DX11 since DX11 will have much more going on inside the GPU. Of course it depends on how well devs make use of DX11 in games.

I agree, DX11 will benifit most from GPU with Directx compute especially for GPGPU. For games, I believe that tesselation will be big and give us a lot better textures.

Nvidia have on occasion made GPUs that don't make much sense for gaming, like the 5700 that had way higher clock speed than the Ti 4200, but couldn't perform any better than it due to very minimal memory bandwidth. I sincerely hope they haven't gone down that road again with the GT300 series because there's no way I'm sticking with ATI.

I think Nvidia should have learned that by now. :) We've seen little of its gaming capabilities yet, but they will come when Nvidia choose to focus more on that. For now, it was more about programs (realtime rendering ray-tracing programs for professionals and similar that Nvidia offered)
http://www.trustedreviews.com/graphics/news/2008/08/15/nVidia-Demonstrates-GPU-Acclerated-Ray-Tracing/p1

Basic software that shows that GPGPU in general can do.

What I am curios about, is what the cGPU from Nvidia can do extra besides software and what benifits the cGPU can give us compared to an ordinary CPU+GPU setup (Like I7 with gtx285). :)

[Frag Maniac]

Argggh, I don't like that Nvidia are catering first to the business crowd with these workstation graphics accelerator demos instead of their bigger market niche, consumers. As far as ray tracing goes, I agree with John Carmack that Intel are being a bit unrealistic wanting to use exclusive ray tracing vs a reasonable mix of using it for the more obvious light and shadow effects and rasterizing for the rest. No one would even see the difference unless they wandered slowly through every map and gawked at everything as if in a slow poke adventure game anyway.

That being said, it is rather remarkable that Nvidia somewhat outdid Intel by showing an HD full ray tracing demo at 30 FPS. Carmack had speculated some time ago it may take a 3GHz GPU to pull that off. What it really comes down to is whether the average gamer notices the difference enough to warrant it and the expense of the hardware R&D and game development behind it, not to mention if consumers would want to pay potentially higher prices because of it. I also fear it would further perpetuate the trend of devs favoring eye candy over substance in story and gameplay in their titles.

[whismerhill]

For games, I believe that tesselation will be big and give us a lot better textures.

herrr .... not really, textures are not improved at all by tesselation ..
it's the number of polygons that's increasing while keeping less triangle data & compute time around ..
which means better 3D, better "bumpmap" (kind of) ...

[Paradigm Shifter]

Well, tesselation means more realistic world modelling without upping the polygon count of the world and characters to ridiculous levels - see some of the enhanced Crysis model mods that PCGH has showcased occasionally that up the characters to a million polys... framerate nosedive... so in that respect I can't wait for tesselation as it can only mean things looking better.

Frag, I don't understand nVidia's demoing of this for business/CAD etc either - looking at nVidia's revenue figures, they made $1bn. last year... only $10mn. of that was from Tesla. That said, more people will be willing to take Tesla and CUDA up, now that the chips have ECC in them. That was the stumbling block before now.

Did you notice that that ray-tracing demo wasn't run on Fermi? In fact, it wasn't run on GT200 either, from everything I can figure out. From their spec sheet it seems like it's effectively a 7950GX2? :? That can't be right... :? I'm confused.

From what I've read, there was no confirmed working silicon at that demo for Fermi. Even the pro-nVidia camp at a couple of forums I lurk on are doubtful that Fermi is working... or at least working at clockspeeds that make it worth buying. Still, nVidia have got a few months before the crunch comes. It's entirely feasible that this paper launch (because make no mistake, that's what it was) was a kind of 'look what you'll be able to do, give us more time' thing.

Basically, Fusion, Larrabee, Fermi... all approach the same idea from different directions: Intel and AMD are approaching a unified CPU/GPU from the CPU side as that's what they do. nVidia is approaching a unified CPU/GPU from the GPU side as that's what they do.

nVidia have done an ATi on the GDDR5 front. They've starved the GPU of memory bandwidth. The 2900 series had too much memory bandwidth, while the 4870 was better balanced. The 5000 series has increased speed, but hasn't increased bandwidth enough. The 5870 gains even more speed when you overclock the VRAM a bit, from what little overclocking of it I've seen. Time will tell on the mix of shaders/ROPs - ATi have had excess shaders for a while, but given that they do shader-based anti-aliasing that's not too surprising.

nVidia, you're totally right, has made some odd decisions in the past with regards to their GPUs. The GeForce FX series was a complete mess from the get go. It was late. It was a new design on a new process which always leads to problems. This time, nVidia cut their teeth on the 40nm process with some mobile chips, but I don't think it's enough to give them a solid idea on how their architectures 'play' on a 40nm process.

If GT300 isn't everything for gaming, you can still get a GTX285/295/whatever, Frag. It's not like they're bad cards. Unless you're wanting to game on a 30" panel then except in a couple of instances there is no need for much more than a GTX285. If it wasn't for the fact that their latest drivers break PhysX when ATi cards are present, I'd probably go 5870 (perhaps CF when EyeFinity and CrossFire play together) with EyeFinity and an 8800GT I've got in another rig for physics... at least until they decide that having it platform-agnostic is better (I don't see the problem - it'd be like the ultimate teeth-kicker - let it run on ATi cards as well, but emblazon 'PhysX by nVidia' by the PhysX option in the game menu, and the Control Panel, etc etc... :lol:)

[Paradigm Shifter]

http://www.techpowerup.com/105052/NVIDIA_Fermi__Tesla_Board_Pictured_in_Greater_Detail_Non-Functional_Dummy_Unveiled.html

http://www.semiaccurate.com/2009/10/01/nvidia-fakes-fermi-boards-gtc/

Ouch.

Now, while I'm certainly not going to believe anything Charlie Dimerjian says about nVidia without basically having the proof sitting right in front of me, when TechPowerUp say it, I'm more inclined to believe it.

Justification for my suspicions about that card Huang was holding up wasn't a working card, despite what he said.

[Mach1.9pants]

Yeah the pic that got my was the one showing the back end of the back side of the card.
The big blobs of solder for power connectors don't line up (like they do on my GTX280 & 8800GTX) and the 6 pin doesn't even have any!
And there appears to be 2 stickers just cut through in the d!

LOL I won't be waiting for this card, like my original plan, I reckon late Q1 at the earliest and it may not even have 'nFinity'

[Frag Maniac]

Honestly, I think the talk about the mock-up and deceit are overblown. Nvidia were clearly intending primarily to demo the GPU itself onscreen, not unveil a card as if ready for market. The article explains they had trouble getting some of the parts for the production model, so they probably cobbled something together with less than ideal parts for onscreen demo purposes on a non production board, and showed a mock-up that looks closer to the final product on stage.

Keep in mind people, when Nvidia said "This puppy is Fermi", they're talking primarily about the radically changed GPU architecture itself. The actual workstation card using that chip is in fact called Tesla, as Nvidia cards of that type have been previously called. I suspect most of this knee jerk reaction to what is primarily a bit of an embarrassing setback, will die down once the Fermi and Juniper based cards are benched and ready for market.

While reading through that article where the guy was showing someone else's pics of the mock-up, I couldn't help but think he'd had too much of his favorite coffee roast while spending several paragraphs ranting on and on about the screws and power connectors on the mock-up. The guy just sounds totally full of himself. Don't pat him on the back people, he might spill some of that coffee on his E-penis.

[Mach1.9pants]

I think it was deceitful, it was put out as here is a Fermi card. It was not, it was a mock up and they should have said so. I just think lying by omission is just as bad, a real dirty politician type trick!

http://www.fudzilla.com/content/view/15798/1/ 'confirmed fake' if Fud can be taken as confirmed ;)

[Gilly]

I think it was deceitful, it was put out as here is a Fermi card. It was not, it was a mock up and they should have said so. I just think lying by omission is just as bad, a real dirty politician type trick!

http://www.fudzilla.com/content/view/15798/1/ 'confirmed fake' if Fud can be taken as confirmed ;)

he was lying, i could of just held up a bundle of wires, said "hi this is Fermi card, honest there is a PCB under there somewhere", and everyone would of laughed, but at least i wouldnt of been lying

[Paradigm Shifter]

That extremely poor effort on the mockup card makes me doubt any information nVidia provides about Fermi until I see cards in the wild.

Also, reports in other forums indicate that that system running 'Fermi' doing that demo was, in fact, G200b based cards. Not Fermi.

We'll see when they're released. Currently I'm not expecting them any time before mid/late-Q1 2010. We might get a paper launch in January. We might even get a couple of cards to review sites. But I doubt there will be any in the wild until February at the earliest. Which is a shame, as I'd have liked to see the 5870's maybe drop in price a bit before Christmas. Oh well, I suppose in the end holding off on the purchase saves me money...

...

Testing/engineering sample cards are usually covered with wires. He could have held up a GTX285 with a coupla hundred wires soldered on and we wouldn't have known different as long as the chip wasn't seen. But the effort into that mockup was just so terrible it'd have been better had they not bothered.

...

Although I must confess a strong sense of schadenfreude watching all the fanboys from both sides go into spasms... :lol:

[whismerhill]

on this page : http://www.hardware.fr/articles/772-7/nvidia-fermi-revolution-gpu-computing.html

there's a graphic which shows: "theoretical calculation power"
it's based on the fact that Fermi CUDA Cores would run @1600Mhz which is pretty conservative.
As you can see the progression from G200 is enormous
And even though cypress is superior in Floating Point simple precision, one has to keep in mind that Cypress vec4+1 architecture is less efficient and can't always be fully utilised ...

In double precision, Fermi wins.


I thought this could interest you guys, I translated the interesting bits to the best of my knowledge.

[Frag Maniac]

Also, reports in other forums indicate that that system running 'Fermi' doing that demo was, in fact, G200b based cards. Not Fermi.

If those were truly reports (eg: substantiated vs rumor), they'd have been backed up with some kind of proof vs sounding like more mere speculation. Let's be realistic here, you can't say the demo was run on a GT200 unless you know it was without sounding just as deceitful as you're accusing them of being.

It's obvious Nvidia were caught with their pants down, but I really think they were more trying to avoid people thinking they couldn't pull off Fermi at all by holding up that mock card after an unforeseen problem getting parts in time. The last thing you want to risk at such a conference is potential customers thinking you lack the confidence to provide what you've promised.

Everyone seems to be glossing over that they mentioned they couldn't get all the parts in time to build a complete working card, or that Fermi is in fact the GPU itself, not the entire card, which is Tesla. A political dodge none the less to say this is Fermi, when technically the card is supposed to be Tesla, but still, more a clever speech tactic than a lie.

In all of this what is perhaps most absurd though is despite it now being rather obvious Fermi and Tesla are aimed at HPC vs gaming, many including some of the more knowledgeable here are still clinging to the idea of using Fermi for gaming. Do you really think after what you've heard that any card with a Fermi is going to be anywhere near consumer grade affordable or well driver supported for games?

How many times does it have to be said, JUNIPER is going to be the consumer level chip. Regardless of how long it actiually takes Fermi to hit the market, we may see a scaled down GPU with all the gaming power and none of the HPC power before that happens at a much lower price.

I tell ya man, I'm really getting sick of the angst here over nothing more than what is mostly speculation. I thought people on this forum were more intelligent than that. It just goes to show that even some of the smarter consumers lose their heads when they don't look at things from a business perspective. :roll:

[whismerhill]

Actually hardware.fr mentioned several times that nvidia wouldn't comment on fermi 3D performances but they seemed to imply we could expect fermi to be the high end of the spectrum
At least that's how I understood it...

"Fermi", "Tesla" or "Fermi Tesla" ?
well hardware.fr mentioned that there's several "Tesla" out there:
-Fermi is the codename for the new GPU
-Tesla was the codename for the GT200
-Tesla also became a brand for Nvidia products aimed at the "massive computing market"

don't know if it's what you meant...