I wanted to just click WTF but I didn't want to shoot the messenger. Seriously. WTF
Would Volta be better than Pascal at graphics? Given that Volta puts more emphasis on compute while Pascal is more narrowly focused on graphics, it could easily be worse. \
Titan V is available today, if you want to pay for it ($3,000). So there have been comparisons of Volta versus Pascal.
There’s no denying that NVIDIA TITAN V smashes GeForce GTX 1080 Ti FE gaming performance at 4K and 1440p. We experienced performance differences up to 40%, most were around the 30% mark, with varying games below or above that average.
Yeah, Volta will be a decent bump over Pascal, if we are just looking at Architecture vs Architecture and not specific card versus specific card.
Yeah, but you won't see a Volta card tricked out like the Titan is, I would surprised if they managed a 20-25% bump. I think it will be more like 10-15% on average with the higher end cards slightly more.
So basically, it's an 815 mm^2 die on a supposedly better process node offering 30% more performance than a 471 mm^2 die. If it takes 73% more die size to offer 30% more performance, that's not an architectural advantage. Scale it down to the same die size (which might still cost more for Volta because of the process node) and will it still be faster?
Couple of things here:
nVidia has never shied away from big dies.
The Volta in question has a lot of whatever-the-heck tensor cores are... which are presumably not all that useful for gaming, and presumably aren't affecting the gaming benchmark tests that have been done. Remove those, die size ~probably~ comes down to something more manageable. Speculation on my part, I know, but probably more true than not.
It could just be a matter of scale: Titan Xp has 3840 shading units, V has 5120 shading units -- which is 30% more shading units for 30% more performance.... which sounds like a wash, except your getting that 30% more for no net increase in power draw, so there's the combined benefit of the 12nm vs 16nm node and Volta vs Pascal architecture. How much is node and how much is architecture - I couldn't say.
I'm not claiming that Volta would certainly be worse at graphics than Pascal. Maybe it would be better. Maybe it would be about even. Maybe if you strip the extra compute stuff out of Volta, it would just be Pascal without any meaningful changes.
But there's no point in doing a ton of work to make a new generation unless it will be a lot better than the old. Even if they did do all of that work, would a Volta-based GeForce be better than a Pascal ported to 12 nm FFN based GeForce? If not, then it could make more sense for a new generation to do the latter than the former.
As for tensor cores, I could tell you what those are. They're dedicated logic for doing a matrix multiply-add on 4x4 matrices of half-precision (16-bit) floats. And yes, that's useless for graphics. And useless for just about everything else, for that matter.
I wanted to just click WTF but I didn't want to shoot the messenger. Seriously. WTF
Would Volta be better than Pascal at graphics? Given that Volta puts more emphasis on compute while Pascal is more narrowly focused on graphics, it could easily be worse. \
Titan V is available today, if you want to pay for it ($3,000). So there have been comparisons of Volta versus Pascal.
There’s no denying that NVIDIA TITAN V smashes GeForce GTX 1080 Ti FE gaming performance at 4K and 1440p. We experienced performance differences up to 40%, most were around the 30% mark, with varying games below or above that average.
Yeah, Volta will be a decent bump over Pascal, if we are just looking at Architecture vs Architecture and not specific card versus specific card.
Yeah, but you won't see a Volta card tricked out like the Titan is, I would surprised if they managed a 20-25% bump. I think it will be more like 10-15% on average with the higher end cards slightly more.
I'm confused. The current Volta card is a tricked out Titan. Or do you mean we won't see another Titan based on Volta... because we already have a Titan based on Volta.
Historically, in the past, x80Ti editions and Titan cards have shared a lot (usually the former has slightly fewer cores running at slightly faster clocks, and maybe only half the precision, but both based on the same die).
Your right that the 1250, 1260, 1270, and 1280 won't be the same thing as Titan V. But they never have on any previous generation either, and no one here is claiming otherwise.
But there's no point in doing a ton of work to make a new generation unless it will be a lot better than the old. Even if they did do all of that work, ...
Well, regardless of how well Volta plays out versus Pascal, all that work you talk about... nVidia has already done it. Volta already exists, it's not like they have to continue a bunch of R&D to bring it to market. It's already on the market, in a few different SKUs.
But there's no point in doing a ton of work to make a new generation unless it will be a lot better than the old. Even if they did do all of that work, ...
Well, regardless of how well Volta plays out versus Pascal, all that work you talk about... nVidia has already done it. Volta already exists, it's not like they have to continue a bunch of R&D to bring it to market. It's already on the market, in a few different SKUs.
Just not under GeForce branding.
And perhaps more to the point, not under GeForce pricing, either. Selling GPUs based on an 815 mm^2 die for $300 each is not a way to make a profit.
Comments
But there's no point in doing a ton of work to make a new generation unless it will be a lot better than the old. Even if they did do all of that work, would a Volta-based GeForce be better than a Pascal ported to 12 nm FFN based GeForce? If not, then it could make more sense for a new generation to do the latter than the former.
As for tensor cores, I could tell you what those are. They're dedicated logic for doing a matrix multiply-add on 4x4 matrices of half-precision (16-bit) floats. And yes, that's useless for graphics. And useless for just about everything else, for that matter.
Historically, in the past, x80Ti editions and Titan cards have shared a lot (usually the former has slightly fewer cores running at slightly faster clocks, and maybe only half the precision, but both based on the same die).
Your right that the 1250, 1260, 1270, and 1280 won't be the same thing as Titan V. But they never have on any previous generation either, and no one here is claiming otherwise.
Just not under GeForce branding.