Nvidia Pascal

[xMishax]

Is it going to be groundbreaking jump in GPUS history or what? There is so much hype around the internet about it. I know that it will support HBM2 but what else about it?

[Ciccioo]

it will most probably be released in a very cut down version because nvidia doesn't need any performance boost

it will have like 5 cuda cores

it will still beat the fury X

by far

sad

[slightlyjaded]

Is it going to be groundbreaking jump in GPUS history or what? There is so much hype around the internet about it. I know that it will support HBM2 but what else about it?

Last I heard it got put way back because their approach to 3D memory just wasn't turning out to be economically viable. I've heard stuff all about pascal since.

[Unlikelyjoker]

INB4 They only release the 1060 because its more than enough.

[Glenwing]

Pascal is still very far away, and it seems to be designed as an improvement in compute, not so much in gaming.

HBM is not really a big deal for gaming, as memory bandwidth is not typically an issue unless the bus width is disproportionately small for the power of the GPU and memory capacity. I think the performance of the Fury X has demonstrated this well enough.

Compute applications are more memory intensive than games though, and it's said that double precision will be coming back on Pascal after it didn't appear heavily on Maxwell. According to NVIDIA Pascal will have 10x the compute performance of Maxwell (which sounds cool but it's not particularly impressive since Maxwell is horrible at FP64, worse than even a crippled Kepler core let alone a fully enabled one). If we knew whether that FP64 capability was 1/3 or 1/4 FP32 then we could estimate gaming performance against Maxwell too, but we don't have that information. If it's 1/3 FP32 like Kepler then it will only be about the same performance as Maxwell, at equal clock speeds and core counts anyway, but we don't have any information about those numbers either, so it's all up in the air really. If Pascal is grouped into 128-core units like Maxwell though, then it's more likely 1/4 or 1/2.

[Curufinwe_wins]

Pascal is still very far away, and it seems to be designed as an improvement in compute, not so much in gaming.

HBM is not really a big deal for gaming, as memory bandwidth is not typically an issue unless the bus width is disproportionately small for the power of the GPU and memory capacity. I think the performance of the Fury X has demonstrated this well enough.

Compute applications are more memory intensive than games though, and it's said that double precision will be coming back on Pascal after it didn't appear heavily on Maxwell. According to NVIDIA Pascal will have 10x the compute performance of Maxwell (which sounds cool but it's not particularly impressive since Maxwell is horrible at FP64, worse than even a crippled Kepler core let alone a fully enabled one). If we knew whether that FP64 capability was 1/3 or 1/4 FP32 then we could estimate gaming performance against Maxwell too, but we don't have that information. If it's 1/3 FP32 like Kepler then it will only be about the same performance as Maxwell, at equal clock speeds and core counts anyway, but we don't have any information about those numbers either, so it's all up in the air really. If Pascal is grouped into 128-core units like Maxwell though, then it's more likely 1/4 or 1/2.

But if compute prwformance increases dramatically and their own architecture doesn't take a massive shit in this new cycle, then it should be a very sizable jump in gaming performance.

Also nvidia is releasing Maxwell compute cards soonish so they must have done something with the admittedly terrible double precision performance.

Btw less than a year for release isn't that far away in my mind.

[Glenwing]

But if compute prwformance increases dramatically and their own architecture doesn't take a massive shit in this new cycle, then it should be a very sizable jump in gaming performance.

Also nvidia is releasing Maxwell compute cards soonish so they must have done something with the admittedly terrible double precision performance.

Btw less than a year for release isn't that far away in my mind.

No, because games don't use FP64 at all. Keep in mind when NVIDIA says "compute" they mean FP64. Maxwell was reduced to 1/32 FP32 compared to Kepler's 1/3. Maxwell simply doesn't have more FP64 cores, it's not like Kepler where they were just locked. Even the Quadro M6000 has 1/32, the K6000 or K5200 and Kepler-based Teslas walk all over it in compute. You could theoretically just remake Maxwell with more FP64 cores, bring it up to, say, 1/4 FP32. With that alone you'd have an 8x compute performance increase with zero gaming performance gain. I imagine Pascal will have 1/4 FP32 and then either some architectural improvements, additional cores, or a new fabrication process that allows for higher clockspeeds, or a combination of those, to get the last 2x. That'll be a pretty mild increase to the gaming performance, around 25% or so.

[Curufinwe_wins]

No, because games don't use FP64 at all. Keep in mind when NVIDIA says "compute" they mean FP64. Maxwell was reduced to 1/32 FP32 compared to Kepler's 1/3. You could theoretically just remake Maxwell with more FP64 cores, bring it up to, say, 1/4 FP32. With that alone you'd have an 8x compute performance increase with zero gaming performance gain. I imagine Pascal will have 1/4 FP32 and then either some architectural improvements, additional cores, or a new fabrication process that allows for higher clockspeeds, or a combination of those, to get the last 2x. That'll be a pretty mild increase to the gaming performance, around 25% or so.

Again. Even 25% isn't what I would consider mild. Anyways gm204 is being made into a compute chip at the moment, so while it would be possible to claim it is 10x Maxwell that wouldnt quite be on par with their own methodology.

Also we still are looking at a 4x increase in mixed precision (which not particularly impressive on its own).

Honestly I also don't agree that hbm hasn't shown a large improvement... A 980 overclocked can come in pretty near on synthetics and 1080p gaming to the fury x with the same amount of memory but it gets obliterated at higher resolutions. Also the fury x does preform above its memory and compute class (the compute (sp) ratio between amd and nvidia cards of the last generation has been pretty near steady at 1.3-1.4 for cards with similar in game performance) at higher resolutions where it should struggle instead.

All in all I am personally pessimistically expecting about a 50-150% improvement in current games, with obvious increases for future titles. That said I hope the smaller process doesn't turn out to be too poor of an overclocker (as for some reason nvidia insists on selling cards at base clocks so far below their real potential).

[Glenwing]

Again. Even 25% isn't what I would consider mild. Anyways gm204 is being made into a compute chip at the moment, so while it would be possible to claim it is 10x Maxwell that wouldnt quite be on par with their own methodology.

Also we still are looking at a 4x increase in mixed precision (which not particularly impressive on its own).

Honestly I also don't agree that hbm hasn't shown a large improvement... A 980 overclocked can come in pretty near on synthetics and 1080p gaming to the fury x with the same amount of memory but it gets obliterated at higher resolutions. Also the fury x does preform above its memory and compute class (the compute (sp) ratio between amd and nvidia cards of the last generation has been pretty near steady at 1.3-1.4 for cards with similar in game performance) at higher resolutions where it should struggle instead.

All in all I am personally pessimistically expecting about a 50-150% improvement in current games, with obvious increases for future titles. That said I hope the smaller process doesn't turn out to be too poor of an overclocker (as for some reason nvidia insists on selling cards at base clocks so far below their real potential).

I guess we'll just have to wait and see.

[Curufinwe_wins]

I guess we'll just have to wait and see.

indeed. Who knows when Volta is coming out amd might have a competitive gpu....