Nvidia thinks "Pascal is just unbeatable" and already decided not to move Volta ahead

[laminutederire]

2 hours ago, Taf the Ghost said:

 RTG's Navi is going to really shift the market around.

Let's hope so, the market suffers from the stale state occurring! 

[Dylanc1500]

7 minutes ago, valdyrgramr said:

9370 and 9590 at launch were priced high.

$999 for the 9590 at launch

[Jito463]

1 hour ago, valdyrgramr said:

9370 and 9590 at launch were priced high.

Yep, and that was a mistake.  They had to drop the price down pretty quickly, as I recall.  They pushed the architecture to it's very limits, and thought they could justify the extra cost because of the clock speed, but the market proved them wrong.

 

It wasn't like the old Socket 939 days, when they could justify a $1,000+ CPU because they actually had the better architecture.

[LAwLz]

2 hours ago, laminutederire said:

Where have you seen those figures though?

Basing my calculations on this review:

With power saving mode, you get around 25% more performance per watt for the 1070, not 40%. That's nearly half your figures.

Besides, I said closer, compared to Polaris etc. At least in Hitman there is a significant improvement in perf/watt under power saving mode (13% improvement in that game). I don't have the figures for every game, but that's a probably valid trend regarding the improvement from Polaris to vega (which could be worse vs the 1070 as hitman is a good game for Amd, but closer anyway)

I got my numbers from TechPowerUp which takes 20 games into consideration (not just one) which I think will be a pretty good average without any cherry picking. But to be fair, they did not use power saving BIOS. But on the other hand, will anyone actually use that? I doubt many people will buy an expensive high end GPU, and then put it in power saving mode (keeping it in standard mode will give you about 9% higher performance, for the 64, as you can see here).

Here is their chart:

[laminutederire]

9 minutes ago, LAwLz said:

I got my numbers from TechPowerUp which takes 20 games into consideration (not just one) which I think will be a pretty good average without any cherry picking. But to be fair, they did not use power saving BIOS. But on the other hand, will anyone actually use that? I doubt many people will buy an expensive high end GPU, and then put it in power saving mode (keeping it in standard mode will give you about 9% higher performance, for the 64, as you can see here).

Here is their chart:

For future designs it's interesting though. They were put of the efficiency range to push extra perf, but they're capable of way better in terms of efficiency.

9% compared to turbo or balanced mode?

[LAwLz]

Just now, laminutederire said:

For future designs it's interesting though. They were put of the efficiency range to push extra perf, but they're capable of way better in terms of efficiency.

9% compared to turbo or balanced mode?

9% higher performance going from the power saving BIOS to balanced mode. Turbo is within the margin of error from balanced so it's completely worthless.

[Marinatall_Ironside]

9 hours ago, laminutederire said:

The 56 draws less power than my fury, but I'm fine with my 530W PSU, so 750W is a bit extreme, 600W should be enough for mostly everyone granted that my cpu isn't that power hungry (6600k). With the power saving mode, the 56 is on par with reference 1070s and it draws something around 30W more, that's not that bad for efficiency in itself.

The power draw figure I mentioned is coming from both AMD and assuming a high-performance gaming PC using an Intel Core i7 or high-end Ryzen 7.

 

Yes it seems overkill to start there, but it's also never good to load a power supply to 80% or more of its total rated capacity, as that shortens the lifespan.

[laminutederire]

2 hours ago, JurunceNK said:

The power draw figure I mentioned is coming from both AMD and assuming a high-performance gaming PC using an Intel Core i7 or high-end Ryzen 7.

 

Yes it seems overkill to start there, but it's also never good to load a power supply to 80% or more of its total rated capacity, as that shortens the lifespan.

The calculated minimum power supply wattage for my system is 495W and I have a 530W PSU, and it usually draws something like 400W on load. That's nearly 80 percent but under it. 600W is probably plenty enough for most systems

[leadeater]

16 hours ago, Dylanc1500 said:

Details on the Volta arch improvements: Here

Some comparions between P100 and V100: Here

Everyone loves slides so here's a couple slides: Slides

Have to wonder how the gaming GPUs will be configured. Looking at all that information there are two standout things to me: Most performance comparisons/boosts shown are for the Tensor cores, Volta SM are now split and have dedicated INT32 and FP32 cores.

 

Going in to the second point more, SM cores, which is more applicable to us for gaming I have to wonder how good this could be. So the SM count has increased to 80 but the core makeup has totally changed. Each SM has 64 FP32 Cores, 64 INT32 Cores, 32 FP64 Cores, and 8 Tensor Cores. We know Nvidia cuts down the FP64 cores for the gaming cards but will they also keep the separate INT32 cores? How useful would dedicated INT32 cores be for gaming?

 

What if Nvidia swaps out the INT32 cores for FP32 cores instead giving 128 of them for 10240 CUDA cores, does this mean Volta could have well over double the performance of Pascal?!?

 

I'm also very sure Tensor cores will be removed for GV102 and below, for cost/die size reasons.

 

[Paragon_X]

Either nVidia is pulling an Intel, or they have to delay and use this as an excuse because the maturity of the node they want to use isnt where it needs to be yet. So that means we are stuck with the 10 series atleast up to next summer.

[Dylanc1500]

7 hours ago, leadeater said:

Have to wonder how the gaming GPUs will be configured. Looking at all that information there are two standout things to me: Most performance comparisons/boosts shown are for the Tensor cores, Volta SM are now split and have dedicated INT32 and FP32 cores.

 

Going in to the second point more, SM cores, which is more applicable to us for gaming I have to wonder how good this could be. So the SM count has increased to 80 but the core makeup has totally changed. Each SM has 64 FP32 Cores, 64 INT32 Cores, 32 FP64 Cores, and 8 Tensor Cores. We know Nvidia cuts down the FP64 cores for the gaming cards but will they also keep the separate INT32 cores? How useful would dedicated INT32 cores be for gaming?

 

What if Nvidia swaps out the INT32 cores for FP32 cores instead giving 128 of them for 10240 CUDA cores, does this mean Volta could have well over double the performance of Pascal?!?

 

I'm also very sure Tensor cores will be removed for GV102 and below, for cost/die size reasons.

 

From what I understand and how it was explained to us, the tensor cores have the ability to  also operate as two FP16 units or a single FP32 unit doing two FP32 operations. I could be misunderstanding it entirely.

[leadeater]

5 minutes ago, Dylanc1500 said:

From what I understand and how it was explained to us, the tensor cores can also operate as two FP16 units or a single FP32 unit doing two FP32 operations. I could be misunderstanding it entirely.

Something like that, in your link there's comments down the bottom and someone asks that question and it gets answered. Still I can't really see Tensor core staying outside of the slim chance the Titan's get them and not the other GeForce products. Nvidia won't be putting gaming GPUs on the market above 610mm2 that's for sure.

 

Edit:

Quadro's share the same die with minor revisions to them as the GeForce products but Telsa is the dedicated product line for the tasks Tensor is good at so Quadro doesn't need them either.

[Dylanc1500]

3 minutes ago, leadeater said:

Something like that, in your link there's comments down the bottom and someone asks that question and it gets answered. Still I can't really see Tensor core staying outside of the slim chance the Titan's get them and not the other GeForce products. Nvidia won't be putting gaming GPUs on the market above 610mm2 that's for sure.

 

Edit:

Quadro's share the same die with minor revisions to them as the GeForce products but Telsa is the dedicated product line for the tasks Tensor is good at so Quadro doesn't need them either.

Oh it will definitely be cut down, it'll be interesting what they will for that, that is for sure.

 

I see a possibility of them having at least one Quadro with the full V100 (similar to the Quadro P100)  but I guess we will just have to wait and see. The rest of the line-up will be, just as you said, sharing the same dies with the consumer line with specific revisions to them.

 

On a separate thought could you imagine them putting a fully enabled GV100 on a Titan? My wonder is whether games could take full advantage of something like the tensor cores. I haven't dealt with games (last game I played was portal 2 lol) , and what not, to know.

[leadeater]

9 minutes ago, Dylanc1500 said:

I see a possibility of them having at least one Quadro with the full V100 (similar to the Quadro P100)  but I guess we will just have to wait and see. The rest of the line-up will be, just as you said, sharing the same dies with the consumer line with specific revisions to them.

Oh right, I totally forgot there was a Tesla P100 and a Quadro P100.

 

9 minutes ago, Dylanc1500 said:

On a separate thought could you imagine them putting a fully enabled GV100 on a Titan? My wonder is whether games could take full advantage of something like the tensor cores. I haven't dealt with games (last game I played was portal 2 lol) , and what not, to know.

I'd think based on how much faster they are supposed to be performance wise if they were any good for gaming and/or other workloads Volta would only have Tensor cores, but to be honest my knowledge of Tensor and TensorFlow stops at knowing it exists lol.

[Dylanc1500]

Just now, leadeater said:

Oh right, I totally forgot there was a Tesla P100 and a Quadro P100.

 

I'd think based on how faster they are supposed to be performance wise if they were any good for gaming and/or other workloads Volta would only have Tensor cores, but to be honest my knowledge of Tensor and programming stops at knowing it exists lol.

Ya I know exactly ZERO about Tensor. I design and create databases, I didn't know what Tensor was till we got info last year about Volta lol.

[Eroda]

On 15/08/2017 at 6:52 PM, TrigrH said:

Did anyone honestly expect Volta to come out soon anyway?

well if Vega Crushed Pascal then yeah maybe

[Taf the Ghost]

2 hours ago, Eroda said:

well if Vega Crushed Pascal then yeah maybe

Nvidia has known how far Vega could be pushed since last year. That's why they pushed out the 1080 Ti & 1050 Ti when they did, along with some price cuts. Though even Nvidia expected AMD to make the April window.

[AlwaysFSX]

5 hours ago, Dylanc1500 said:

Ya I know exactly ZERO about Tensor. I design and create databases, I didn't know what Tensor was till we got info last year about Volta lol.

What's Tensor for us dumb people?

[Master Disaster]

1 hour ago, Taf the Ghost said:

Nvidia has known how far Vega could be pushed since last year. That's why they pushed out the 1080 Ti & 1050 Ti when they did, along with some price cuts. Though even Nvidia expected AMD to make the April window.

Yep, Linus has stated multiple times on different WAN shows that its pretty normal for the competition to have the oppositions new hardware in their labs weeks (if not months) before general release.

 

Its a weird concept and one which I struggle to understand (like how the hell do they get hold of the hardware so early?) but Linus knows his shit when it comes to how these companies operate.

[Taf the Ghost]

41 minutes ago, Master Disaster said:

Yep, Linus has stated multiple times on different WAN shows that its pretty normal for the competition to have the oppositions new hardware in their labs weeks (if not months) before general release.

 

Its a weird concept and one which I struggle to understand (like how the hell do they get hold of the hardware so early?) but Linus knows his shit when it comes to how these companies operate.

GPUs are a different space than CPUs. I believe Board Partners aren't bound in the same way with the GPUs than with the CPUs. There's also the issue that, with GPUs, they normally both need access to the same Memory. This means starting a pricing war hurts both of them, so they're a little more chummy. 

 

With CPUs? They ship early test chips by armed guard.

[BirdyTheMighty]

nice means my 980 ti will solider on, granted i dont play any games (except bf1, fallout and world of tanks/warships) soo 980ti is a wasted space