NVIDIA Pascal Thread

[AtenRa]

There is a rumor that GP100 will not be released for the consumer market (PC and gamers).

 

[antihelten]

There is a rumor that GP100 will not be released for the consumer market (PC and gamers).

It's really more speculation than rumor (at least there has been plenty of speculation on this possibility in this thread), and it's really quite a reasonable theory.

First off there's the lack of any mention of ROPs in the diagrams and articles on GP100. This might just be an omission from Nvidia, but if the ROPs are in fact missing, then that would instantly preclude it from launching as a consumer card (i.e. non-tesla card).

The other thing is that GP100 just wouldn't be very competitive as a gaming card in the the time frame it could potentially launch in. Right now GP100 would be roughly 35% faster than an aftermarket 980Ti (3584 cores at 1480 MHz, versus 2816 cores at ~1400 MHz). 35% faster would be plenty competitive if they launched a GP100 equipped Geforce card today, but the Tesla equipped cards are not set to launch until Q1 2017, so Geforce version would probably not be available until Q2-Q3 2017. 12-15 months from now and only 35% faster, would almost certainly put the card somewhere around midrange to upper midrange, and thus a price range of $300-400. Selling a 610mm2 GPU with HBM2, for $300-400, isn't really a viable prospect for Nvidia.

 

[Adored]

Yeah I don't believe we'll see GP100 for consumers. Any die candidates would likely end up quite a lot slower than GP102 (assuming that exists) and probably not worth the extra cost of interposer and HBM2.

 

[Silverforce11]

The other thing is that GP100 just wouldn't be very competitive as a gaming card in the the time frame it could potentially launch in. Right now GP100 would be roughly 35% faster than an aftermarket 980Ti.

You are severely underestimating +IPC and the new 2x Warp 64 layout that means Pascal is better suited for the era of GCN-optimized console port.

GP100 will be no slouch for gaming. NV has never made a big chip that was weak at gaming, period.

The FP64 units just make it lose some perf/mm2 and perf/w for gaming, but the gaming performance is still oustanding, always.

Just look back at the 480, 580, Titan/780Ti etc. These were compute focused chips with excellent gaming performance. NV here just takes it up a notch, as those prior gen had 1:3 FP64, while big Pascal has 1:2 FP64. It's also relatively bigger, so it makes up for that loss in perf/mm2 already.

I fully expect big Pascal to be ~60% better than Titan X for gaming, rising to ~80% for GCN optimized titles.

 

[antihelten]

You are severely underestimating +IPC and the new 2x Warp 64 layout that means Pascal is better suited for the era of GCN-optimized console port.

Or maybe you're just overestimating those things, seeing as there has been absolutely no indications that those things even exist (Pascal going from 4 to two processing blocks compared to Maxwell, and thus from 4 to 2 warp schedulers doesn't mean it's changing to a warp 64 layout, just like Maxwell didn't have a warp 128 layout. Instead Maxwell has a 4x warp 32 layout, and Pascal has a 2x warp 32 layout)

Secondly I never said GP100 wasn't a good gaming GPU, and it will probably beat an aftermarket 980 Ti by roughly 35%, which is about the same as what Titan beat 680 by. I'm simply saying that by the time GP100 eventually finds it's way into a Geforce card it won't be competing with the 980 Ti, it will be competing with Vega and possibly GP102, and at best it will only be able to match the performance of those GPUs. Imagine if the original Titan (GK110) had launched into a market where Hawaii already existed, in that case I very much doubt Nvidia would have been able to take anywhere near $1000 for it.

 

[dangerman1337]

I'm also in the "GP100 won't be used for consumers" camp, a cut down one would have to be offered and would be very neutered in terms of SP performance. I think a GP102 will be offered with a lot more CUDA cores along with either GDDR5X or HBM2 with some if not all the HPC stuff slashed out. More compute friendly than Maxwell at the very least IMO.

 

[tential]

If they were smarter they wouldn't release the reference model to the press, but only custom models with huge OC.

That's the nice thing about NV's boost, it's just so easy. No tweaks required, move the slider to the right, viola!

Hence why I say amd should ditch reference cards and instead designate an oem card as reference. Far more incentive to have oems compete to deliver the best design which oem will then ship out to reviewers as the "reference".

Or just designate sapphire..... Seriously.


That + auto boost oc would make day 1 reviews much better.

 

[Silverforce11]

Secondly I never said GP100 wasn't a good gaming GPU, and it will probably beat an aftermarket 980 Ti by roughly 35%, which is about the same as what Titan beat 680 by.

That doesn't make any sense. Titan and 680 are the same family, same gen and node. GK104 and GK110.

We're talking about GP100, you have to compare it to Titan beating the 580.

 

[Kris194]

People here are really underestimating GP100. 30-40% higher clocks, ~17% more units, improved IPC. That alone gives you at least 50-60%. Not to mention that some of these people compare stock GP100 to very highly overclocked GM200...

 

[Genx87]

I'd be surprised if this chip shows up in a gaming card. Those FP64 units cost a lot of silicon and have little use in gaming. I would imagine a card like GP100 in the gaming space will replace those FP64 units with 1 or 2 FP32 units.

 

[antihelten]

That doesn't make any sense. Titan and 680 are the same family, same gen and node. GK104 and GK110.

We're talking about GP100, you have to compare it to Titan beating the 580.

No I don't, people don't by GPUs based upon their node or architecture, they buy them based upon their performance relative to the competition. Titan didn't compete with 580 when it launched, it competed with the 680 (and beat it by roughly 35%).

GP100 would compete with the 980 Ti if it launched today, and with Vega/GP102 if it launches in Q2/Q3 2017. So those are the GPUs you have to compare it to.

People here are really underestimating GP100. 30-40% higher clocks, ~17% more units, improved IPC. That alone gives you at least 50-60%. Not to mention that some of these people compare stock GP100 to very highly overclocked GM200...

Why do you think it will have higher IPC (outside of compute)?

And the reason why I'm comparing it to an overclocked GM200, is because that's what people can buy today out of the box. If you can pay $50 more for 20% more performance, you would be crazy not to do so (unless you're planning to overclock yourself of course)

 

[Adored]

People here are really underestimating GP100. 30-40% higher clocks, ~17% more units, improved IPC. That alone gives you at least 50-60%. Not to mention that some of these people compare stock GP100 to very highly overclocked GM200...

30-40% higher clocks vs a stock 980 Ti...how many of those do you see though? I'll eat my hat if big Vega doesn't blow that away tbh. Nvidia must have a GP102 or something in the works.

 

[Stuka87]

GP100 will be no slouch for gaming. NV has never made a big chip that was weak at gaming, period.

How is a card with no ROPs going to be a gaming card at all?

Pretty sure the big chip gaming card will be the GP102 we have heard about.

 

[MrTeal]

I don't get why people keep saying it has no ROPs. There's absolutely no evidence of that. They showed some block diagrams of the part being used as a compute chip, and the omission of a render block there doesn't mean it doesn't exist on the silicon.

While the module is P100 and doesn't have video outputs, nVidia specifically refers to the chip as GP100. Unless someone can point to a credible source at nVidia saying that the chip can't do graphics, saying there's no ROPs is wildly premature.

 

[antihelten]

I don't get why people keep saying it has no ROPs. There's absolutely no evidence of that. They showed some block diagrams of the part being used as a compute chip, and the omission of a render block there doesn't mean it doesn't exist on the silicon.

While the module is P100 and doesn't have video outputs, nVidia specifically refers to the chip as GP100. Unless someone can point to a credible source at nVidia saying that the chip can't do graphics, saying there's no ROPs is wildly premature.

Fun fact, I just had a look at the old GK110 press announcement article here on Anandtech, which happened to include a diagram of GK110, and guess what, no ROPs here either:

later diagrams of GK110 did include the ROPs (blue boxes next to the L2 cache):

So yeah, definitely premature.

 

[Adored]

I don't think it's the lack of ROPs that are the issue, more the lack of specs. Given the relative immaturity of the node we can only really expect to see a GP100 Titan with 2 disabled SMX and a voltage bump.

How close will that be to a normal gaming focused 384-bit GDDR5 GP102? The extra cost of the HBM2 and interposer will be too much to make it worth it.

Then how fast is that likely to be compared to big Vega? What if AMD goes with a 600mm2 Vega as well, except a pure gaming GPU (their last "big" GPU)? A cut-down GP100 will get annihilated.

 

[antihelten]

I don't think it's the lack of ROPs that are the issue, more the lack of specs. Given the relative immaturity of the node we can only really expect to see a GP100 Titan with 2 disabled SMX and a voltage bump.

How close will that be to a normal gaming focused 384-bit GDDR5 GP102? The extra cost of the HBM2 and interposer will be too much to make it worth it.

Then how fast is that likely to be compared to big Vega? What if AMD goes with a 600mm2 Vega as well, except a pure gaming GPU (their last "big" GPU)? A cut-down GP100 will get annihilated.

Well some people were arguing that the apparent lack of ROPs would exclude the GPU from ever showing up as a Geforce card.

I do agree with you that the specs relative to the price of the GPU (610mm2 plus HBM2 doesn't come cheap), would largely prevent it from being competitive as a gaming card within the timeframe that it would be likely to be released.

 

[Despoiler]

The specs for P100 do not list a ROP count either. That's not to say they don't exist, but they aren't being revealed as a spec.
https://devblogs.nvidia.com/parallelforall/inside-pascal/

 

[MrTeal]

The specs for P100 do not list a ROP count either. That's not to say they don't exist, but they aren't being revealed as a spec.
https://devblogs.nvidia.com/parallelforall/inside-pascal/

ROP count would be relatively meaningless for P100 or any Tesla card, so that's not surprising.

 

[antihelten]

The specs for P100 do not list a ROP count either. That's not to say they don't exist, but they aren't being revealed as a spec.
https://devblogs.nvidia.com/parallelforall/inside-pascal/

GK110 didn't have it's ROP count listed either when it was first announced

Original GK110 white paper.

 

[Headfoot]

Absence of proof is not proof of absence

 

[Sweepr]

Inside Pascal Presentation - GTC 2016

http://on-demand.gputechconf.com/gtc/2016/presentation/S6176-Mark-Harris-Lars-Nyland.pdf

 

[RussianSensation]

Secondly I never said GP100 wasn't a good gaming GPU, and it will probably beat an aftermarket 980 Ti by roughly 35%, which is about the same as what Titan beat 680 by.

What kind of math is that?

Fully unlocked GP100 is 3840 CCs. Let's assume it comes clocked at 1480mhz and the full 1TB/sec HBM2. Out of the box, most reference 980Ti cards Boost to 1202mhz or so.

3840 * 1480 / (2816 * 1202) = 68%.

That's not accounting for:

1) NV's concentrated focus on Pascal driver support, which coincidentally means less focus on Maxwell (*Kepler gen hint hint*). With the architecture mimicking many parts of GCN, it also means GCN PS4/XB1 console ports will run much better on Pascal vs. Maxwell with minimal optimization.

2) increase in IPC of Pascal over Maxwell.

I have no clue where you got your 35% number from. I am pretty sure you have GP104 mixed up with GP100.

FYI, on paper, the performance increase from 780Ti to 980Ti is LESS than going from a 980Ti to GP100 and people tout 980TI as the greatest thing since sliced bread this gen. Tons and tons of 780Ti users upgraded to 980Ti for WAY less than a 68% boost in performance.

 

[RussianSensation]

People here are really underestimating GP100. 30-40% higher clocks, ~17% more units, improved IPC. That alone gives you at least 50-60%. Not to mention that some of these people compare stock GP100 to very highly overclocked GM200...

:thumbsup: That's 50-60% against the fully unlocked Titan X. Most people here purchased 980Ti, not the TX. In that context, GP100 on paper is almost 70% faster than the card most enthusiasts will be upgrading from. I guess most were expecting a leap of 2X based on GTX580->780Ti. I am sure NV has tricks up its sleeve such as upping the clocks from 1480mhz since by the time 1080Ti launches, the 16nm will be more mature. Secondly, the above analysis assigns absolutely 0% increase in IPC of Pascal over Maxwell. If NV squeezes 20% IPC on top of that, that's more or less 95%+ faster than 980Ti.

I don't get why people keep saying it has no ROPs. There's absolutely no evidence of that.

Because it flows better with the theory that NV GP100 is weak sauce and that it's not a real flagship. So the real flagship will throw out all the compute junk and instead have 5120-6144 CUDA cores clocked @ 1.4-1.5Ghz. This of course makes WAY more sense when we do the math:

5120 CUDA cores @ 1400mhz (The Real Big Pascal)
vs.
3072 CUDA cores @ 1075mhz (The Real Big Maxwell Titan X)
=> Gives us a 2.17X increase in raw performance. No wait, that's still too slow. Needs MOAR Cuda Cores.

6144 CUDA cores with 1480mhz clocks it is then. :thumbsup:

 

[jpiniero]

Even if there are ROPs, there won't be a GP100 Titan at 610 mm2, even severely cut. It'd be waaaay too expensive. Maybe a Quadro but that would be several K.

If the GP102 exists it's likely something like 300-350 mm2, and that would still be something that would probably be 800-1100. Vega is likely the same deal I guess, where Vega 10 is basically compute only where Vega 11 is the extreme end for gamers and FirePros.