NVIDIA Pascal Thread

[jpiniero]

I could see nVidia calling the full GP104 1080 Ti, especially if the price is $649 or more. Really, the main target of GP104 is anyone with Kepler or earlier along with 970/980, and people with really flaccid e-penises. The people with 980 Tis... nVidia will get them with GP102 or with the next round.

 

[Fallen Kell]

But they will keep hiking the prices and ultimately kill their own market altogether, cause the price of a new gaming PC will become outrageously high compared to consoles. Its already happening.

To that I will say a high end gaming PC was always outrageously high compared to the consoles. It cost approx $1600 in 2000 to build a high end PC, while the PS2 cost between 1/5th and 1/6th that. RIght now, a high end PC will be in the $2k range, but a PS4 or XbOne will cost 1/4th that amount. If anything, the prices of the consoles have actually come up vs the price of a high end gaming computer.

 

[Silverforce11]

No way that rumored 2560CC will be barely 20% faster than stock 980 TI, it would be ridiculous and it would mean that it is only 10% faster than Titan X. If full GP104 have 2560CC it would be like

~30 - 40% faster clocks than Titan X
~83% units of Titan X
At least 10% more efficient SM what would be equal to 2816 CUDA Cores (GTX 980 Ti has exactly that amount and look what ~1,3GHz GTX 980 Ti is doing with stock Titan X

http://pclab.pl/zdjecia/artykuly/chaostheory/2016/01/gfx/OC/oc_bf4_2560u.png

Even if it boosted to 1.5ghz, that isn't 30-40% faster clocks than Titan X.

You guys have fallen for the marketing PR trick of boost clocks. I see it often, people think their Maxwell OC to 1.5ghz is like a HUGE 40% OC... when in reality its close to 20% because it already hits 1.25ghz out of the box even in reference cards. In custom cards, it gets to >1.4ghz, so the jump to 1.5ghz is tiny.

Now, let's get real.

So far with Pascal, 15-20% higher clocks vs reference. IPC we don't know about. It could be huge, it could be little. But given that even NVIDIA themselves have said Pascal ~= Maxwell + FP64 + Mixed Mode + NVlink etc... it would be closer to the lower end on IPC gains as the uarch are very similar.

The full GP104 could well be Titan X + 20% at much less power usage, that's a good mid-range for a next-gen part.

 

[Adored]

They will push the power usage in order to make a performance gap. It's much easier to sell extra performance than lower power usage. I wouldn't be surprised if this is where the dual 8-pin thing came from. OK I would be surprised if it had a dual 8-pin, but I won't be surprised if the 1080 Ti uses a lot more power than the average true "midrange" GPU of recent times.

 

[Magee_MC]

But Nvidia is not charging more for more (at least as far as the initial launch of Kepler is concerned), GTX 580 cost $500 and GTX 680 cost $500.

The only one who charged more for more was AMD (6970 cost $370, 7970 cost $550), but for some reason no one seems to be complaining about that one in this thread.

Sure you could argue that Nvidia is not providing enough "more" for the price, but that's different, and it's not really a question of people receiving less value for what they spend (680 isn't slower than 580), but rather that they are not receiving a big enough increase in value for what they spend. For what it's worth I absolutely agree with this, which is also why I didn't buy a 680 or a 7970, since to me simply not buying the product in question seemed a perfectly reasonable solution to this problem.

Maybe at that point in time, they didn't charge more for more, however, in the next generation, the prices for the top cards of the 700 series went to $649 (780), $699 (780Ti) and $1000 (Titan). They were charging more for more performance. It may have been that it was the Titan driving the price of the top of the GTX line to double the previous top of the line's price that started to galvanize the perception that NV was pushing prices higher for the same tier in their previous generation's stack.

 

[antihelten]

Maybe at that point in time, they didn't charge more for more, however, in the next generation, the prices for the top cards of the 700 series went to $649 (780), $699 (780Ti) and $1000 (Titan). They were charging more for more performance. It may have been that it was the Titan driving the price of the top of the GTX line to double the previous top of the line's price that started to galvanize the perception that NV was pushing prices higher for the same tier in their previous generation's stack.

The GK110 and GM200 cards did indeed feature price increases (GM204 and the $550 GTX980 also did, albeit only a slight one), I never said otherwise. My point was purely about the 600 series and GTX 680 in particular since that was the series relevant to the discussion at hand, which was focused on cost and it's effect on prices (cost affecting issues being the small die size of GK104 and the increased cost of the 28nm node).

 

[3DVagabond]

This is what you believe? Can you expand on this? Say a company develops a GPU that can double the performance of anything out there today, but can do so at half the cost. Why, exactly, do you feel they should charge half the price?

Where is this mythical chip? Try and keep it a little bit real. Besides I already told you what price is based on.

 

[Silverforce11]

Companies exist to make as much profit as possible. It's just quite normal for capitalism.

The problem I have is there are people who deny it, they don't want to see that prices have increased and they have this distorted view to try and defend it.

You don't need to defend it. NV and AMD are companies and they will milk the heck out of consumers if it's possible.

 

[cytg111]

Companies exist to make as much profit as possible. It's just quite normal for capitalism.

The problem I have is there are people who deny it, they don't want to see that prices have increased and they have this distorted view to try and defend it.

You don't need to defend it. NV and AMD are companies and they will milk the heck out of consumers if it's possible.

True that. On top of that you have goverment, taxes, regulations and unions to keep the capatalistic construct from overreaching, it is an equilibrium.

 

[Kris194]

To be honest, I don't think that we will see GTX 1080Ti on GP104. GP104 looks more like GTX 1060Ti, GTX 1070 and GTX 1080 just like GK104.

 

[antihelten]

To be honest, I don't think that we will see GTX 1080Ti on GP104. GP104 looks more like GTX 1060Ti, GTX 1070 and GTX 1080 just like GK104.

Agreed, and if the performance of those three parts are roughly equivalent to aftermarket 980 Ti, stock 980 Ti and stock 980 respectively, then they would would provide roughly 40-45% performance boost over the parts they replace i.e. 980, 970 and a theoretical 960 Ti (with performance halfway between 960 and 970).

40-45% performance boost seems to be pretty much the standard for node jumps from Nvidia, so it would make sense for them to aim for that. Then prices would probably follow the standard $500, $400 and $300 (which would be a slight increase overall vs. the 900 series at $550, $330 and a theoretical $250-280 for a 960 Ti).

 

[96Firebird]

Where is this mythical chip? Try and keep it a little bit real. Besides I already told you what price is based on.

I'd say the same to you... You know, like how the real world works. D:

 

[3DVagabond]

I'd say the same to you... You know, like how the real world works. D:

So, you think a chip that's 1/2 the size and 2x as fast is realistic?

 

[96Firebird]

So, you think a chip that's 1/2 the size and 2x as fast is realistic?

Hey, it could happen, but you seem to be avoiding answering me. Probably because you know how stupid the answer sounds.

I'll make it easier for ya, since you have a hard time understanding a simplistic example...

Say a company develops a GPU that is 150% the performance of anything out there today, but can do so at the same cost. Why, exactly, do you feel they should charge the same price?

Ready... GO!

 

[Glo.]

192 Pascal cores have approximately 250% performance of 192 Kepler cores.

64 Pascal cores have around 90% of 128 of Maxwell Cores.
Looks like the cores are also bigger than Maxwell's, thats why in 600mm2 monster there will be only 3840 CUDA cores. Only 30% more than Titan X on a node that is at least 50% denser.

 

[antihelten]

So, you think a chip that's 1/2 the size and 2x as fast is realistic?

That scenario isn't just realistic, it's downright inevitable. It's just a question of time, wont happen on 16nm obviously, but 10nm should get close.

192 Pascal cores have approximately 250% performance of 192 Kepler cores.

64 Pascal cores have around 90% of 128 of Maxwell Cores.
Looks like the cores are also bigger than Maxwell's, thats why in 600mm2 monster there will be only 3840 CUDA cores. Only 30% more than Titan X on a node that is at least 50% denser.

What are you basing this on?

Also remember that GP100 has to use die space on all the FP64 units, which isn't the case for Titan X, so the lower CUDA core count per mm2, doesn't necessarily indicate that the cores are bigger.

 

[Glo.]

https://devblogs.nvidia.com/parallelforall/inside-pascal/ Second table with Compute Capability. 64 Cores have exactly the same performance as 128 Maxwell cores, but have smaller cache. It is also important that the cores are compared here clock-to-clock. But Pascal ones will have higher clocks.

128 Maxwell Cores have had 90% of performance of 192 Kepler cores. You know why, by seeing that table.

 

[antihelten]

https://devblogs.nvidia.com/parallelforall/inside-pascal/ Second table with Compute Capability. 64 Cores have exactly the same performance as 128 Maxwell cores, but have smaller cache. It is also important that the cores are compared here clock-to-clock. But Pascal ones will have higher clocks.

128 Maxwell Cores have had 90% of performance of 192 Kepler cores. You know why, by seeing that table.

There is nothing in that table to indicate that Pascal will 250% the performance per core of Maxwell.

The reason why Maxwell was capable of achieving a ~35% IPC improvement over Kepler was because Kepler had problems being fully fed and thus achieved sub-optimal occupancy (often only a bit more than 2/3). There is not really anything to indicate that this is the case with Maxwell, at least nowhere near enough to achieve a 150% performance boost (even when accounting for higher clocks)

 

[Head1985]

Pascal should have much better "IPC" than maxwell
GTX980-16SMX
TITANX-24SMX
GP104- 40SMX?

GP100’s SM incorporates 64 single-precision (FP32) CUDA Cores. In contrast, the Maxwell and Kepler SMs had 128 and 192 FP32 CUDA Cores, respectively. The GP100 SM is partitioned into two processing blocks, each having 32 single-precision CUDA Cores, an instruction buffer, a warp scheduler, and two dispatch units. While a GP100 SM has half the total number of CUDA Cores of a Maxwell SM, it maintains the same register file size and supports similar occupancy of warps and thread blocks.GP100’s SM has the same number of registers as Maxwell GM200 and Kepler GK110 SMs, but the entire GP100 GPU has far more SMs, and thus many more registers overall. This means threads across the GPU have access to more registers, and GP100 supports more threads, warps, and thread blocks in flight compared to prior GPU generations.

 

[Glo.]

Yes, that is another thing if the cores will be fed enough. Thank for interesting observation, I didn't put it into account.

 

[Glo.]

Pascal should have much better "IPC" than maxwell
GTX980-16SMX
TITANX-24SMX
GP104- 40SMX?

Exactly.

 

[Silverforce11]

https://devblogs.nvidia.com/parallelforall/inside-pascal/ Second table with Compute Capability. 64 Cores have exactly the same performance as 128 Maxwell cores, but have smaller cache. It is also important that the cores are compared here clock-to-clock. But Pascal ones will have higher clocks.

128 Maxwell Cores have had 90% of performance of 192 Kepler cores. You know why, by seeing that table.

That's not how you calculate that btw.

That table focused on FP64 performance.

We can see with Kepler GK110, it has 960 FP64 CC @ 875mhz for 1,680 GFlops.

Pascal has 1,792 FP64 CC @ 1480mhz for 5,304 GFlops.

GFlops per FP64 CC:

GK110 = 1.75
GP100 = 2.95

But the clock speed is vastly higher, ~69%!

Factoring in identical clock speeds, ie, if GK110 operated at 1480mhz, it's GFlops per CC = ~2.95!

No difference in actual IPC. The gains come from having more CC at a higher clock.

If you want to talk about gaming performance per CC, then compare the 980 vs the 780Ti, 2048 CC vs 2880 CC, with vram capacities that are enough to compare.

http://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-review

The 980 when it was released was around 5-10% faster than the 780Ti, reference model.

Here's the kicker, the 980 reference boosts to above 1.25ghz, whereas the reference 780Ti boosted to ~950mhz. Immediately there's a ~30% clock improvement for the 980.

So 2048 x 1.3 = 2,662 effective CC vs Kepler. Pretty close to the 2880 CC in the 780Ti. Overall IPC gain is ~20%.

Anandtech's article talks a lot about the perf/w gains of Maxwell and they claim its to do with NV optimizing specifically for the 28nm node from TSMC, on a transistor level, because they knew they would be stuck on 28nm for longer than anticipated. Will this carry over to 16nm FF? Don't know!

ps. Where Pascal could gain a lot is in next-gen games that are GCN optimized, because it's gone with a very GCN-like layout, with optimal warp size of 64 identical to GCN. This means it is capable of hitting peak utilization of its 64CC per SM cluster when games are GCN optimized. Something Maxwell cannot do, and Kepler just eats dirt.

 

[Glo.]

You took the wrong table . The lower one was one I was referring .

 

[Silverforce11]

You took the wrong table . The lower one was one I was referring .

Seriously which table, just take a pic and upload it, would have saved me the hassle of that analysis.

 

[Glo.]

It is quite funny because you have posted that table in this very thread

http://forums.anandtech.com/showpost.php?p=38146747&postcount=1065

64 Cores from Pascal Arch have exactly the same performance clock to clock as 128 Cores from Maxwell and 192 from Kepler.