Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Nvidia announces GP100, the big Pascal
Key specs:
3584 shaders (for comparison, Fury X has 4096, Titan X has 3072)
14 MB registers (Fury X has 16 MB, Titan X has 6 MB)
4 MB L2 cache (Fury X and Titan X both have 2 MB)
10.6 TFLOPS single-precision (Fury X has 8.6 TF, Titan X has 6.6 TF)
21.2 TFLOPS half-precision (no other recent GPUs do half precision faster than single)
5.3 TFLOPS double-precision (next best: Hawaii has 2.6 TF)
16 GB HBM2 (FirePro S9170 has 32 GB GDDR5; only HBM card available is Fury line with 4 GB)
720 GB/sec bandwidth (Fury X has 512 GB/sec, Titan X has 336 GB/sec)
1.48 GHz boost clock (highest of any GPU ever; Fury X is 1.05 GHz, Titan X is 1.075 MHz)
300 W TDP (Fury X is 275 W, Titan X is 250 W)
If we take their specs at face value, the big Pascal will be faster than any GPU currently on the market for nearly all purposes. That shouldn't be surprising, as the top end GPUs of one generation usually are faster than the previous generation. But they're only claiming a 25% energy efficiency gain over a Titan X, in spite of shrinking all the way from 28 nm to 16 nm. That's rather less than I was hoping for, though in their defense, Maxwell is awfully efficient for 28 nm.
But that's if we take their specs at face value, and maybe we shouldn't. A 600 mm^2 chip as your first chip on a new process node is something that may never been tried before. I can't even think of any situation where someone's first chip on a new process node was even 400 mm^2. I don't know if they make huge chips for test SRAM or not, and FPGAs are ridiculously repairable, but neither of those are at all like making a huge GPU that is a lot less tolerant of defects.
The last time we saw a huge chip of a new architecture on a new-ish process node was Fermi, with the GeForce GTX 480 being 529 mm^2. That was a fairly new process node, but Nvidia had made several small chips on 40 nm (GeForce G 210, GT 220, and GT 240) before Fermi. Several months before it launched, Nvidia promised Tesla cards with:
512 shaders
750 MHz
175 W
Impressive specs for its day. By the time the chip actually launched, the specs were:
448 shaders
600 MHz
225 W
That's a change of 30% less performance for 30% more power consumption. And the 225 W was a lie, too. And that's why we regard the GeForce GTX 480 (GeForce version of the same card) as having been such a disaster.
The next best example that comes to mind is Tahiti, in the Radeon HD 7970. That was AMD's first chip of a new architecture on a new process node, but it was also 378 mm^2--a whole lot less than 529, let alone 610. And it was also the least efficient GPU ever made on 28 nm by a substantial margin. It also helped that TSMC's 28 nm process node was a whole lot less troubled than 40 nm, 16 nm, or especially 20 nm have been.
Now, this won't be the only Pascal chip, of course, but the big one is the interesting one. The next chip down in the lineup will likely be slower than what you could buy today. Now we just wait for AMD to offer more details on Polaris and Vega, and to see if Nvidia can actually build anything with the promised specs or if they'll have to cut way back.
3584 shaders (for comparison, Fury X has 4096, Titan X has 3072)
14 MB registers (Fury X has 16 MB, Titan X has 6 MB)
4 MB L2 cache (Fury X and Titan X both have 2 MB)
10.6 TFLOPS single-precision (Fury X has 8.6 TF, Titan X has 6.6 TF)
21.2 TFLOPS half-precision (no other recent GPUs do half precision faster than single)
5.3 TFLOPS double-precision (next best: Hawaii has 2.6 TF)
16 GB HBM2 (FirePro S9170 has 32 GB GDDR5; only HBM card available is Fury line with 4 GB)
720 GB/sec bandwidth (Fury X has 512 GB/sec, Titan X has 336 GB/sec)
1.48 GHz boost clock (highest of any GPU ever; Fury X is 1.05 GHz, Titan X is 1.075 MHz)
300 W TDP (Fury X is 275 W, Titan X is 250 W)
If we take their specs at face value, the big Pascal will be faster than any GPU currently on the market for nearly all purposes. That shouldn't be surprising, as the top end GPUs of one generation usually are faster than the previous generation. But they're only claiming a 25% energy efficiency gain over a Titan X, in spite of shrinking all the way from 28 nm to 16 nm. That's rather less than I was hoping for, though in their defense, Maxwell is awfully efficient for 28 nm.
But that's if we take their specs at face value, and maybe we shouldn't. A 600 mm^2 chip as your first chip on a new process node is something that may never been tried before. I can't even think of any situation where someone's first chip on a new process node was even 400 mm^2. I don't know if they make huge chips for test SRAM or not, and FPGAs are ridiculously repairable, but neither of those are at all like making a huge GPU that is a lot less tolerant of defects.
The last time we saw a huge chip of a new architecture on a new-ish process node was Fermi, with the GeForce GTX 480 being 529 mm^2. That was a fairly new process node, but Nvidia had made several small chips on 40 nm (GeForce G 210, GT 220, and GT 240) before Fermi. Several months before it launched, Nvidia promised Tesla cards with:
512 shaders
750 MHz
175 W
Impressive specs for its day. By the time the chip actually launched, the specs were:
448 shaders
600 MHz
225 W
That's a change of 30% less performance for 30% more power consumption. And the 225 W was a lie, too. And that's why we regard the GeForce GTX 480 (GeForce version of the same card) as having been such a disaster.
The next best example that comes to mind is Tahiti, in the Radeon HD 7970. That was AMD's first chip of a new architecture on a new process node, but it was also 378 mm^2--a whole lot less than 529, let alone 610. And it was also the least efficient GPU ever made on 28 nm by a substantial margin. It also helped that TSMC's 28 nm process node was a whole lot less troubled than 40 nm, 16 nm, or especially 20 nm have been.
Now, this won't be the only Pascal chip, of course, but the big one is the interesting one. The next chip down in the lineup will likely be slower than what you could buy today. Now we just wait for AMD to offer more details on Polaris and Vega, and to see if Nvidia can actually build anything with the promised specs or if they'll have to cut way back.
Comments
its pretty rapid from this point forward
Please do not respond to me, even if I ask you a question, its rhetorical.
Please do not respond to me
http://anandtech.com/show/10222/nvidia-announces-tesla-p100-accelerator-pascal-power-for-hpc
http://www.tomshardware.com/news/nvidia-pascal-tesla-p100-gpu,31557.html
http://techreport.com/news/29946/pascal-makes-its-debut-on-nvidia-tesla-p100-hpc-card
The Q1 2017 date being reported may be only for Tesla cards. Sometimes GeForce variants come sooner, but as this is a very HPC-focused card, that's not guaranteed to happen. The two previous HPC-focused chips have been those of the GeForce GTX 480 (Tesla M2050) and GeForce GTX Titan (Tesla K20), and the Tesla cards seem to have released about the same time as the GeForce cards for the former and before the GeForce cards for the latter.
AMD is apparently promising Polaris around the middle of this year and Vega early next year. It's not clear whether Polaris will be an upgrade over cards you can buy today, however, except that it will have better energy efficiency. There will surely be some lower end Pascal chips coming, too, but I haven't seen any Nvidia announcement about them.
21.2 TFLOPS half-precision (no other recent GPUs do half precision faster than single)
5.3 TFLOPS double-precision (next best: Hawaii has 2.6 TF)
16 GB HBM2 (FirePro S9170 has 32 GB GDDR5; only HBM card available is Fury line with 4 GB)
720 GB/sec bandwidth (Fury X has 512 GB/sec, Titan X has 336 GB/sec)
1.48 GHz boost clock (highest of any GPU ever; Fury X is 1.05 GHz, Titan X is 1.075 MHz)
300 W TDP (Fury X is 275 W, Titan X is 250 W)
Stats are quite dissapointing. 300W TDP. Thats what happens when you add DP hardware.
Hawaii is what, 459mm^2 28nm 2,6 TF DP at 1000 MHz. so 3,9 at 1500 like this one. Double the output by 2 for node shrink and youre looking at 7,8 TF for 459mm^2 vs 5,3 TF for 610mm^2 for Pascal.
Also you can "preorder" 8 of these for 129 000 $. Yields on this must be abbysmal.
They also made very "clever" PR stunt where 2xperf/W was for Kepler, not Maxwell....
1) the first discrete GPU chip on a new process node, the first time Nvidia has attempted this (rather than letting AMD go first) in over a decade,
2) the largest publicly known chip on the process node by about a factor of four,
3) the largest GPU chip (by mm^2) ever made, on any process node, ever,
4) the highest clocked GPU chip ever, higher than any Radeon or GeForce, in spite of Tesla cards normally being clocked much lower than consumer graphics cards, and
5) the highest wattage GPU chip ever.
Nvidia had a slide promising "five miracles", and they'll probably need more than five to pull off all of that. For all that risk, even if everything pans out the way Nvidia is hoping, it might be 50% faster than a Titan X, if that. And if things don't work how Nvidia hopes, there's no real guarantee that it will be meaningfully faster than a Titan X at all.
Or that it will have widespread availability a year from today. If you promise a hard launch in a month, you've already got a ton of working chips and realistically, you know that you're going to deliver barring some very low probability events like a fab getting destroyed. If you promise something six months out, you're hoping that the next respin goes well with no real reason to believe that it will. I'm not sure if a major GPU chip has ever had specs publicly announced this far before availability; the only one that I can think of in the same ballpark is the GeForce GTX 480, hardly an auspicious omen.
If you're looking for a new desktop card and waiting for Pascal, I'd stop waiting. GP100 is not the chip you want for graphics, and there's no real guarantee that the next chip down in the line will be faster than a GTX 980 Ti or a Fury X that you could buy today.
Waiting for Vega is a different matter. Still plan on waiting a year, but maybe it will be a great gaming chip. Or maybe not. We have no clue, really, but will at least get a better idea of what AMD has up their sleeve when Polaris comes, promised for around the middle of this year.
But, NVidia still hasnt said single word on consumer chips, all they tout is "self driving cars" and now this "Q1 (maybe) 2017" Tesla that you can preorder 6-7-? months in advance....without even showing a working card.
PS: they finally showed a working card. Wonder why they didnt do it at Huang gig.
So next time try not to be funny, it doesnt suit you.
거북이는 목을 내밀 때 안 움직입니다
I'm not sure where you are getting this 2017 release date. None of the articles mention a release date, and all the 2017 stuff is speculation.
It would be completely unprecedented for NVidia or AMD to release info that far in advance of an actual hardware release.
You're making quite a lot of inferences from quite a small amount of information.
"The surest way to corrupt a youth is to instruct him to hold in higher esteem those who think alike than those who think differently."
- Friedrich Nietzsche
That the stock price didn't change very much really only means that the announcement wasn't massively better or worse than expected. Had Nvidia announced that it's clearly worse than previous generation parts, or that they already have over $1 billion in pre-orders for the Tesla cards, you'd see some real stock price movement.
I'm not disagreeing with you on the stock situation. I'm sure that's a significant factor. But releasing info this early with expected release 8-16 months out would actually be bad for their long term stock prices. It's just not wise business particularly in this field. Intel has had that bite them in the ass in the past when they did paper launches of products and then had massive supply issues when they hard launched.
My comment about the speculation was mostly in regards to the insinuations that they were bull faced lying to everyone about the card, and that it would be substantially changed/delayed/etc, and more importantly that they knew this upfront and were lying on purpose.
"The surest way to corrupt a youth is to instruct him to hold in higher esteem those who think alike than those who think differently."
- Friedrich Nietzsche
Remember the Fermi debacle where Nvidia bit off more than they could chew and it backfired disastrously? The entire GeForce 400 series was a train wreck, and by the time they fixed it with the GeForce 500 series, AMD had had their DirectX 11 lineup out for a year.
What Nvidia is trying with Pascal is massively more ambitious than what they tried with Fermi. Compared to big Pascal, big Fermi was a smaller, lower power chip that got to follow both moderately large AMD chips to help TSMC work out process node issues and three other small chips from Nvidia themselves to help Nvidia understand the process node. Pascal has none of those advantages.
I thought Nvidia had learned their lesson. Kepler started with a medium sized chip, and the big Kepler came nearly a year later. Maxwell started with a small chip, before working up to medium and eventually a huge one. Maybe Pascal will end up doing that and Nvidia just isn't talking about the smaller chips that will precede GP100 by half a year or whatever. But if that's the case, then big Pascal is a long, long way off.
Don't count on anything with GDDR5X or HBM2 launching this summer. Fall is more plausible, but they'll have to wait until they can get the new memory types in large volume. GDDR5 cards could come sooner, and that seems to be the direction AMD is going with Polaris. GDDR5 is not viable for next-generation, top end cards, but it will do for the middle to low end cards.
http://www.tomshardware.com/news/nvidia-pascal-tesla-p100-gpu,31557.html
"The Tesla P100 is in volume production today and will ship “soon." Jen-Hsun Huang said that really means it will show up in the cloud first, and ship via OEMs in Q1 2017.
Of course, as a consumer, you can’t have this chip yet. That’s no surprise. In the past, Nvidia also first implemented its new architectures and GPUs in scientific super computing before bringing it to the consumer market. Nvidia did not announce when we would be seeing a consumer-oriented Pascal chip."
------------
2025: 48 years on the Net.
I will instruct you, because you seem incapable of comprehending yourself. You should have said: "it is back up now" instead trying to be funny (but lookign idiotic). As i already said funny doesnt suit you.
But this is very specific chip and buying it for anything else than inteded is...
OTOH, what if they indeed bring this chip to consumers as Titan. Fully enabled SP shaders (38xx) with DP part diabled (Nvidia has to have dedicated DP hardware on their chip which is useless for gaming). But that would kinda be a failure because they would disable 1/3 of the chip and have roughly 400mm^2 of 610mm working. Ti would have even more of the chip disabled.
It is a theory, but how viable is completely new chip without DP part?
Keeping DP is very costly, you basically take chips from potentionally very low yields to put in much cheaper parts rather than 10+k$ Tesla/Quadro. If that is true next Titan may be slated for GTC next year when node matures and they have enough faulty chips from Tesla/Quadro.
b) AMD has nothing to do with this tread, NVidias GTC or GP100
Please, dont embarass yourself further (even if its funny as hell) Its YOU who dont have a clue of the context yet keep babbling nonsense.