nvidia GTC 2016:Nvidia Pascal

[Colonel_Gerdauf]

59 minutes ago, DocSwag said:

Let's look at half precision first:

P100 has 21 tflops half precision

Radeon Pro Duo has 16 tflops single and supports half precision do I'm assuming it has 16 tflops half precision.

That's 1.33 times

 

At single:

P100 is 10 tflops

Radeon pro duo is 16 tflops

Radeon pro duo wins by 1.66 times

 

At double: 

P100 is 5 tflops

Firepro w9100 (or whatever the firepro 290x is called) is 2.5 tflops

P100 wins by 2 times

 

If those numbers are true, then it is the best floating-point ratio (2:1) for NVidia GPUs in quite a while.

[Kinda Bottlenecked]

Just now, Castdeath97 said:

So Nvidia still haven't gotten bored from trying to fit tegra's in cars, *sigh*. All these rumours about mobile GPUs coming out turned out to be false then.

 

I think they're trying to put pascal(or tesla, I don't remember) inside it now 

[Van Weapon]

If there's one thing we can all take away from this... Deep Learning 

[Castdeath97]

3 minutes ago, Pohernori said:

 

I think they're trying to put pascal(or tesla, I don't remember) inside it now 

What's next, Gameworks in cars!

[ivhyn]

Just now, Castdeath97 said:

What's next, Gameworks in cars!

There's driveworks iirc

[Castdeath97]

1 minute ago, ivhyn said:

There's driveworks iirc

Maybe we can have hairworks in cars!

[Van Weapon]

19 minutes ago, Castdeath97 said:

So can anyone give me a TL DR for the event since the OP hasn't updated his post with anything.

Deep Learning

[alamox]

10 minutes ago, i_build_nanosuits said:

yeah, one post above it's all there...these are the specs for the testla P100 that will start shipping soon for high performance computing...and will most likely be the next big pascal titan gaming GPU for next year of something.

And all the rest is about deep learning, neuro networks...and other skynet related stuff

no the tesla P100 will start shipping in Q1 2017, what they are shipping now is sample for ppl who make servers to prepare for next year, nvidia need a year of saving up chips to sell, because the yields of a new process with a 600mm² should be so bad, that most of them will be defective, and since the cost include the defects and the good chips, they cannot make it a consumer product like titan, unless you are prepared to pay 10-15k$ for a gpu.

[i_build_nanosuits]

3 minutes ago, alamox said:

no the tesla P100 will start shipping in Q1 2017, what they are shipping now is sample for ppl who make servers to prepare for next year, nvidia need a year of saving up chips to sell, because the yields of a new process with a 600mm² should be so bad, that most of them will be defective, and since the cost include the defects and the good chips, they cannot make it a consumer product like titan, unless you are prepared to pay 10-15k$ for a gpu.

so how long do you think before the big pascal chip meant for consumer grade cards (next titan) could be hitting the market?

2 years from now?!

[VagabondWraith]

The real question is: how many ACE's will Pascal have?

[NumLock21]

5 hours ago, larsi239 said:

Source?

http://wccftech.com/nvidia-pascal-gpu-deep-neural-ai-performance/

edit: @drinkO beat me to it.

[zMeul]

1 hour ago, VagabondWraith said:

The real question is: how many ACE's will Pascal have?

none

ACE's are AMD's territory, GeForces have CUDA Cores

[MMKing]

10 minutes ago, zMeul said:

none

ACE's are AMD's territory, GeForces have CUDA Cores

ACE refers to  ''Asynchronous Compute Engine'', of which the AMD Fury x has 8. For the express purpose of having different tasks run... you guessed it, asynchronously. On the Fury x diagram, there is also a single ''Graphics Command Proccesor''.

 

Nvidias counterpart on the GM200 chip (titan / 980Ti) would be the ''Gigathread Engine''.

 

 

I doubt Nvidia would use the same name if they opted for running graphics task asynchronously, but i have not heard any information about it.

[Agosto]

Anything new about the ROP count?

If - and I say IF - AMD manages to increase the performance of single next-gen GCN cores (higher than Maxwell levels) and pack enough ROPs, Vega could beat GP100... I sincerely hope so.

[Big-G]

LOL... I seriously have to laugh at the failed attempts from AMD fangirls to try and discredit Pascal...

 

O and those complaining at no consumer stuff... try keep your panties on till Computex

[i_build_nanosuits]

4 hours ago, Agost said:

Anything new about the ROP count?

If - and I say IF - AMD manages to increase the performance of single next-gen GCN cores (higher than Maxwell levels) and pack enough ROPs, Vega could beat GP100... I sincerely hope so.

it's a MONSTER:

 

http://wccftech.com/nvidia-pascal-gpu-gtc-2016/

 

The GP100 GPU is comprised of  3840 CUDA cores, 240 texture units and a 4096bit memory interface. The 3840 CUDA cores are arranged in six Graphics Processing Clusters, or GPCs for short. Each of these has 10 Pascal Streaming Multiprocessors. As mentioned earlier in the article the Tesla P100 features a cut down GP100 GPU. This cut back version has 3584 CUDA cores and 224 texture mapping units.

 

Each Pascal streaming multiprocessor includes 64 FP32 CUDA cores, half that of Maxwell. Within each Pascal streaming multirprocessor there are two 32 CUDA core partitions, two dispatch units, a warp scheduler and a fairly large instruction buffer, matching that of Maxwell.

 

The massive GP100 GPU has significantly more pascal streaming multiprocessors, or CUDA core blocks.  Because each of these has access to a register file that’s the same size of Maxwell’s 128 CUDA core SMM. This means that each Pascal CUDA core has access to twice the register files. In turn we should expect even more performance out of each Pascal CUDA cores compared to Maxwell.

• 5.3 teraflops double-precision performance, 10.6 teraflops single-precision performance and 21.2 teraflops half-precision performance with NVIDIA GPU BOOST™ technology

• 160GB/sec bi-directional interconnect bandwidth with NVIDIA NVLink

  Advertisements

• 16GB of CoWoS HBM2 stacked memory

• 720GB/sec memory bandwidth with CoWoS HBM2 stacked memory

• Enhanced programmability with page migration engine and unified memory

• ECC protection for increased reliability

• Server-optimized for highest data center throughput and reliability

  Tesla P100 Boosts To Nearly 1.5Ghz

  Perhaps one of the most exciting, yet perhaps predictable, revaluations about the GP100 Pascal flagship GPU is that it can achieve clocks even higher than Maxwell. Despite Nvidia opting for very conservative clock speeds on its professional GPUs like the Tesla & Quadro products the P100 actually has a base clock speed of 1328mhz and a boost clock speed of 1480mhz. Considering that GPU Boost 2.0 actually allows these cards to operate at even higher clock speeds than the nominal boost clock.

  We’re looking at actual frequencies of upwards of 1500Mhz on the GeForce equivalent of the P100. What is inevitably going to launch as the next GTX Titan. This means boost clocks of even upwards of 1600Mhz on factory overclocked models, and perhaps 2Ghz+ manual overclocks. This should be extremely exciting news to all GeForce fans.

Nvidia Pascal – 2X Perf/Watt With 16nm FinFET, Stacked Memory ( HBM2 ), NV-Link And Mixed Precision Compute

[alamox]

37 minutes ago, i_build_nanosuits said:

it's a MONSTER:

 

http://wccftech.com/nvidia-pascal-gpu-gtc-2016/

 

The GP100 GPU is comprised of  3840 CUDA cores, 240 texture units and a 4096bit memory interface. The 3840 CUDA cores are arranged in six Graphics Processing Clusters, or GPCs for short. Each of these has 10 Pascal Streaming Multiprocessors. As mentioned earlier in the article the Tesla P100 features a cut down GP100 GPU. This cut back version has 3584 CUDA cores and 224 texture mapping units.

 

Each Pascal streaming multiprocessor includes 64 FP32 CUDA cores, half that of Maxwell. Within each Pascal streaming multirprocessor there are two 32 CUDA core partitions, two dispatch units, a warp scheduler and a fairly large instruction buffer, matching that of Maxwell.

 

The massive GP100 GPU has significantly more pascal streaming multiprocessors, or CUDA core blocks.  Because each of these has access to a register file that’s the same size of Maxwell’s 128 CUDA core SMM. This means that each Pascal CUDA core has access to twice the register files. In turn we should expect even more performance out of each Pascal CUDA cores compared to Maxwell.

• 5.3 teraflops double-precision performance, 10.6 teraflops single-precision performance and 21.2 teraflops half-precision performance with NVIDIA GPU BOOST™ technology

• 160GB/sec bi-directional interconnect bandwidth with NVIDIA NVLink

  Advertisements

• 16GB of CoWoS HBM2 stacked memory

• 720GB/sec memory bandwidth with CoWoS HBM2 stacked memory

• Enhanced programmability with page migration engine and unified memory

• ECC protection for increased reliability

• Server-optimized for highest data center throughput and reliability

  Tesla P100 Boosts To Nearly 1.5Ghz

  Perhaps one of the most exciting, yet perhaps predictable, revaluations about the GP100 Pascal flagship GPU is that it can achieve clocks even higher than Maxwell. Despite Nvidia opting for very conservative clock speeds on its professional GPUs like the Tesla & Quadro products the P100 actually has a base clock speed of 1328mhz and a boost clock speed of 1480mhz. Considering that GPU Boost 2.0 actually allows these cards to operate at even higher clock speeds than the nominal boost clock.

  We’re looking at actual frequencies of upwards of 1500Mhz on the GeForce equivalent of the P100. What is inevitably going to launch as the next GTX Titan. This means boost clocks of even upwards of 1600Mhz on factory overclocked models, and perhaps 2Ghz+ manual overclocks. This should be extremely exciting news to all GeForce fans.

Nvidia Pascal – 2X Perf/Watt With 16nm FinFET, Stacked Memory ( HBM2 ), NV-Link And Mixed Precision Compute

you understand that the numbers are bad for a chip of that size ? it's a freaking 600mm² chip with barely 20% more compute than a furyX in single precision about the same double precision as a 295x2, on a node twice as small, if the tesla P100 was 300mm² size that would be impressive, but those numbers are bad news for a 16nm 600mm² chip, especialy at that clock 1328mhz with 300TDP, seriously how big a chip should be to have double the compute ?

the expected compute from 16nm 600mm² chip should be around 18/20teraflops single precision, and about 10teraflops double precision, and even more because of the clock speed

[79wjd]

19 minutes ago, alamox said:

you understand that the numbers are bad for a chip of that size ? it's a freaking 600mm² chip with barely 20% more compute than a furyX in single precision about the same double precision as a 295x2, on a node twice as small, if the tesla P100 was 300mm² size that would be impressive, but those numbers are bad news for a 16nm 600mm² chip, especialy at that clock 1328mhz with 300TDP, seriously how big a chip should be to have double the compute ?

the expected compute from 16nm 600mm² chip should be around 18/20teraflops single precision, and about 10teraflops double precision, and even more because of the clock speed

A 295x2 is 50% larger at 900mm and consumes twice as much power while providing half the performance. 

 

So I don't really think that pascal's performance is "bad news".

[Notional]

11 minutes ago, djdwosk97 said:

A 295x2 is 50% larger at 900mm and consumes twice as much power while providing half the performance. 

 

So I don't really think that pascal's performance is "bad news".

Considering it's a 2013 architecture, then yeah kinda. You have to compare to contemporary architectures. Fiji and Vega makes this look less than impressive.

[79wjd]

1 minute ago, Notional said:

Considering it's a 2013 architecture, then yeah kinda. You have to compare to contemporary architectures. Fiji and Vega makes this look less than impressive.

I never said I was impressed by pascal. Fiji has 4.3 tflops, is 600mm, and consumes about 300w. So pascal is about 20% faster than Fiji, that's still a decent generational improvement, although I still expect Polaris to pull ahead in compute. 

 

[Notional]

Just now, djdwosk97 said:

I never said I was impressed by pascal. Fiji has 4.3 tflops, is 600mm, and consumes about 300w. So pascal is about 20% faster than Fiji, that's still a decent generational improvement, although I still expect Polaris to pull ahead in compute. 

 

Never claimed you did?

However considering this chip is larger than Fiji, and using a node almost half as large as Fiji's, including a much higher clockrate, it's really not impressive. The more I think about it, the less impressive Pascal is.

[79wjd]

5 minutes ago, Notional said:

Never claimed you did?

However considering this chip is larger than Fiji, and using a node almost half as large as Fiji's, including a much higher clockrate, it's really not impressive. The more I think about it, the less impressive Pascal is.

The chip itself is virtually the same size -- both are about 600mm. And I don't think clock speeds mean all that much as long as power consumption (and heat) and performance are decent. Now the fact that it's clocked so high and on a significantly smaller node and only yields a 20% performance gain doesn't necessarily bode well for gaming performance (assuming gaming performance follows in compute's footsteps -- which isn't necessarily the case).

 

Also, are there other factors to consider aside from raw tflops/power consumption/heat when it comes to compute that could affect performance? 

[Notional]

3 minutes ago, djdwosk97 said:

The chip itself is virtually the same size -- both are about 600mm. And I don't think clock speeds mean all that much as long as power consumption (and heat) and performance are decent. Now the fact that it's clocked so high and on a significantly smaller node and only yields a 20% performance gain doesn't necessarily bode well for gaming performance (assuming gaming performance follows in compute's footsteps -- which isn't necessarily the case).

 

Also, are there other factors to consider aside from raw tflops/power consumption/heat when it comes to compute that could affect performance? 

In gaming, sure, compute won't be the be all end all. We've seen this with Maxwell contra Fiji in DX11, but DX12 will use async compute more and more, so gaming cards are wise to support it extensively.

 

As for compute, it also depends on how you program for it, and what type of programming you use, etc, etc. Are we talking general compute or async compute for gaming?

 

Point is, that if Pascal is "only" Maxwell with some added compute functionality and a die shrink, then it's not improbably to see AMD with a new architecture, excellent compute capabilities and a die shrink, beating out Pascal.

 

But afaik, we won't see Big Pascal or Vega until December/January, so it's difficult to say anything about the koth cards.

[goaway]

I really hope the only reason Nvidia did not announce any new pascal gaming cards is that they did not want to harm the sales of their 900-series, and they want to wait with the announcement right until they are ready to ship the cards - crossing my fingers for computex. I mean, they probably could have said everything about pascal if they wanted to, they even had two pascal modules ready for the drive px 2 system they showed. 

[Spectrez]

So basically not a word on their gaming cards?