Nvidia's next-gen Pascal GPU 10x the performance 8-way SLI

[ChiefGunny]

If Nvidia's beastly new Titan X graphics card wows you—and it should—well, hold onto your pants: The company said it's next-gen GPU technology, codenamed Pascal, will offer roughly ten times the performance of the Titan X, at least in compute tasks.

 

http://www.pcworld.com/article/2898175/nvidias-next-gen-pascal-gpu-will-offer-10x-the-performance-of-titan-x-8-way-sli.html

[The Official Czex]

If Nvidia's beastly new Titan X graphics card wows you—and it should—well, hold onto your pants: The company said it's next-gen GPU technology, codenamed Pascal, will offer roughly ten times the performance of the Titan X, at least in compute tasks.

 

http://www.pcworld.com/article/2898175/nvidias-next-gen-pascal-gpu-will-offer-10x-the-performance-of-titan-x-8-way-sli.html

 

No fucking way.

 

What's its release date?

[ChiefGunny]

2016

 

NVIDIA Pascal GPU Might Feature 17 Billion Transistors, Almost Twice The Transistors of Fiji

Read more: http://wccftech.com/nvidia-pascal-gpu-17-billion-transistors-32-gb-hbm2-vram-arrives-in-2016/#ixzz3lxLqsM10

[Aytex]

10x my ass size

[Casperian]

If Nvidia's beastly new Titan X graphics card wows you—and it should—well, hold onto your pants: The company said it's next-gen GPU technology, codenamed Pascal, will offer roughly ten times the performance of the Titan X, at least in compute tasks.

 

http://www.pcworld.com/article/2898175/nvidias-next-gen-pascal-gpu-will-offer-10x-the-performance-of-titan-x-8-way-sli.html

I watched that at GPU Tech conf (I think it was there) and NVidia's CEO said 10x was "CEO math".

It comes from 5x the neural network compute performance combined with the fp16. Since 2x the fp16s can be processed at once, it gives 2x the performance. 5x2 = 10 so 10x.

And IIRC is was only for neural networks not just everything.

 

Also NVLink is not SLI.

[Leonard]

Wowo....then in 2017, Nvidia announces.......THIS IS THE BADDEST FASTEST COOLEST GPU EVER.......on to 2018.....and so on and so on......

[Notional]

Lol those 10x was literally "CEO Math", as Jen himself said. It's in extremely specific tasks, that just proves how gimped Kepler and Maxwell really are. Right now, there is no indication that Pascal will be useful for async shading either, but hopefully it will be.

 

Also this "news" is from MARCH!​

[carolkarine]

Lol those 10x was literally "CEO Math", as Jen himself said. It's in extremely specific tasks, that just proves how gimped Kepler and Maxwell really are. Right now, there is no indication that Pascal will be useful for async shading either, but hopefully it will be.

 

Also this "news" is from MARCH!​

 

kepler was fine, it's just maxwell that's got really terrible compute perf.

[azurezeed]

Because Maxwell is just bad at compute

[VagabondWraith]

I'll be happy with 2x performance to be honest. Although I don't think anyone would complain if Pascal delivered Ultra 60fps 4K across the board. In reality, that might be Volta or possibly the one thereafter.

[patrickjp93]

I watched that at GPU Tech conf (I think it was there) and NVidia's CEO said 10x was "CEO math".

It comes from 5x the neural network compute performance combined with the fp16. Since 2x the fp16s can be processed at once, it gives 2x the performance. 5x2 = 10 so 10x.

And IIRC is was only for neural networks not just everything.

 

Also NVLink is not SLI.

Yes it is. NVLink is coming in 2 forms, the one IBM will be using for direct CPU-GPU interaction for enterprise computing, and the one Nvidia is using to replace the inter-GPU communication protocol that is SLI.

[Xorbot]

Finally, the new GTX 1080 based on Pascal will give us the ability to run games at 1080p!

[Casperian]

Pascal was more tailored for neural networks. A convolutional neural network isn't an ordinary compute workload so it is actually possible that Pascal will be 5x faster (CUDNN will likely also be getting optimisation updates for Pascal, so there is that). 

In graphics performance, I'd expect another ~35% increase over Maxwell.

 

 

Yes it is. NVLink is coming in 2 forms, the one IBM will be using for direct CPU-GPU interaction for enterprise computing, and the one Nvidia is using to replace the inter-GPU communication protocol that is SLI.

Well, that is true but we won't be having 8-way SLI in desktop machines for a while even if it is simply due to PCIe lanes / physical connectors. Since NV requires 8x Gen3.0 to have SLI verification, 8-way would require 64 PCIe lanes which no CPU currently supports(I don't even think high end Xeons have that many, but they can run multiCPU). And NVLink does more than SLI. SLI shares framebuffers, NVLink is a high speed interconnect.

[patrickjp93]

Pascal was more tailored for neural networks. A convolutional neural network isn't an ordinary compute workload so it is actually possible that Pascal will be 5x faster (CUDNN will likely also be getting optimisation updates for Pascal, so there is that). 

In graphics performance, I'd expect another ~35% increase over Maxwell.

 

 

Well, that is true but we won't be having 8-way SLI in desktop machines for a while even if it is simply due to PCIe lanes / physical connectors. Since NV requires 8x Gen3.0 to have SLI verification, 8-way would require 64 PCIe lanes which no CPU currently supports(I don't even think high end Xeons have that many, but they can run multiCPU). And NVLink does more than SLI. SLI shares framebuffers, NVLink is a high speed interconnect.

I'd expect 65+% over Maxwell purely because it's a double die shrink and a movement to superior memory.

 

SLI in its current form requires the GPUs are all talking to the CPU to share those buffers. Remove that overhead and you remove the need for the CPU to even talk to the extra GPUs, especially with an external high-speed interconnect.

[iHardware Shelden]

Finally, the new GTX 1080 based on Pascal will give us the ability to run games at 1080p!

 

9800p Master Race  B)

[SteveGrabowski0]

So is Pascal for sure going to be 16 nm? They're going to have to offer at least 50% higher performance at the same price and similar TDP on one of their models to get me to upgrade my 970. So maybe I won't buy until Volta (or AMD equivalent). 

[TroubleKlef]

Pascal looks to be so good.

[BuckGup]

Hey whoever has it out by my birthday is the one I am buying Greenland sounds nice

[Suika]

NVIDIA also stated Maxwell could offer 2x the performance per watt. The 970 consumes 25% less power than the 780 yet both yield very similar results.

[Psykomantis00]

If Nvidia's beastly new Titan X graphics card wows you—and it should—well, hold onto your pants: The company said it's next-gen GPU technology, codenamed Pascal, will offer roughly ten times the performance of the Titan X, at least in compute tasks.

 

http://www.pcworld.com/article/2898175/nvidias-next-gen-pascal-gpu-will-offer-10x-the-performance-of-titan-x-8-way-sli.html

 

I came. 

[Zah]

Didn't we already talk about this?

[Briggsy]

I'm not ready to start believing anything from Nvidia about their own hardware. Fool me twice, shame on me...

[BobbyG2]

9800p Master Race B)

I owned that exact card back in the day.

Nvlink interests me. Is it superior to talking over the pcie lanes like AMD? Or only really in enterprise environments with massive amounts of them?

[Casperian]

I owned that exact card back in the day.

Nvlink interests me. Is it superior to talking over the pcie lanes like AMD? Or only really in enterprise environments with massive amounts of them?

Assuming my information is current, NVLink is a 64GB/s interconnect so yes it is superior to PCIe Gen3.0. And IIRC, it should also allow GPU memory to be pooled.

[BobbyG2]

With memory bandwidth approaching 1tb/s with hbm2, that still doesn't sound like enough to make it worth it? Or do I just misunderstand how the memory pooling works?