Overall, I'm a bit underwhelmed.

31.34 SP and 15.67 DP GFLOPS/W for Pascal Tesla with HBM RAM don't sound that great when the 28nm FirePros are at ~20 SP and ~10 DP GFLOPS/W . Seems like all the machinery to do double precision is costing them a lot of transistors and power consumption. Of course, CUDA has been the preferred architecture for scientific workloads for a while, but that could shift a bit of marketshare, if AMD manages more than a 50% performance per watt increase with Polaris Server GPUs.

Self-Driving Cars will kill us.

The GP100 is looking sexy ...

Indeed.

Nvidia did show Pascal, but only in their Tesla configurations. At least now we know what the new Titan could be like .

Btw, have you seen that Realities VR game video? I posted it in the news and thought you may like it given that you made that thread about UE4 and photogrammetry pics.

Nvidia has given some details and a comparison chart of Pascal through their blog:

https://devblogs.nvidia.com/parallelforall/inside-pascal/

Like previous Tesla GPUs, GP100 is composed of an array of Graphics Processing Clusters (GPCs), Streaming Multiprocessors (SMs), and memory controllers. GP100 achieves its colossal throughput by providing six GPCs, up to 60 SMs, and eight 512-bit memory controllers (4096 bits total). The Pascal architecture’s computational prowess is more than just brute force: it increases performance not only by adding more SMs than previous GPUs, but by making each SM more efficient. Each SM has 64 CUDA cores and four texture units, for a total of 3840 CUDA cores and 240 texture units.

_{*I've put in bold the aspects that I think are more important for gaming.}

Also, videocardz says that this Tesla P100 is not using the full chip.

My excitement died down after reading the specs ...

fatslob-:O said: My excitement died down after reading the specs ...

Don't be so negative.

It's true that Pascal focuses more on Nvidia's computing business than Maxwell, but I'm confident it will still bring some serious improvements over the current cards.

*Edit: I'm working on a better table to be able to compare Pascal/P100 with previous Nvidia cards, unless I find something on the web.

This reminds me of that 3D NES emulator that came up a few weeks before. Doesn't look that bad.

I wonder if Nintendo could implement these things in their VC system, and enhance the roms with different modes such as these voxels or implement vector graphics?

It would be cool if they did. But since it's Nintendo, the ywon't.

Ok, so here's a kind of recap about Pascal and the Tesla P100.

This is Tesla P100:

And this is Pascal's P100 core diagram:

First things first: The chip has 8x512-bit memory controllers for a total of 4096-bit memory bus width. This is because the P100 uses 4xHBM2 memory stacks instead of GDDR5.

The GPU itself is compromised of 6 GPCs (Graphics Processor Cluster), with each cluster comprised of 10 SMs (Streaming Multiprocessor). This is a departure from Maxwell and its 4 SMs for each GPC, of which also had 6 GPUnits.

Now let's take a closer look at a SM unit

As you can see, each Pascal SM is comprised of 64 CUDA cores (or shaders) and 4 Texture Units, whereas Maxwell also had 4 Texture Units but the Shader Units count was twice as much with 128.

That's because Nvidia has focused a lot on the computing side of things, improving its single precision (FP32) performance but specially it's Double Precision (FP64) results, achieving a whooping 1:2 ratio between single and double precision compared to the paltry 1:32 ratio of Maxwell (that was beated by Kepler).

Finally, the chip features 14MB of shared register files and 4MB of L2 cache.

And here is a table comparing the Tesla P100 with the previous Tesla products and top-end cards:

Of course, anyone expecting the upcoming GTX 1080/1070 (or whatever they are called) to be as big as P100 will be disappointed. Expect the 1080/1070 to be based on a GP 104 chip with the number of Graphics Processor Clusters reduced to 4, and the 1060/1050 to have 2 GPCs.