News Nvidia's H100 Hopper Compute GPU Benchmarked in Games, Found Lacking

[Admin]

Nvidia's H100 card tested in 3DMark, Red Dead Redemption 2, performs worse than integrated GPUs.

Nvidia's H100 Hopper Compute GPU Benchmarked in Games, Found Lacking : Read more

 

[Deleted member 2731765]

H100 is actually structured so that only 2 of its TPCs/(Texture Processing Clusters are available for the standard graphics processing tasks, while the entire GPU block is primarily dedicated to compute tasks so that can lead to adverse results in gaming.

The poor performance was due to under-utilization and the non-optimized nature of the drivers which should be expected since the H100 is an HPC/AI-first solution and the company has no official gaming drivers made for the card.

We can also see that the power of the card is under 100W, which means it is having major under-utilization of the H100 GPU in games.

 

[setx]

So, what is "improper" in Nvidia's consumer GPUs support of FP64/FP16?

 

[Deleted member 2731765]

It appears that the tester was actually running the chip on a standard PC in a 4-Way configuration.

For testing a 3D-Printed duct had to be made to deliver cooling to the card, since it comes with a passive heatsink which means there's no active cooling solution onboard the card.

It was not necessary to use a 4-way setup at least to test games, IMO. For display he had to use the GTX 1650 Ti graphics card as a secondary display card, since the H100 lacks display outputs.

If you watch the video carefully, it looks like using two H100s offered a 43% boost in content creation apps, but, 3-way and 4-way results showed diminishing returns and negative scaling. Looks like standard PCs just can't take advantage of multiple H100 GPUs.

 

[bit_user]

H100 is actually structured so that only 2 of its TPCs/(Texture Processing Clusters are available for the standard graphics processing tasks, while the entire GPU block is primarily dedicated to compute tasks so that can lead to adverse results in gaming.

Similar to the A100, from what I'd heard. My guess is that Nvidia wanted to enable some degree of OpenGL support for HPC users to visualize the data from their simulations.

AMD has gone to the extreme, with their CDNA architecture, and completely omitted texturing, ROPs, etc.

If you watch the video carefully, it looks like using two H100s offered a 43% boost in content creation apps, but, 3-way and 4-way results showed diminishing returns and negative scaling. Looks like standard PCs just can't take advantage of multiple H100 GPUs.

In that picture, there's no over-the-top connectivity (i.e. NVLink) being used. Was there, in the video?

Also... OMG, Legos!
(lower left)

 

[Deleted member 2731765]

Similar to the A100, from what I'd heard. My guess is that Nvidia wanted to enable some degree of OpenGL support for HPC users to visualize the data from their simulations.

AMD has gone to the extreme, with their CDNA architecture, and completely omitted texturing, ROPs, etc.


In that picture, there's no over-the-top connectivity (i.e. NVLink) being used. Was there, in the video?

Also... OMG, Legos!
(lower left)

Yes, NVidia only enabled two TPCs in both the SXM5 and PCIe H100 GPUs which are graphics-capable, so that they can run vertex, geometry, and pixel shaders, which means 2 TPCs (4 SMs) out of 57 TPCs (114 SMs).

But by tradition, Nvidia still calls the H100 a graphics processing unit, but the term is clearly on its last legs: because just two out of the 50+ texture processing clusters (TPCs) in the device are actually cable of running vertex, geometry, and pixel shader maths required to render 3D graphics.

Hey, is that a real LEGO ?