NVIDIA H100 Hopper GPU Tested for Gaming, Slower Than Integrated GPU

[AleksandarK]

NVIDIA's H100 Hopper GPU is a device designed for pure AI and other compute workloads, with the least amount of consideration for gaming workloads that involve graphics processing. However, it is still interesting to see how this 30,000 USD GPU fairs in comparison to other gaming GPUs and whether it is even possible to run games on it. It turns out that it is technically feasible but not making much sense, as the Chinese YouTube channel Geekerwan notes. Based on the GH100 GPU SKU with 14,592 CUDA, the H100 PCIe version tested here can achieve 204.9 TeraFLOPS at FP16, 51.22 TeraFLOPS at FP32, and 25.61 TeraFLOPS at FP64, with its natural power laying in accelerating AI workloads.

However, how does it fare in gaming benchmarks? Not very well, as the testing shows. It scored 2681 points in 3DMark Time Spy, which is lower than AMD's integrated Radeon 680M, which managed to score 2710 points. Interestingly, the GH100 has only 24 ROPs (render output units), while the gaming-oriented GA102 (highest-end gaming GPU SKU) has 112 ROPs. This is self-explanatory and provides a clear picture as to why the H100 GPU is used for computing only. Since it doesn't have any display outputs, the system needed another regular GPU to provide the picture, while the computation happened on the H100 GPU.



View at TechPowerUp Main Site | Source

 

[Dirt Chip]

That's like measuring how much good a Ferrari at squeezing oranges..

 

[bob3002]

That's like measuring how much good a Ferrari at squeezing oranges..

If we're going by automotive analogies, it's like buying the most expensive cargo-hauliing semi-truck... and then taking it to the dragstrip and measuring how fast it does a quarter mile.

 

[fancucker]

speaking of datacenter gpus, any benchmarks for the super hyped up mi300? will it be competitive while being restricted to on-die memory/cxl?

 

[ZoneDymo]

I mean from the article it seems everyone is well aware of this....it's just for fun, so I don't really get these comments

 

[P4-630]

I'm sold, I'm gonna buy one!!

 

[kondamin]

other than the differently abled demand for these cards.
What is the reasoning behind 30000 for a card like this?

 

[Udyr]

I mean from the article it seems everyone is well aware of this....it's just for fun, so I don't really get these comments

The comments are just for fun... like the article.

 

[Double-Click]

other than the differently abled demand for these cards.
What is the reasoning behind 30000 for a card like this?

 

[wg106]

This guy mastered the art of doing useless work with maximum efficiency and precision!

 

[DemonicRyzen666]

The R.O.P Thing kind proves that all GPU's currently out right now are bottlenecked by lack of R.O.P count.
If we want really good cards like they should be for 2160p/4K then they need about 256 R.O.Ps. minuim.

 

[TheoneandonlyMrK]

The R.O.P Thing kind proves that all GPU's currently out right now are bottlenecked by lack of R.O.P count.
If we want really good cards like they should be for 2160p/4K then they need about 256 R.O.Ps. minuim.

Those Rop's need to be passed the right data at the right time, so you can't really benefit from exceeding the shader engine's ability to output that Data, also more Rop's, would up the anti equally in all areas, twice the rops as usual, twice the Shader's then, and cache.

But your. Right, next 5060 and 8600Xt better both have 256 Rop's and 30k shaders or gtfo
I do think the Rop thing a bit overplayed though, this thing was outputting through another cards output, and had Zero optimization or official gaming driver support , they're lucky it worked.


Edit £#@n phone's.

And I too will buy one if poss, 30K though :/

 

[thegnome]

This is like testing the AMD threadripper/epyc chips on older games. It's stupidly unnecessary but it sure makes a fun experience

 

[Wirko]

They were holding it wrong. The H100 should be used to run the AI engine in games, not to render the graphics.

 

[Minus Infinity]

Well i just cancelled my order at Newegg. I'm getting the MI300 instead.

 

[DemonicRyzen666]

Those Rop's need to be passed the right data at the right time, so you can't really benefit from exceeding the shader engine's ability to output that Data, also more Rop's, would up the anti equally in all areas, twice the rops as usual, twice the Shader's then, and cache.

But your. Right, next 5060 and 8600Xt better both have 256 Rop's and 30k shaders or gtfo
I do think the Rop thing a bit overplayed though, this thing was outputting through another cards output, and had Zero optimization or official gaming driver support , they're lucky it worked.


Edit £#@n phone's.

And I too will buy one if poss, 30K though :/

You don't need more shaders at all. The majority of shaders sit idle in current designs. That's why Nvidia is using boring things like D.L.S.S & FG to use up shader & wasted Tensor cores also opticalflow acclerators that sit idle without that software. Nearly all the RT & pixel fill work is done with & through the R.O.P'S. Jim at AdordTV has pretty good video explaining it. More RT units & ROPs right now cards like the RTX 4090 should already have gotten to 256 ROPS. AMD's 7900 XT had increase in ROPS of 50% but only 20% increase in shaders. The RTX 4080 matching the 3090 TI is another example of this it has 112 ROP but less shaders /Tensor cores/ RT cores than the 3090 ti. It is proof that shader count doesn't really help.

 

[pf100]

Jim at AdordTV has pretty good video explaining it.

Got a link?

 

[DemonicRyzen666]

Got a link?

I'm not sure if in this one or an older one. >.>......
Investigating Nvidia's Raytracing Performance - YouTube

 

[Dirt Chip]

If we're going by automotive analogies, it's like buying the most expensive cargo-hauliing semi-truck... and then taking it to the dragstrip and measuring how fast it does a quarter mile.

Your analogy makes no sense- you pretend as if both competitors have 4 wheels, when the H100 have non (no display outputs..)
I preferer to space-shuttle analogy it: take an out-of-the-stratosphere-rocket and use it to get to the top floor of the Burj-khalifa. Than compere it to taking the elevator.

 

[pf100]

Your analogy makes no sense- you pretend as if both competitors have 4 wheels, when the H100 have non (no display outputs..)

It doesn't need a display output if you can make it output through another gpu.

 

[Dirt Chip]

It doesn't need a display output if you can make it output through another gpu.

You are right- a car doesn't need 4 wheels as long as you can put it on another car to move around in tandem.

 

[pf100]

You are right- a car doesn't need 4 wheels as long as you can put it on another car to move around in tandem.

The H100 does all the work, then it outputs the video through the other gpu that isn't doing any work and is only displaying the screen. Either you aren't understanding that, or you're just being a troll, and right now I can't tell which one.

 

[Dirt Chip]

Fully understand, that's why you connect the drive-shaft of the above car to the lower car wheels.
And this is a 'fun thread' from top to button- take it easy

 

[pf100]

Fully understand, that's why you connect the drive-shaft of the above car to the lower car wheels.
And this is a 'fun thread' from top to button- take it easy

I apologize if I misunderstood. I must be getting old...

 

[wolf]

@AleksandarK

while the gaming-oriented GA102 (highest-end gaming GPU SKU) has 112 ROPs

I believe the highest end gaming GPU SKU is based on AD102 and has 176 ROPs, unless this comparison is strictly against Ampere in which case that's not clear.