Images of NVIDIA Hopper H100 GPU Compute Processor Surface

Published by

Click here to post a comment for Images of NVIDIA Hopper H100 GPU Compute Processor Surface on our message forum
https://forums.guru3d.com/data/avatars/m/268/268248.jpg
Oh wow that's hbm memory and all the rest phases ! 600 watts start to seem like a real leak :/
https://forums.guru3d.com/data/avatars/m/246/246171.jpg
It's funny, in servers, the higher wattage doesn't really matter so long as the performance-per-watt is better (which it most likely is). On the other hand, most servers don't need that much GPU compute power in a single chip. It would be much more affordable and practical to have multiple dies per PCB, and in most cases, there wouldn't be a significant performance loss. Of course, there are use-cases where a single gargantuan die would make sense.
https://forums.guru3d.com/data/avatars/m/270/270008.jpg
Venix:

Oh wow that's hbm memory and all the rest phases ! 600 watts start to seem like a real leak :/
Those leaks were for graphics cards not this beast. The HPC cards are and will be massively power hungry beasts. This is likely a $10,000 card that I wouldn't be surprised if it is closer to 1000w.
https://forums.guru3d.com/data/avatars/m/270/270008.jpg
schmidtbag:

It's funny, in servers, the higher wattage doesn't really matter so long as the performance-per-watt is better (which it most likely is). On the other hand, most servers don't need that much GPU compute power in a single chip. It would be much more affordable and practical to have multiple dies per PCB, and in most cases, there wouldn't be a significant performance loss. Of course, there are use-cases where a single gargantuan die would make sense.
Higher wattage is if it increases density and performance-per-watt is vastly better. Floor space is a huge premium as well electricity.
https://forums.guru3d.com/data/avatars/m/246/246171.jpg
JamesSneed:

Higher wattage is if it increases density and performance-per-watt is vastly better. Floor space is a huge premium as well electricity.
I know. I feel like you didn't read what I said. If you do multiple dies on a single AIB, you can still get roughly the same performance (better in some cases, worse in others) as a single die but for a significantly lower cost, without using more floor space. Server tasks don't have to deal with the same mGPU difficulties as games, which is why Nvidia still heavily invests in mGPU interconnections in the server space. Having dies this huge is a real unnecessary expense in most cases. I know there are some applications that would warrant a GPU die this huge, but there aren't many. I am not aware of Nvidia releasing any mGPU AIBs recently, and I think they would sell well.
https://forums.guru3d.com/data/avatars/m/288/288652.jpg
28 Phases?? Or did I lose count? That is insane! 😱
https://forums.guru3d.com/data/avatars/m/268/268248.jpg
JamesSneed:

Those leaks were for graphics cards not this beast. The HPC cards are and will be massively power hungry beasts. This is likely a $10,000 card that I wouldn't be surprised if it is closer to 1000w.
well yeah still ... i never saw as many phases so close together :P
https://forums.guru3d.com/data/avatars/m/225/225084.jpg
More Phases than the Enterprise D.........
https://forums.guru3d.com/data/avatars/m/270/270008.jpg
Venix:

well yeah still ... i never saw as many phases so close together 😛
Yeah its a beast.