NVIDIA NVLink Switch Chips Change to the HGX B200

[erek]

Hmm

"We asked NVIDIA to confirm these are the new NVSwitches, we received this response via e-mail “We haven’t disclosed the positioning of the chips, but we did go from 4 chips to 2.” At the same time, the company also corrected our branding when we called those two “NVSwitches” since “… the two chips in the middle are called “NVLink Switch” chips.

While NVIDIA has not disclosed the location of the NVLink Switch chips, there are two of these chips in the middle of the HGX B200 baseboard, and it was also confirmed on the HGX B100.

Final Words​

The NVLink Switch chips are bigger now and moving them to the center of the board likely reduced trace lengths versus having them all at one end. That has benefits when doing high-speed signaling. Further, the NVIDIA B200 GPUs have the sets of four GPUs flipped on either side, so it is likely NVIDIA also is doing that to further reduce trace lengths to the NVLink Switches."

Source: https://www.servethehome.com/ingras...ink-switch-chips-change-to-the-hgx-b200-b100/

 

[Zarathustra[H]]

Does anyone actually use Nvlink for anything? Outside of - you know - the data enter?

 

[Lakados]

Does anyone actually use Nvlink for anything? Outside of - you know - the data enter?

There are plenty of workstation uses for it to connect 2-4 cards together. CUDA does more than just AI, and the physics and simulation libraries in there are still leaps and bounds ahead of most of the CL based alternatives out there.

But as individual cards get faster the need for it decreases.

I mean if. Single card does it in 10-12 hours, but 2 cards manage it in 5-6 hours that’s fantastic! But useless if it’s a job unsupervised run after hours because you’re still not going to be there. Granted that means you could use 4 get it done over lunch but the costs associated with that improvement are large and the benefit nearly nothing. So it’s an upgrade that’s very hard to justify. But if it lets you take your job from a weekend to a single night that’s another.
So it really just depends on the scope of work.

 

[Ski]

Does anyone actually use Nvlink for anything? Outside of - you know - the data enter?

3d artists who want to maximize the VRAM and bandwidth it provides because that is one of GPU's biggest weakest links is its limted RAM limitations. If you're an architectural visualization artist like I am, it's not hard to push over 128GB of RAM for certain scenes. So that's one example.