At OCP 2024, we managed to grab a few new shots of the NVIDIA HGX B200. While a lot of the focus has been on the GH200 NVL72, GB200 NVL2, and so forth, many of the high-end training systems will still be based on the HGX B200 as the successor to the NVIDIA HGX H100/ H200.
New Shots of the NVIDIA HGX B200
Here is the board that we saw at OCP Summit 2024. As you can see, one of the big changes with this versus the NVIDIA HGX A100/ H100/ H200 is that the NVLink Switch chips moved to the center of the assembly instead of being on one of the sides. This minimizes the maximum link distances between the GPUs and the NVLink Switch chips. You can also see that unlike previous versions that we have seen, there are no heat spreaders on this board’s switch chips.
In case you were wondering, this is an Umbriel GB200 SXM6 8 GPU baseboard with a part number: 675-26287-00A0-TS53.
In addition to the NVIDIA B200 GPus that we have seen before, here are the NVLink Switch chips we had seen before covered with heatspreaders.
Here is the OCP Summit 2024 one, where we can see the big NVLink Switch chip.
The other very notable collection of chips here is the PCIe retimers. We can see that these are Astera Labs retimers, who seem to be the incumbent on the NVIDIA HGX platforms now.
Overall, a few cool new features.
Final Words
We already have two NVIDIA HGX H200 systems in the lab that we will review in the coming weeks, and it looks like another one or two eight GPU servers are inbound. Still, it is cool to see the next-generation parts coming out, especially without their massive coolers. The NVLink Switch chips are a change that some will sleep on as there are only two now, and they have moved around the HGX 8-GPU assembly for the B200 generation.
what’s the difference between HGX & MGX?
@Benito
MGX is a set of reference designs that can be used by OEMs as basis of the systems they provide. It encompasses multiple generations and types of GPUs (both “big” like A/H/B100 and consumer-based like L40), form factors and interconnects. According to the NVIDIA MGX Whitepaper they provide over 100 configurations.
HGX is a more strict set configuration options for a system composed of one or more boxes with 4 to 8 “big” GPUs like A/H/B100(200) with internal NVSwitches.