As long as we’re having fun speculating about where AEHR might be headed, can anyone shed light on the use of photonics in NVIDIA Infiniband interconnections?
Infiniband is an international standard for very high speed connection between computers. It competes with 10 Gigabit Ethernet, which is more widely used in supercomputers.
You may have heard the word “Infiniband” tossed around in the course of an analysis of what’s going on for NVIDIA. Here’s how they tie together:
NVIDIA makes Graphics Processing Units (GPUs), that are able to do the kind of highly parallel operations that are needed in Artificial Intelligence (AI). NVIDIA had the vision to see where AI might be headed, and has consequently built their architecture – hardware and software – to support the transformation of entire data centers into a single enormous AI compute cluster.
“Huh, what??!”, you say?
A single NVIDIA H200 system combines a boatload (that’s a technical term) of GPUs into a single compute platform for doing AI computation: 256 Grace Hopper SuperChip GPUs, all tightly interconnected for maximum performance.
However, NVIDIA’s architecture doesn’t stop there. If you want your entire data center full of NVIDIA H200s to operate as a single compute platform, you can tie them together with Infiniband to multiply the size of your compute platform. Now, you’ve got all your GPUs across your entire data center tightly coupled together, operating like a single compute platform for your AI.
Infiniband is the communications mechanism that provides this interconnection between the different H200 systems in your data center. Infiniband operates much faster than 10 Gigabit Ethernet – and when using fiber optic cable, can connect systems that are up to 10 km distant from each other (as opposed to the copper cable version of Infinband, which can only span 10 meters (aka 30 feet).
So if you want to have a datacenter filled to the brim with NVIDIA GPUs operating as a single huge system, you’ll interconnect them with NVIDIA’s Infiniband offering over fiber optic cable (they bought the last independent Infiniband vendor, Mellanox, in 2019).
If demand for interconnects between NVIDIA systems jumps as sharply as NVIDIA seems to expect, this could mean a new market opportunity for AEHR, to test the photoelectronic capabilities of the NVIDIA wafers.
It’s been at least 20 years since I last heard anyone talking about Infiniband, so could someone who is more up to date confirm what I’m saying here?
And can any of the folks who are more knowledgeable than I am on NVIDIA and AEHR chime in? Speculating is fun, but I prefer hard facts, where they’re available.