At Computex 2024, the team saw Wiwynn’s Intel Gaudi 3 server as well as the AMD Instinct MI300X. Both share the OCP universal baseboard (UBB) design, but Intel has a slight wrinkle for scale-out fabric. We figured we would show these off from the show, and as a bonus have some ZutaCore 2-phase direct liquid cooling as well.
Wiwynn GS1800G Intel Gaudi 3 AI Server
Here is the front of the Wiwynn GS1800G. This is a more standard hyper-scale system with easily serviceable trays that are more on a tier with a top-bin AI server, with a twist.
First, this is an Intel Gaudi solution. So as a result, we know the 8x Gaudi 3 UBB assembly is $125K list pricing. That was part of the recent Intel Gaudi 2 8x OAM UBB $65K Gaudi 3 $125K announcement.
That price, on a Gaudi system also includes networking. Here we can see the 6x OSFP connectors that take 200GbE from the AI accelerators and use that to scale its fabric. See our Intel Gaudi 3 piece for more. We even have a short video.
Wiwynn then does something very different. It has different power supplies for the GPU/ AI accelerator portion and then for the main server. You can see that here with the two orange tab power supplies for the CPUs, memory, and NICs and the blue tab power supplies for the AI accelerators.
That allows Wiwynn to tune output voltage for the different subsystems in an AI server and is a really neat solution.
Wiwynn AMD Instinct MI300X
Wiwynn also had another variant of this server with the AMD Instinct MI300X 8x GPU server.
This uses the same OCP UBB tray area, just swapping out the GPUs. Wiwynn also has a NVIDIA HGX version of this chassis.
Wiwynn’s idea is to offer a single chassis design and then allow customers to pick the AI accelerator.
Wiwynn ZutaCore 2-phase DLC Shown
As a quick bonus, the team say the two-phase direct liquid cooling solution.
ZutaCore’s cooling solution allows cooler liquid to then vaporize undergoing a phase change. It can then use a heat exchanger and condense the vapor back to liquid.
This is a way to get some of the benefits of 2-phase immersion cooling, without having to switch to immersion tanks. That means less fluid is needed.
Final Words
Overall, these are neat AI servers. We saw the Gaudi 2 version of these in our Touring the Intel AI Playground – Inside the Intel Developer Cloud piece. If you want to see those in action in Intel’s Developer Cloud, then we also have a video for you:
the title shoud be ‘computex’
The “AI” server chassis are not standard 19″ wide? It looks like it’s placed on a steel plate only, not rack mounted.
Two phase cooling systems can’t die out fast enough. They are toxic to produce and handle. We have good single phase fluid tech now, in cold-plate and full immersion configurations. Even 3M is exiting the PFAS production business by the end of 2025.