This is perhaps one of the coolest server releases that we have seen so far this year. Gigabyte has a new Arm-based server for NVIDIA GPUs. The Gigabyte G492-PD0 combines the Ampere Altra Max (up to 128 cores) plus 8x NVIDIA A100’s for an AI training server.
Gigabyte G492-PD0 Ampere Arm and NVIDIA Server
The new Gigabyte G492-PD0 is a 4U server that takes a lot of design features from the company’s x86 offerings. It essentially has the 8x NVIDIA A100 Delta platform that we saw in the Inspur NF5488A5 8x NVIDIA A100 HGX Platform and the two we saw in our Liquid Cooling Next-Gen Servers Getting Hands-on with 3 Options pieces and swaps out the CPU side.
Like other NVIDIA HGX A100 Delta platforms, we have the NVIDIA provided 8x GPU HGX board on the bottom with its own cooling. On the top, we have the CPU and storage. This is a very common configuration for OEM systems, as is the GPU tray that holds the entire HGX A100 assembly.
The server itself has a single Ampere Altra Max processor for up to 128 cores, but that is notable. The x86 systems usually have two processors because a lot of AI training workloads have pre-processing steps that use high-performance cores. Still, the use of the Ampere Altra is an Arm-based high-end server that is leaning heavily on accelerators. For those thinking who may want a server like this, remember NVIDIA Grace is coming out next year but between now and then, running high-performance GPUs with an Arm processor is a challenge. This is perhaps the server to work on between now and Grace.
On the rear of the system, we see 4x 3kW PSUs and 6x fan modules, but the other feature that is notable is the array of eight PCIe Gen4 x16 slots. These are notable since filling these slots with PCIe lanes would mean no more lanes for the GPUs or storage on an Ampere Altra.
What Gigabyte is actually doing is using Broadcom PEX88096 switches. Each switch has two GPUs, up to two NVMe SSDs, and two of these rear slots.
NVIDIA has done a lot of work to offload tasks from CPUs with features like GPUdirect storage, but this is another level. This is a system specifically targeted to de-emphasize the CPU in its operation, which is why the Ampere Altra is a unique choice here.
Gigabyte has historically focused a lot of effort on Arm server designs. From the Gigabyte Cavium ThunderX days and even the Annapurna Labs Arm storage server days. As a result, the shift to more Arm servers as cloud workloads seek lower performance per core, but more cores is one that the company is uniquely positioned to target. We hope that this system is a precursor to future Grace + Hopper systems in 2023 since that is when we expect Arm and NVIDIA GPUs to really start to take off. Still, if you need something today that is Arm + NVIDIA GPU, and you want something bigger than NVIDIA Orin, then this is an exciting option.