With the CPU space being a bit quiet lately, AMD came out with a couple of new announcements regarding its Instinct family of GPUs and accelerators, specifically the Instinct MI430X and Instinct MI350P.
First, the MI430X was announced about half a year ago, where Team Red’s continuous chase in high-precision computing has come to fruition, especially in FP64 performance, heavily used in simulations, physics models, and scientific workloads where accuracy matters just as much as speed.
The latest news, however, is them “dumpstering” NVIDIA’s upcoming Rubin architecture by up to 600% in classic-HPC workloads. How? Up to 200 TFLOPs of raw FP64 power coupled with 432GB HBM4 memory with nearly 20TB/s bandwidth. And they are pretty in their position and strategy, as the MI430X isn’t strictly about “AI Factory”, but rather it is lined up for next-gen supercomputers and exascale systems that prioritize both memory capacity and double-precision compute.
The accelerator will be going into the Discovery Supercomputer at Oak Ridge National Laboratory in the USA, slated for 2028, as well as projects involving the DOE in fields of biology, energy, advanced materials, manufacturing, and even national security.
Meanwhile, the Instinct MI350P is about practicality because the main point is about the PCIe interface, which means any existing servers can just get it, drop it in, and get a boost in computing power thanks to its dual-slot, air-cooled design – no custom infrastructure required.
It is not shabby, either, as it is given 144GB of HBM3E memory with up to 4TB/s bandwidth that stretches AI performance close to the multi-PFLOP range via FP4 and FP8, which makes it highly applicable for large inference workloads, including modern LLM deployments.
Ease of installation and powerful performance within this form factor also make it a great choice for on-site enterprise AI deployment with efficiency and throughput in mind.
