AMD is currently in the spotlight in the AI sector, thanks in part to a recent agreement with OpenAI. This agreement includes the purchase of numerous AMD graphics cards to equip the AI company’s future servers. Among these cards is the Mi450, which Jensen Huang ironized by saying that they (AMD) weren’t yet producing, but was already selling the house. Anyway, today we learn that the Reds are teaming up with Samsung and opting for their HBM4 memory chips!
The Mi450 will run on Samsung’s HBM4?
Broadly speaking, the board will be aimed at the artificial intelligence sector, and should feature a very large configuration. Although the latter is not yet known, rumors point to a massive memory capacity of up to 432 GB of HBM4 type. The bandwidth is said to be massive, at 19.6 TB/s, and Samsung is said to be the main supplier of this memory. The composition of the GPU is not known, but production is expected to be handled by TSMC, using a 2 nm process.
As for performance, rumors point to 20 PFLOP in FP8, and up to 40 PFLOPS in FP4. On the other hand, power consumption is said to be high: 1400W at peak. In itself, the first GW of cards is likely to be quickly supplied. Finally, to keep the heat under control, we’re talking about an aircooling or even liquid cooling system, which would enable the deployment of configurations with up to 128 GPUs. While NVIDIA takes advantage of NVLink for GPU interconnection, AMD is said to be relying on an open standard called UALink.
With this board, AMD seems to have confidence in its product, stating on September 12 that “there’s no excuse for choosing Team Red over NVIDIA in AI”. We’ll see.
