AMD adds MI350P PCIe GPU and 256-core EPYC 9996 at Advancing AI

Alongside the Helios launch, AMD introduced the Instinct MI350P, a PCIe card meant to add AI acceleration to existing servers, and the EPYC 9996 with 256 cores and 512 threads. AMD's own tests put the MI350P at up to 4.2x the tokens per second per dollar of one NVIDIA H200 NVL GPU, and the release gives no unit or revenue figures for either product.

• Instinct MI350P: PCIe card for adding AI compute to existing infrastructure
• AMD claim: up to 4.2x tokens per second per dollar vs one NVIDIA H200 NVL (AMD internal test, Llama 3.3 70B FP8)
• EPYC 9996: 256 cores / 512 threads, the highest core count AMD says is publicly disclosed for a 1P CPU
• No unit volume, pricing or revenue disclosed