AI BriefingAMD Unveils MI350P PCIe Card
Key point
AMD has unveiled the CDNA 4-based Instinct MI350P PCIe accelerator.
Details
AMD has unveiled the Instinct MI350P based on CDNA 4, bringing back an Instinct PCIe card for the first time in nearly 5 years.
- The goal is on-premise AI inference for environments where rack-scale AI nodes are too much of a burden.
- Scaling down the MI350X by half, it offers 128 CU, 512 Matrix Cores, 144GB HBM3E, and 4TB/s.
- TBP is 600W, with a 450W mode also supported for server constraints.
- The form factor is PCIe Gen5 x16, 10.5-inch FHFL dual slot, and air-cooled passive design.
AMD explained that this product was made as a separate scaled-down design rather than simply a binned part. With a 1 IOD + 4 XCD configuration, it provides half the performance and memory capacity of the MI350X, but the Infinity Fabric link is not exposed, so card-to-card communication is limited to PCIe.
- Rather than combining multiple cards to run one giant model, it's better suited for running multiple models in parallel.
- Partitioned execution is maintained, supporting up to 4 partitions.
- This time, AMD disclosed not just theoretical figures but also delivered performance numbers.
Pricing and release schedule have not yet been announced.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
AMD Publishes Paper on Numerical Behavior Modeling of GPU Matrix CoresvLLM Publishes Performance Benchmarks for 5 Speculative Decoding Methods on AMD GPUs
Hot Chips 2026: CUDA Targets RISC-VGoogle Collaborating with AMD on Next-Gen TPU Design — Potential Hybrid AI ASIC with On-Package CPU Cores for Reinforcement Learning