AMD Instinct MI350P Brings CDNA 4 in PCIe Format with 144 GB of HBM3e

AMD has announced the new Instinct MI350P, a solution aimed at the artificial intelligence market designed for traditional server systems equipped with PCIe slots and air cooling. The product expands the Instinct MI350 family by introducing a different variant compared to the OAM modules previously used by the family of accelerators based on the CDNA 4 architecture.

The new MI350P arrives just weeks before the expected debut of the upcoming MI400 family and represents AMD's return to the high-performance PCIe AI GPU segment. The last similar offering dates back to 2022 with the Instinct MI210, a solution based on the CDNA 2 architecture characterized by a TDP of 300 watts.

With the MI350P, power consumption rises to 600 watts, a value that AMD still considers compatible with air-cooled server infrastructures already present in data centers. The card is indeed designed to operate by utilizing the internal airflow of large enterprise chassis, avoiding the need for liquid systems or proprietary configurations typical of more advanced AI platforms.

From an architectural standpoint, the MI350P adopts an approach already seen in the past within the Instinct line. AMD has developed the product using only half of the full MI350 package. The original GPU integrates two sections called IOD0 and IOD1, each connected via Infinity Fabric and equipped with four XCDs. In the MI350P, only the IOD0 section is used, effectively reducing the available computational resources and memory.

The technical specifications directly reflect this design choice. The new accelerator offers 128 Compute Units compared to the 256 CUs of the complete MI350 versions, while the HBM3e memory decreases to 144 GB, exactly half compared to the higher models in the range.

This downscaling inevitably results in a performance drop compared to the MI355X solutions, which can reach power consumption of up to 1400 watts, but allows AMD to offer an accelerator that is more easily integrable into existing servers. The company also focuses its communication on the FP4 performance dedicated to modern AI workloads, while FP64 scenarios are not highlighted, historically central in the previous Instinct generations oriented towards traditional HPC computing.

AMD has not announced official prices for the new Instinct MI350P, but availability among partners should begin shortly.