AMD has officially revealed its Instinct MI350X series GPU, targeting the high-performance artificial intelligence market. The MI350X is built on AMD´s latest CDNA 4 compute architecture, positioning it squarely against NVIDIA´s B200 ´Blackwell´ lineup, particularly with the top-tier Instinct MI355X compared directly to the B200. This launch not only introduces the new silicon architecture but also debuts ROCm 7, AMD´s fresh software stack, as well as a hardware ecosystem adhering to the Open Compute Project specification. Collectively, this ecosystem brings together AMD´s EPYC Zen 5 CPUs, MI350 series GPUs, Pensando Pollara Ultra-Ethernet-capable NICs, and standards-aligned racks and nodes for both air- and liquid-cooled deployments.
The Instinct MI350 is defined by its complex, chiplet-based architecture and stacked silicon design. Central to the GPU are two I/O dies (IODs), engineered using the 6 nm TSMC N6 process, which orchestrate the connectivity for up to four Accelerator Compute Die (XCD) tiles each. These XCDs, fabricated on the advanced 3 nm TSMC N3P node, each house a 4 MB L2 cache and encompass four shader engines with a total of 36 compute units (CU) per XCD. With four XCDs stacked per IOD, each IOD contains 144 CUs, summing to a remarkable total of 288 CUs across the package.
Supporting this computational muscle, each IOD manages four HBM3E memory stacks, amounting to 144 GB per IOD and an impressive 288 GB of high-speed memory for the complete package. Interconnectivity between the two IODs is achieved via a 5.5 TB/s bidirectional link, ensuring full cache coherency and rapid data movement. Additional hardware features include 256 MB of Infinity Cache, a robust Infinity Fabric interface, and a PCI-Express 5.0 x16 root complex, underscoring the platform´s readiness for demanding artificial intelligence workloads in diverse data center environments.
