AMD has initiated sampling of its new Instinct MI450 GPU to leading customers. The company plans to begin production shipments for the Heliros rack platform in the second half of 2026. CEO Lisa Su stated that demand for the MI450 has already exceeded internal expectations for 2027. OpenAI and Meta have signed multi-gigawatt contracts to deploy the chip.
The MI450 is built on the CDNA 5 architecture. It features 432GB of HBM4 memory, which represents a 50 percent increase over the previous 288GB HBM3e capacity. The memory provides a bandwidth of 19.6 TB/s. The GPU delivers 40 PFLOP of FP4 performance and 20 PFLOP of FP8 performance. Expansion bandwidth reaches 300 GB/s.
AMD claims the MI450 offers 1.5 times the memory capacity of competing products. The company also asserts the chip provides 1.5 times the expansion out bandwidth compared to rivals. These figures position the MI450 to compete directly with NVIDIA's Vera Rubin GPU. Anthropic is also building AI infrastructure using the MI400 series.
The broader server market is experiencing supply constraints due to surging demand for AI processors. Intel is upgrading its Xeon server processors to address this gap. AMD expects its data center CPU shipments to grow by more than 40 percent in 2026. This growth is driven by a shift toward higher-margin media processing platforms.
AMD is currently engaging with customers on the next-generation MI500 series. The company has not confirmed the specific launch window for the MI450 beyond the 2026 H2 production start. Details regarding the MI500 series remain in early development stages.
Discussion
0 comments
Log in to join the thread with a thoughtful take, question, or correction.