AMD adds 288 GB Instinct MI450X GPU to its Helios lineup

AMD introduced the Instinct MI450X data center accelerator on November 12. The GPU has 288 GB of high-bandwidth memory and gives buyers of AMD’s Helios systems a second configuration alongside the existing 432 GB MI455X. AMD lists the same peak memory bandwidth for both models, at 23.3 terabytes per second per GPU.

Capacity determines how much room a GPU has for a model and the data it is processing. Bandwidth measures how quickly that data can move. A team serving long requests or many users at once may need extra room even when the memory can move data at the same speed. A deployment that fits within 288 GB now has another option on the same rack architecture and software. AMD did not disclose pricing or delivery timing for the MI450X in its announcement or product specifications.

Two memory capacities in the same rack design

AMD’s specifications show eight HBM4 memory stacks on the MI450X and 12 on the MI455X. A stack combines several memory chips in layers. The new GPU has 144 GB less memory, a third below the MI455X. AMD lists the same peak bandwidth for moving data to and from memory, and the same peak bandwidth for links used to connect GPUs. Those figures describe the hardware’s rated limits.

Both products are direct-liquid-cooled modules designed for running and training models, including fine-tuning. Their specifications list support for frameworks such as PyTorch and JAX. When AMD introduced Helios in July, it described a rack with 72 MI455X GPUs and 18 EPYC server CPUs. The company says the MI450X will use the same open rack architecture and ROCm software. The November announcement gave no rack-wide memory total or detailed sales configuration for the MI450X.

The DeepSeek figure comes from the MI455X

AMD also released a test result for the existing MI455X in Helios. Its performance lab ran DeepSeek-R1 in October and reported 21,333 tokens per second per GPU using an MLPerf Inference Offline configuration. AMD said this was up to 30% above a published result for NVIDIA’s Vera Rubin NVL72.

AMD identified its reference as entry 6.1-0106 in MLPerf Inference 6.1. NVIDIA’s developer table lists 1,183,327 tokens per second offline across 72 Rubin GPUs, or about 16,435 per GPU. Comparing that calculated figure with AMD’s stated result yields a difference of about 29.8%. MLCommons classified the NVIDIA result as a preview submission. AMD did not provide a submission ID or disclose how many GPUs ran its own test.

The Offline scenario gives the system a batch of inputs and measures how quickly it processes the workload. MLPerf’s Server scenario simulates requests arriving over time and applies a latency requirement. The figures AMD published describe throughput for a specific model and scenario. The release gave no first-response latency or power results for the run and included no DeepSeek-R1 test on the new MI450X.

AMD also updated its public material on the MI430X, a member of the same family for scientific computing. Its specifications list 432 GB of HBM4 and native support for double-precision arithmetic, which is used in simulations that need to preserve small numerical differences. AMD’s editor’s note says that the MI430X information was added to the release after it first appeared. In a separate technical post, AMD described a node for scientific workloads that connects four of those accelerators directly with a server CPU.

AMD has begun supplying the earlier Helios configuration. In its November 3 earnings remarks, the company said it had started shipping the MI455X in volume during the third quarter. It named Anthropic, Meta, Microsoft, OpenAI and Oracle as customers that had brought their first Helios racks online. AMD’s product page lists November 12 as the MI450X launch date. AMD’s earnings remarks set out plans to increase production and shipments across Helios. The November product announcement gave no date for the first MI450X deliveries.