← News archive

Chips · NEWS BRIEF

AMD adds a memory option to MI400 GPUs

Glass-fronted AMD headquarters at 2485 Augustine Drive in Santa Clara in June 2020; corporate context, not new GPU equipment.
Photo: Coolcaesar, CC BY-SA 4.0. AMD headquarters at 2485 Augustine Drive, Santa Clara, photographed June 21, 2020; corporate context, not an MI400 GPU or Helios installation. Resized and converted; photo adaptations under the same license. · CC BY-SA 4.0 licence

Published

Another configuration for Helios

AMD announced an additional MI400-series option on November 12: the Instinct MI450X GPU for its Helios rack-scale platform. The same update discusses scientific computing on MI430X and an AMD-run inference test using MI455X.

These are distinct parts of the portfolio. Reading them as one interchangeable performance claim would obscure the different memory and calculation requirements that AMD says the products address.

Source: AMD — November 12, 2026 ·

Memory capacity becomes a purchasing choice

MI450X has 288 gigabytes of HBM4 memory per GPU. AMD says it retains the core compute, memory bandwidth and connectivity capabilities of MI455X, while MI455X offers more memory headroom for larger models, longer contexts or more simultaneous work.

Capacity and bandwidth answer different questions: how much data fits close to the processor, and how quickly that data can move. A workload that fits in one configuration may not need the capacity of another. That makes the application’s actual requirements central to any comparison.

Scientific calculations have different demands

For MI430X, AMD highlights native double-precision calculation, 432 gigabytes of HBM4 memory and memory bandwidth of up to 23.3 terabytes per second. The company positions these features around scientific simulation and AI for science.

That emphasis helps explain why a GPU family can contain several configurations. Scientific calculation, model serving and large-model workloads put different pressures on the system; a useful purchasing assessment starts with the intended job.

Keep the benchmark attached to its conditions

AMD’s MI455X comparison uses DeepSeek-R1 in an MLPerf Inference Offline configuration. The release attributes the testing to AMD and warns that results can vary with system configuration, software and other factors. It should be read as a particular test, not a promise about every application.

For operators evaluating the new option, the next practical questions are workload fit, complete-system cost and performance under their own operating conditions. The announcement supplies product and test details; it does not establish a universal winner across all those choices.

Understand the infrastructure

What a GPU does inside an AI system

Sources linked below.