Loading page…
The MI450X offers less memory headroom without cutting core compute capabilities, AMD says. Its separate MI455X performance claim compares an internal test with Nvidia’s published benchmark result.
Listen to this story
AMD is positioning MI450X as a lower-memory configuration within its Helios platform, not as a cut-down compute option: the company says it retains MI455X’s compute, bandwidth, and connectivity. The accompanying performance case is narrower than a general chip lead: AMD reports a 30% throughput advantage for MI455X in one DeepSeek-R1 configuration against Nvidia’s published result. For buyers, the choice turns on workload memory needs, while the benchmark does not establish an advantage across models or software setups.
AMD measured 21,333 tokens per second per GPU on MI455X in an October 2026 MLPerf Inference Offline test.
The Nvidia comparison uses Vera Rubin NVL72’s published MLPerf Inference v6.1 result, retrieved by AMD on November 11.
AMD says MI450X’s 288 GB of HBM4 suits workloads that need less memory; it discloses no operating-cost savings figure.
AMD is giving Helios customers another AI chip choice while making a new performance claim against Nvidia. In its November 12, 2026 announcement, AMD introduced the MI450X GPU with 288 GB of memory. It also said the MI455X delivered up to 30% more DeepSeek-R1 throughput than Nvidia Vera Rubin’s latest published MLPerf Inference Offline result.
AMD Performance Labs measured 21,333 tokens per second per GPU on DeepSeek-R1 using an AMD-tested MLPerf Inference Offline configuration. The testing took place in October 2026, with the MI455X powering AMD’s Helios rack-scale system. The figure measures throughput in that configuration, rather than establishing a uniform advantage across AI models.
For the Nvidia comparison, AMD used the published MLPerf Inference v6.1 Closed Division result for Vera Rubin NVL72, entry 6.1-0106. AMD retrieved that result on November 11.
AMD cautions that results can vary with the model, system configuration and software version.
The new MI450X addresses a different question from the benchmark: how much memory a customer’s workload needs. AMD says its 288 GB of HBM4 high-bandwidth memory comes with the same core compute and memory bandwidth as MI455X. It also retains the same connectivity capabilities within and between systems, rather than trading those capabilities for a smaller memory configuration.
Both GPUs use the same Helios rack-scale architecture and ROCm software foundation. That makes the announcement a choice within one platform, rather than a separate system and software offering. AMD says the MI455X supplies additional memory headroom for larger models, longer contexts or higher concurrency—the number of requests handled at once.
For workloads that fit inside MI450X’s memory, AMD pitches the chip as another way to meet throughput and response-time targets while optimizing operating expenses. That is a workload-dependent proposition, not a disclosed savings figure.
Loading discussion...
Join the conversation
Explain when extra capacity is worth buying before you need it.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.