AMD expands Instinct MI400 Series with new GPU options

Key Takeaways

• AMD expands the Instinct MI400 series with three distinct GPUs: MI430X, MI450X, and MI455X.
• MI430X targets double‑precision scientific workloads, offering 432 GB HBM4 and 23.3 TB/s memory bandwidth.
• MI455X delivers record inference speed—21,333 tokens / s on DeepSeek‑R1—up to 30 % faster than Nvidia’s Vera Rubin in MLPerf Offline tests.
• MI450X adds a 288 GB HBM4 memory option while keeping the same compute core and bandwidth as the MI455X, giving Helios customers a lower‑cost, high‑throughput alternative.
• All three GPUs run on AMD’s open ROCm stack and share the same rack‑scale Helios architecture, simplifying integration and scaling.

Overview of the Updated Instinct MI400 Portfolio

AMD has announced detailed specifications for the latest members of its Instinct MI400 series, a family built to satisfy the divergent needs of scientific simulation, frontier AI, and interactive inference. Each model retains native FP64 capability for high‑precision tasks, while offering scalable memory capacities and bandwidth to accommodate ever‑larger AI models.

MI430X – Engineered for Double‑Precision Science

The MI430X is positioned as the flagship for compute‑intensive research workloads that require true double‑precision arithmetic. Key attributes include:

• FP64 native compute – eliminates the performance penalty of software emulation.
• 432 GB of HBM4 memory, the largest capacity in the series.
• 23.3 TB/s memory bandwidth, enabling rapid access to massive datasets.

AMD’s internal study examined 47 configuration permutations across 11 scientific applications, confirming that native FP64 delivers measurable speedups for core physics, chemistry, and climate models where emulation falls short.

MI455X – Record‑Setting Inference Performance

In the Helios rack‑scale system, the MI455X achieved 21,333 tokens per second per GPU on the DeepSeek‑R1 model using the MLPerf Inference Offline benchmark. This figure translates to:

• ~30 % higher throughput than Nvidia’s Vera Rubin (latest published MLPerf result).
• ~20 % greater HBM bandwidth and superior scale‑up interconnect bandwidth versus Vera Rubin, according to AMD’s measurements.

These advantages stem from a combination of wider memory channels, optimized routing, and the latest AMD CDNA 3 architecture.

MI450X – Flexible Memory‑Capacity Option

The MI450X mirrors the MI455X’s compute engine and bandwidth but offers a more modest 288 GB HBM4 memory pool. This configuration is aimed at:

• Reducing capital expense for workloads that fit within the smaller memory envelope.
• Maintaining identical scale‑out connectivity, allowing seamless substitution in existing Helios deployments.

Built on AMD’s open compute foundation, the MI450X simplifies integration for OEMs and system integrators, minimizing firmware and driver changes.

Specification Comparison

Implications for Helios Deployments

The Helios rack‑scale platform now supports two inference‑focused GPUs (MI450X & MI455X) alongside the science‑oriented MI430X. Because all three share the same PCIe/Infinity Fabric topology and ROCm software stack, data centers can mix and match GPUs within a single chassis without redesigning cooling, power, or networking. This flexibility helps organizations balance performance, memory, and total cost of ownership (TCO) as AI model sizes evolve.

Bottom Line

AMD’s expanded Instinct MI400 lineup gives customers a clear, tiered approach to modern HPC and AI challenges. The MI430X delivers unparalleled double‑precision capacity for next‑generation scientific codes, while the MI455X sets a new benchmark for inference speed. The MI450X provides a cost‑effective middle ground, preserving the same compute and bandwidth characteristics as the top‑end model. Together, these GPUs reinforce AMD’s strategy of open, scalable architecture, enabling Helios users to tailor hardware to workload demands without sacrificing ecosystem compatibility.