Chips & Compute

AMD adds MI430X and MI450X to Instinct MI400 GPU lineup

AMD expands Instinct MI400 with the FP64-capable MI430X, a 288GB MI450X and an MLPerf claim of 21,333 tokens per second per GPU — up to 30% above Nvidia Vera Rubin.

AMD expands Instinct MI400 line with MI430X, MI450X
AMD expands Instinct MI400 line with MI430X, MI450Xblickpixel / pixabay
By James Calloway5 min read

Why it matters

  • AMD launched the MI450X with 288GB of HBM4 and detailed the MI430X with 432GB of HBM4 and 23.3 TB/s bandwidth
  • MI455X on Helios delivered 21,333 tokens per second per GPU on DeepSeek-R1 in MLPerf Inference Offline, which AMD says is up to 30% higher than Nvidia Vera Rubin's latest published result
  • AMD reviewed 47 workload configurations across 11 applications and benchmarks to assess FP64 emulation versus native double precision
  • AMD claims Helios offers up to 20% higher HBM bandwidth than Nvidia Vera Rubin
  • All three GPUs run on the Helios rack-scale architecture and the ROCm software stack

By James Calloway

AMD has expanded its Instinct MI400 Series GPU lineup with three moves at once: fuller details on the scientific-comcompute-focused MI430X, published inference results for the MI455X, and the launch of a new lower-memory MI450X. All three parts sit inside the Helios rack-scale system, AMD's platform for AI and high-performance computing (HPC) workloads including scientific simulation and interactive inference.

The timing matters. AI infrastructure buying has shifted from a single top-end accelerator for every job toward matching silicon to narrower workload profiles, and AMD is now segmenting its lineup accordingly — within one infrastructure platform rather than across several.

What does the MI430X bring to scientific computing?

The MI430X targets research workloads that mix traditional simulation with AI models. It carries native FP64 support — double-precision arithmetic (64-bit floating point), the numeric format many physics and simulation codes still require — alongside 432GB of HBM4 memory (high-bandwidth memory, the on-chip memory type used in AI accelerators) and memory bandwidth of up to 23.3 terabytes per second.

Those specifications address a real constraint: researchers running large simulations and bigger scientific datasets often need double precision in parts of the pipeline that lower-precision AI accelerators cannot natively supply. AMD said it reviewed 47 workload configurations across 11 applications and benchmarks to determine when FP64 emulation may be sufficient and when native double precision remains necessary.

AMD read those findings as evidence of continued demand for hardware that handles both AI and conventional HPC jobs in the same environment. The study itself is not public in the announcement, so the 47-configuration figure stands on AMD's own description of it.

The takeaway: the MI430X is AMD's bet that scientific buyers will not abandon double precision just because AI models moved into their pipelines.

How fast is the MI455X at inference?

Alongside the scientific computing update, AMD published measured inference figures for the MI455X running on Helios: 21,333 tokens per second per GPU on DeepSeek-R1 in an MLPerf Inference Offline configuration (the industry-standard benchmark suite run in batch, rather than interactive, mode).

AMD said the result represents throughput up to 30% higher than Nvidia Vera Rubin's latest published MLPerf Offline result. The company also claimed Helios offers up to 20% higher HBM bandwidth than Vera Rubin, along with higher scale-up bandwidth — the interconnect speed between GPUs inside a rack — which it linked to the reported inference numbers.

The comparison rests on AMD Performance Labs testing plus Nvidia's published MLPerf submissions, not an independent head-to-head. As with most benchmark disclosures, the figures depend on system configuration, software versions and model setup — a caveat AMD itself flags.

The takeaway: a 30% throughput claim is a strong headline, but it is vendor-tested against a competitor's submitted configuration, and buyers will want matching test conditions before treating it as settled.

What is the MI450X for?

The MI450X is the new configuration in the series: the same core compute, memory bandwidth and connectivity profile as the MI455X, but with 288GB of HBM4 memory per GPU instead of the larger part's capacity. AMD positioned it for workloads that fit within that memory footprint, giving customers another lever to balance throughput, response times and operating costs.

Both the MI450X and MI455X run on the same Helios rack-scale architecture and AMD's ROCm software stack (ROCm is AMD's open software platform for GPU computing, the counterpart to Nvidia's CUDA ecosystem). AMD said the MI455X remains the choice for users needing more memory headroom for larger models, longer context windows or higher concurrency — serving more simultaneous requests before performance degrades.

The takeaway: the MI450X is a cost-and-capacity tier, not a performance tier, and its value depends entirely on whether a customer's models actually fit in 288GB.

Why segmentation is the real story?

The broader signal in the update is strategic. AMD is building a segmented product range inside a single platform, reflecting a market where buyers increasingly pair accelerators to specific workload profiles instead of defaulting to the fastest available part.

Competition in AI infrastructure has also moved past raw processor speed. Memory capacity, interconnect bandwidth and rack-level system assembly now weigh heavily in purchasing decisions, and those factors grow more important as model sizes increase and inference shifts from batch processing to interactive deployment patterns that demand fast response times.

The stakes differ by customer segment. Scientific computing buyers still face memory-and-precision trade-offs, since research institutions and industrial users run applications dependent on native double precision even as AI methods enter their simulation pipelines. AI buyers, by contrast, ask whether a system can serve large models at the right latency and cost. The second memory configuration on Helios targets that operational distinction directly.

"AMD Instinct MI400 Series GPUs bring leadership performance and new capabilities to AI and scientific computing. With advances in high-precision compute, new milestones in AI inference performance and an expanded range of configurations, we're giving customers more options to power their next generation of AI and HPC workloads," said Andrew Dieckmann, Corporate Vice President and General Manager, Data Centre GPU Business, AMD.

The quote carries the standard vendor framing — "leadership performance" is a claim, not a measurement. The measurable substance sits in the numbers around it: 432GB and 288GB memory tiers, 23.3 TB/s bandwidth, 21,333 tokens per second per GPU, and a 30% throughput edge claimed over Nvidia's latest published result.

What comes next?

What the announcement does not include is pricing, availability dates for the MI430X and MI450X, or independent verification of the MLPerf comparison. Those gaps will define the next round of scrutiny: if Vera Rubin posts new MLPerf numbers or AMD publishes its FP64-emulation study in full, the competitive picture could shift quickly. For now, AMD has laid out a segmented lineup that mirrors how the market already buys — one workload profile at a time.

Source: itbrief.com.au

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

722 articles

Related articles

  1. OpenAI Commits to 6 Gigawatts of AMD GPUs in Multi-Year Deal
  2. AMD launches 256-core EPYC 9006 CPUs to fight Intel and Nvidia
  3. Supermicro Ships 20-Plus H15 Servers Built on AMD EPYC 9006 Chips
  4. MiTAC Splits AI Inference Across Two Servers With 2 TB Decode Memory
  5. MemryX Targets 10x Compute Jump With MX3+ Edge AI Chip

« Previous articleNext article »