Skip to content
Friendly disclaimer: flozi00 TechHub is a solo side-project next to a full-time job — personal learning notes, no official statements. Verify critical steps yourself.
flozi00 Logoflozi00 TechHub
ENDE

AMD's MI455X 34× Throughput Claim: What the Disclosure Establishes

AMD reports 34× MI355X token throughput for MI455X on DeepSeek V4 Flash FP4. The published hardware ratios are 2.91× bandwidth and 1.5× capacity; neither explains the benchmark without the missing test configuration.

3 min readflozi00
amdheliosmi455xinferencebenchmarkhardware
Skip to the full article
Watch · explore · understand

A throughput ratio is not a hardware decomposition

Keep AMD’s reported workload result separate from capacity and bandwidth ratios, then identify the evidence needed to explain it.

AI-assisted learning visualization · schematic article explanation

Step 1 / 4Identify the stated comparison

The claim concerns AMD’s selected DeepSeek V4 Flash FP4 test against MI355X. It cannot be transferred to another GPU baseline or an arbitrary workload.

ComparisonCurrent stepSelect a block to explain it
Reported 34× throughput
AMD compares MI455X with MI355X in its selected test.
Disclosed test scope
Model family, FP4 serving and interactivity points.
  1. Reported 34× throughput Disclosed test scope
    read the footnote
What to remember

A benchmark headline and a specification table constrain different questions; missing controls prevent causal attribution.

The article distinguishes a newer product-page bandwidth figure from an earlier engineering projection. Neither peak figure establishes delivered bandwidth or explains the reported throughput ratio.

Terms and the complete walkthrough
Baseline
The exact system or configuration in the denominator of a ratio.
Attribution
Evidence that isolates how a particular change affects the result.
Interactivity
The response-speed target under which a serving result is reported.
  1. Identify the stated comparison

    The claim concerns AMD’s selected DeepSeek V4 Flash FP4 test against MI355X. It cannot be transferred to another GPU baseline or an arbitrary workload.

  2. Keep hardware dimensions separate

    Capacity tells you how much can fit; bandwidth tells you how fast bytes can move. Multiplying their ratios does not identify a combined contribution to token throughput.

  3. Recover the missing experiment

    The disclosed scope does not provide the complete checkpoint, request lengths, concurrency, software and topology. Without those controls, dividing the headline by a hardware ratio does not isolate a software effect.

  4. State only what the evidence establishes

    Compare matched model representations and latency targets, then measure output throughput, memory and power. Until then, the article supports a scoped vendor claim and separate specification ratios.

All connections

  • Reported 34× throughput → Disclosed test scope: read the footnote
  • Reported 34× throughput → Independent peak specifications: separate spec ratios
  • Disclosed test scope → Missing full configurations: unresolved variables
  • Missing full configurations → Matched workload comparison: obtain and control
  • Independent peak specifications → Matched workload comparison: plausibility constraints
  • Matched workload comparison → Bounded conclusion: measured attribution
Set your own pace. Pause at any time or select a step directly.

AMD says its MI455X delivers 34× the token throughput of MI355X. The claim comes from AMD Performance Labs measurements and calculations in July 2026 using DeepSeek V4 Flash with FP4 serving at low, medium, and high interactivity points. It is a vendor-reported result for that workload, not an independently reproduced result or a general GPU speed ratio1.

What can be checked from specifications

Published specificationMI455XMI355XRatio
HBM capacity per GPU432 GB HBM42288 GB HBM3E31.50×
Peak HBM bandwidth per GPU23.3 TB/s28 TB/s32.91×
Dense FP16 matrix peak5 PFLOPS22.5166 PFLOPS31.99×

An earlier AMD Helios brochure uses 19.6 TB/s per MI455X rather than 23.3 TB/s; using that figure makes the peak-bandwidth ratio 2.45×. The brochure labels its specifications engineering projections subject to change. The newer product-page figure should be checked against the delivered system before either value is used for capacity planning2.

These are independent hardware properties. Multiplying the 2.91× bandwidth ratio by the 1.50× capacity ratio does not produce a 4.37× “silicon contribution” to throughput. For a simple memory-bound, single-sequence decode with the same model representation on both GPUs, the ideal bandwidth-only ratio would be 2.91×. More capacity may allow a larger model, longer context, or more concurrent requests; its effect on throughput depends on the workload and scheduler. Compute and communication can become the limiting factors.

What the 34× disclosure does not identify

AMD's footnote names the model family, FP4 serving, comparison GPU, and three interactivity levels. It does not publish the full benchmark harness, checkpoint identity, input and output token lengths, request mix, batch or concurrency settings, serving-engine versions, complete hardware topology, absolute tokens per second, or uncertainty across runs1. The footnote does not let a reader isolate the effect of memory bandwidth, model placement, quantization, speculative decoding, or software tuning.

Dividing 34 by a specification ratio produces a numerical quotient, not an attribution. In particular, the source does not show that speculative decoding or any specific precision mismatch accounts for a stated fraction of the result. Neither the 34× claim nor the hardware ratios can be transferred to an NVIDIA GB300 comparison: AMD's stated baseline is its own MI355X1.

How to use the claim

For a deployment decision, ask for the full MI455X and MI355X test configurations and run the intended checkpoint at the same precision, context lengths, concurrency, and latency targets. Report both output tokens per second and latency (including time to first token), plus memory headroom and power. Until those measurements exist, the defensible conclusion is narrow: AMD reports 34× on its selected DeepSeek V4 Flash FP4 test, while its published peak memory-bandwidth and capacity ratios are 2.91× and 1.50× respectively123.

Footnotes

  1. AMD, AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era, July 23, 2026. Main text states the 34× throughput claim; footnote 7 (MI400-020) specifies AMD Performance Labs, July 2026, DeepSeek V4 Flash with FP4 serving, MI355X baseline, and low/medium/high interactivity. https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era ↩ ↩2 ↩3 ↩4

  2. AMD, Instinct MI455X product specifications: 432 GB HBM4, peak 23.3 TB/s, dense FP16 matrix 5 PFLOPS. AMD's Helios brochure states 19.6 TB/s per GPU. https://www.amd.com/en/products/accelerators/instinct/mi400/mi455x.html ; https://www.amd.com/content/dam/amd/en/documents/solutions/ai/amd-helios-bro.pdf ↩ ↩2 ↩3 ↩4 ↩5

  3. AMD, Instinct MI355X product specifications: 288 GB HBM3E, peak 8 TB/s, dense FP16 matrix 2.5166 PFLOPS. https://www.amd.com/en/products/accelerators/instinct/mi350/mi355x.html ↩ ↩2 ↩3 ↩4