← Frontier Inference Margins · all research reports

Research artifact — each page states its own provenance class in its header: verbatim originals carry SHA-256 stamps; adopted-findings summaries and reconstructions say so explicitly. Conclusions are synthesized (and where needed corrected) in the main report; each page carries its own run date.

Methods note: leave-one-anchor-out validation of the effective-MFU abstraction

Run 2026-07-10, as prescribed by the external review (the empiricist's "load-bearing experiment"). Code: tests/loao.mjs in the project tree. Predictions were fixed before comparison.

Question

The calculator compresses each accelerator's serving behavior into one scalar — an "effective decode MFU" against dense 8-bit FLOPS. Is that a predictive abstraction (one coefficient transfers across hardware), or an interpolating one (each platform needs its own fitted coefficient)?

Design

Fit a single global decode MFU using only the DeepSeek H800 production anchor (1,850 tok/s/GPU on a 37B-active model, 1.98 PF dense FP8 ⇒ 6.91%). Predict every other platform's published decode measurement from its spec sheet and that one coefficient. Pass bar (set in advance by the review): central error ≤ ~25%.

Result

Held-out platform Predicted Measured Error
H20 (Ant/SGLang production, <50 ms tier) 277 675 −59%
GB200 (vLLM published) 4,672 10,100 −54%
Ascend 910C INT8 (CloudMatrix-Infer, optimized) 1,405 1,943 −28%
Ascend 910C INT8 (neutral read: DeepSeek "60% of H100") 1,405 1,303 +8%

Mean |error| 37%, worst 59% — the abstraction FAILS the transfer test.

Interpretation and consequences (adopted in methodology v2)

  1. A scalar compute-MFU does not transfer across platforms, because decode is memory/interconnect-bound: the H20's tiny FLOPS denominator makes its fitted "MFU" (17%) 2.5× the H800's (7%) for the same physical workload. The coefficient absorbs bandwidth, batch regime, serving stack, and latency target — it is not a hardware constant.
  2. The calculator's per-platform MFU values are therefore anchor fits — each reproduces its own published measurement, and the page no longer describes this as "calibration" in the predictive sense. Reproducing the datum used to select a coefficient is an identity, not a validation.
  3. Platforms without a published anchor carry materially lower confidence: TPU v7 Ironwood, Trainium 2/3, and the Rubin projection have no public serving measurement for a frontier MoE and their MFU values are analyst estimates. This is now stated in the calculator.
  4. The model's honest domain of validity: short-to-moderate context, throughput-oriented serving, interpolation near the anchored operating points. The published anchors are all ≤~50B-active models; the flagship scenarios assume 120–300B active — that size axis is the model's largest unanchored extrapolation. Context length, TTFT/TPOT targets, KV-cache lifecycle, and speculative-decoding acceptance are not modeled; the interactivity multipliers (1.0/0.70/0.35) are a coarse stand-in for a latency curve that published data (CloudMatrix: 1,943 → 538 tok/s from 50 ms → 15 ms TPOT) shows is steep.

What would upgrade the model

A roofline formulation (compute + HBM + interconnect terms with per-platform physical constants) fit jointly on all anchors with one shared efficiency residual — then re-run this experiment. Until that passes, cross-platform margin comparisons inherit anchor-fit uncertainty, and the per-provider ranges in §10 should be read accordingly.

Follow-up (2026-07-10, same day): the roofline attempt — a documented negative result

We commissioned the upgrade path named above: a decode roofline with compute, HBM and fabric terms, 1/batch weight amortization, explicit MTP handling, and one shared roof-utilization residual (no per-platform efficiency), fit on the H800 production anchor alone (η = 31.0%). Full derivation: the roofline consultation; runnable diagnostic: /tests/roofline-diagnostic.mjs, deployed with the site.

Result: dramatically better than scalar MFU at throughput-oriented operating points — H20 −11/−16%, GB200 −2% (under an optimistic all-FP4 weight-traffic bound), Ascend@50ms −10%, and it reproduces the CloudMatrix paper's measured ~44% MTP latency increase — but it fails the whole-platform gate: Ascend's documented 15ms operating point misses by −39.8% (≈6–8pp of margin at typical operating points), and prefill transfers worse (worst −70.8%). The favorable numbers are also unstable to undisclosed metadata (H800 batch and MTP status alone swing the worst error between 31% and 57%).

More fundamentally, the preregistered experiment could not be executed as designed on public data (the fit is exactly-determined rather than validated): one fitted parameter against one training platform group violates the identifiability precondition (p < n_train), and no genuinely untouched platform exists — every candidate already informed the model's design. Decision: the anchor-fit model stays; the roofline ships as an experimental diagnostic only. A valid future test needs ≥2 independent training platform groups, a genuinely new platform held out after model freeze, frozen checkpoint/batch/MTP/precision metadata, and a max-error gate over every held-out operating point.