Lambda Claims Over 60 Percent MFU in Llama 3.1 Training on Blackwell
Data center provider Lambda presents numbers showing model training significantly more efficient than what the company describes as the industry norm — but everything comes from its own marketing materials, and the white paper's…
Lambda Claims Over 60 Percent MFU in Llama 3.1 Training on Blackwell
Data center provider Lambda presents numbers showing model training significantly more efficient than what the company describes as the industry norm — but everything comes from its own marketing materials, and the white paper's methodology is not publicly available for review.
Lambda has published a white paper with benchmarks that the company claims show model training significantly more efficient than the industry norm. Heavy caveats apply: the figures come from Lambda's own marketing page for the white paper, which sits behind a download form, and no independent verification exists in the available source material (1938d4a1).
What Lambda Actually Claims
The core of the claim concerns MFU — "model FLOPs utilization" — a measure of how much of a GPU accelerator's theoretical compute capacity is actually utilized during training. According to Lambda, most large-scale training runs operate at 35–45 percent MFU, well below the hardware's potential. The company's own benchmarks of Llama 3.1 models from 8 billion to 405 billion parameters on NVIDIA's Blackwell GPUs reportedly achieved MFU above 60 percent (1938d4a1). Measured against the industry baseline Lambda itself describes, that would amount to a relative improvement of at least roughly a third — but this is AIMag's own calculation, not a figure Lambda reports.
The most concrete single number is an MFU improvement of 2.11x for Llama 70B on 16x NVIDIA HGX B200 (1938d4a1). Lambda further claims the results were achieved without architecture changes, and that every configuration is fully documented and reproducible (1938d4a1).
Why MFU Matters
MFU is a meaningful efficiency measure because GPU time is among the largest cost items in model training. A training run that utilizes 60 percent of the hardware's compute capacity requires significantly fewer GPU hours — and thus lower cost and shorter wait times — than one running at 40 percent, given the same model and dataset. This is why MFU is often used as a basis for comparison between training setups and between GPU cluster operators.
The difference Lambda suggests is significant if it holds: an improvement from 35–45 to over 60 percent MFU means in practice that the same hardware can deliver more training per unit of time. Such numbers also serve as a sales argument for companies offering GPU capacity — which is part of why they should be read with caution.
What Remains to Be Shown
The source material does not provide access to the white paper's methodology — neither the benchmark setup, the measurement method, nor the comparison basis for the 2.11x figure. It is also not known when the white paper was published. No third party has replicated the figures in the source material.
What would make the claims more credible is concrete: making the methodology and configurations available for external review, independent operators repeating the benchmarks with the same results, and other actors publishing comparable MFU figures for Llama 3.1 on Blackwell hardware. Lambda's claim of "fully documented" configurations points in the right direction, but documentation behind a download form is not the same as open verification.
Until then, the story stands as it is: an interesting, but unconfirmed, promise of significantly more efficient training on NVIDIA's newest hardware.