Silicon and Models

Silicon fleet

berth ships spec-sheet profiles for a range of accelerators. Each profile carries peak compute, memory bandwidth, memory capacity, and a list price. The two cards below marked MEASURED have been validated against real hardware traces; the rest carry spec-sheet priors and are labeled accordingly.

Six cells are measured today across three memory technologies and two vendors: L40S under vLLM and again under SGLang, H100 PCIe, H100 SXM at bf16 and at fp8, A100 80GB with a mixture-of-experts model, and MI300X. That last one is the first non-NVIDIA cell in the corpus.

A cell is one accelerator running one model under one serving stack at one precision. A card marked MEASURED has at least one cell; it does not mean every model on it has been measured, and the CLI says which on every line.

Key	Accelerator	Peak TFLOPS (bf16)	HBM BW (TB/s)	Memory	Status
l40s	NVIDIA L40S	362	0.864	48 GB	MEASURED
h100-pcie	NVIDIA H100 PCIe	756	2.00	80 GB	MEASURED
h100-sxm	NVIDIA H100 SXM	989	3.35	80 GB	MEASURED
h200-sxm	NVIDIA H200 SXM	989	4.80	141 GB	prior
a100-80g	NVIDIA A100 80GB	312	2.00	80 GB	MEASURED
mi300x	AMD MI300X	1307	5.30	192 GB	MEASURED
tpu-v6e	Google TPU v6e	918	1.64	32 GB	prior
tpu-v5e	Google TPU v5e	197	0.819	16 GB	prior
trn1	AWS Trainium	190	0.820	32 GB	prior
trn2	AWS Trainium2	650	2.90	96 GB	prior
cpu-spr	Intel Sapphire Rapids	8	0.30	512 GB	prior

A profile with prior status predicts from the spec sheet alone. Predictions for those parts are honest forecasts, not measurements, and berth labels them so.

Two things to know about the peak compute column. It carries the vendors' 2:4 structured-sparsity ratings, which are roughly twice the dense bf16 figure, and the microbenchmarked dense numbers on the measured cards came in far lower: 211 TFLOPS on the L40S against a rated 362, and 388 on the H100 PCIe against 756. On measured silicon this is absorbed by the fitted MFU, so it does not move the published results. On prior silicon there is no fit to absorb it, so a compute-bound prediction on a prior row can be optimistic by up to a factor of two. Decode is memory-bound by two to three orders of magnitude and is therefore unaffected; prefill on an unmeasured accelerator is the exposed case, and the pre-registered forecasts for AMD and TPU parts cover decode only for exactly this reason.

The price column is a list-price default and moves constantly. The measured premium in these docs uses the prices actually paid at capture time ($0.99/hr L40S, $3.35/hr H100 PCIe), which differ from the shipped defaults. Pass your own with --prices or price_hr=; a premium computed on stale list prices is a stale premium.

Model registry

Key	Model	Params	Architecture
llama3-8b	Llama-3-8B	8B	dense
llama3-70b	Llama-3-70B	70B	dense
qwen3-235b-moe	Qwen3-235B-A22B	235B total, 22B active	MoE
mixtral	Mixtral-8x7B	47B total, 13B active	MoE

Provenance status

berth never presents a spec-sheet prior as a measurement. Every accelerator is tagged MEASURED or prior, and every number a prediction produces inherits that status. The distinction is a hard rule, not a stylistic choice: a prediction you can trust is one you can tell apart from a measurement.