Independent testing will determine whether the v100-skinny kernels make NVFP4 weights and speculative decoding a practically high-throughput inference path for Qwen-class models on inexpensive V100 GPUs.

state: expiredheat: lowuncertainty: highknownscott: lownvfp4 speculative-decoding inference-economics

What is this?

A reported v1.0 implementation called “v100-skinny” claims that four Tesla V100 GPUs can serve a Qwen3.6-27B NVFP4 model at up to 366 tokens per second, apparently using speculative decoding. The supplied sources establish that the checkpoint includes an MTP module for self-speculation and that NVFP4 is a compact 4-bit format, but they primarily discuss modern NVIDIA hardware and do not independently verify the V100 benchmark. The project’s authorship, test methodology, quality impact, workload conditions, and practical cost comparison are not established here, so the central throughput claim remains awaiting reproducible, apples-to-apples testing.

Why it matters to Scott

This is another unverified throughput-claim benchmark case in a pattern the radar already tracks extensively — radar:qwen36-quant-specdecode-scaling covers the same Qwen3.6-27B quant/speculative-decoding scaling story, and radar:concept.speculative-decoding, radar:concept.quantization, radar:concept.local-inference, radar:concept.inference-efficiency already aggregate dozens of similar 'independent testing will determine whether X GPU config delivers claimed inference throughput' episodes. It illustrates Scott's local-inference/hardware-aware-inference interests (dev:concept.hardware-aware-local-inference, dev:project.gpt4all) but adds no new claim, technology, or challenge beyond what the radar's existing cluster already holds.

dev:concept.hardware-aware-local-inferenceradar:qwen36-quant-specdecode-scalingradar:concept.speculative-decodingradar:concept.quantizationradar:concept.local-inferenceradar:concept.inference-efficiency

queries asked of Scott's wikis

commodity GPU inference economics
extending legacy GPU life for local inference
quantization quality versus throughput tradeoffs
speculative decoding and self-drafting models
open-model hardware sovereignty
reproducible inference benchmark methodology

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	366 t/s Qwen3.6 27B NVFP4 on v100s LocalLLaMA	Simple_Library_2700	124	114
🟧 echo.github ⭐	The v1.0 commit publishes the benchmark implementation and results. Its README says: “Four Tesla V100 cards serve Qwen3.6-27B at up to 366 t	dnv2003	—	—

Interpretation history

2026-08-17T03:29:07Z
The promised independent testing has not appeared after repeated review, and the latest change is engagement-only. The episode has faded without advancing beyond its original first-party benchmark claim and can be reopened if substantive reproduction arrives.
2026-08-15T03:22:26Z
The latest activity is still comment churn rather than independent reproduction, prefill data, or methodological clarification. The testable first-party claim remains open but cold and should not be revisited on engagement alone.
2026-08-13T02:33:03Z
The refreshed comments add no independent reproduction, prefill result, or methodological clarification. This remains repetitive amplification of a first-party benchmark and should stay cold until substantive testing arrives.
2026-08-13T00:24:19Z
The refreshed discussion still provides no independent reproduction, prefill benchmark, or methodological clarification. Comment churn is repetitive amplification, so the case remains a cold first-party claim pending substantive external testing.
2026-08-12T21:36:50Z
The refreshed discussion adds no independent reproduction, prefill benchmark, or methodological clarification. Repeated comment churn is no longer changing the case; it remains a cold first-party throughput claim awaiting substantive external testing.
2026-08-12T19:27:30Z
The latest comment refresh again adds no independent reproduction, prefill measurement, or methodological clarification. The case remains a cold first-party benchmark claim and should await substantive external testing rather than further discussion updates.
2026-08-12T15:41:54Z
The refreshed comments still provide no independent reproduction, prefill measurement, or methodological clarification. This remains a cold first-party benchmark claim; further discussion-only updates should not trigger review.
2026-08-12T11:41:35Z
The refreshed discussion adds no independent reproduction, prefill measurement, or methodological clarification. Repeated amplification is no longer changing the case’s meaning; it remains a first-party benchmark claim awaiting substantive external testing.
2026-08-12T09:26:50Z
The latest comment refresh still adds no independent reproduction, prefill measurement, or methodological clarification. Repetitive discussion no longer changes the case’s meaning; it remains a first-party benchmark claim awaiting external testing.
2026-08-12T08:39:29Z
Two additional comments add no independent reproduction, prefill measurement, or methodological clarification. The discussion remains repetitive amplification of the first-party benchmark, so the case should stay cold pending substantive test results.
2026-08-12T07:31:44Z
The comment refresh adds no independent reproduction, prefill benchmark, or methodological clarification. Repeated amplification has stopped changing the case’s meaning, so it should remain cold until substantive test results arrive.
2026-08-12T06:34:40Z
The refreshed comments add no independent reproduction, prefill measurement, or methodological clarification; the case remains a first-party benchmark claim despite sustained interest. Repeated discussion updates no longer merit frequent review absent substantive test results.
2026-08-12T05:34:24Z
The refreshed discussion remains enthusiasm and repeated questions about prefill, without the promised independent test or new methodological evidence. The case still represents a reproducible first-party throughput claim awaiting external validation.
2026-08-12T04:30:58Z
The refreshed discussion still supplies no independent benchmark, prefill result, or methodological clarification. It remains a reproducible but first-party throughput claim, with repeated community interest adding no validation.
2026-08-12T03:25:25Z
The velocity spike is amplification of the original benchmark claim, not validation; no independent result, prefill measurement, or methodological clarification changes the case’s meaning.
2026-08-12T02:31:46Z
The refreshed comments remain repetitive enthusiasm and prefill questions; no promised independent benchmark or methodological clarification has appeared. The case still means only a reproducible first-party throughput claim awaiting external validation.
2026-08-11T23:25:27Z
Refreshed discussion remains hardware enthusiasm and questions about omitted prefill performance; the previously mentioned independent test has not produced results. The case is still a reproducible first-party claim awaiting external benchmarks.
2026-08-11T22:24:31Z
Discussion is now probing missing prefill data and includes one commenter’s intent to test, but no independent results have arrived. The case remains an unverified first-party throughput claim rather than emerging corroboration.
2026-08-11T21:52:09Z
Engagement on the original Reddit post increased (19 upvotes, 20 comments), but no new evidence or independent verification; the claim remains unconfirmed.
2026-08-11T21:30:43Z
grounded: known/low — This is another unverified throughput-claim benchmark case in a pattern the radar already tracks extensively — radar:qwen36-quant-specdecode-scaling covers the
2026-08-11T21:27:36Z
origin walked (codex/luna, conf 0.98): anchor reddit.post.1vlt0lj -> echo.github.051df353d3 by dnv2003
2026-08-11T21:25:28Z
case created — Single first-party reddit post with concrete single-stream benchmark numbers for a novel low-cost inference path; no corroboration yet.

Decision trace

08-17 13:29expireThe promised independent testing has not appeared after repeated review, and the latest change is engagement-only. The episode has faded without advancing beyond its original first-party benchmark cla
08-17 13:29alert_silentEleven additional upvotes and one comment add no independent benchmark, prefill result, or methodology disclosure, so there is no consequential delta that merits Scott's attention.
08-17 13:29alert_routeEleven additional upvotes and one comment add no independent benchmark, prefill result, or methodology disclosure, so there is no consequential delta that merits Scott's attention.
08-15 13:22repriceThe latest activity is still comment churn rather than independent reproduction, prefill data, or methodological clarification. The testable first-party claim remains open but cold and should not be r
08-15 13:22alert_silentNo consequential evidence changed the case, so deferring creates no attention regret; revisit only for an independent benchmark or material methodology disclosure.
08-15 13:22alert_routeNo consequential evidence changed the case, so deferring creates no attention regret; revisit only for an independent benchmark or material methodology disclosure.
08-13 12:33repriceThe refreshed comments add no independent reproduction, prefill result, or methodological clarification. This remains repetitive amplification of a first-party benchmark and should stay cold until sub
08-13 12:33alert_silentNo consequential new evidence arrived, so waiting for a later briefing creates no attention regret; revisit only on an independent benchmark or material methodology disclosure.
08-13 12:33alert_routeNo consequential new evidence arrived, so waiting for a later briefing creates no attention regret; revisit only on an independent benchmark or material methodology disclosure.
08-13 12:21sensor_dirtycomment_update
08-13 10:24repriceThe refreshed discussion still provides no independent reproduction, prefill benchmark, or methodological clarification. Comment churn is repetitive amplification, so the case remains a cold first-par
08-13 10:24alert_silentNo consequential evidence arrived, and waiting creates no attention regret; revisit only when an independent benchmark, prefill result, or material methodology disclosure appears.
08-13 10:24alert_routeNo consequential evidence arrived, and waiting creates no attention regret; revisit only when an independent benchmark, prefill result, or material methodology disclosure appears.
08-13 10:21sensor_dirtycomment_update
08-13 07:36repriceThe refreshed discussion adds no independent reproduction, prefill benchmark, or methodological clarification. Repeated comment churn is no longer changing the case; it remains a cold first-party thro
08-13 07:36alert_silentNo consequential evidence arrived, so the next briefing will not be late; revisit only if an independent benchmark, prefill result, or material methodology disclosure appears.
08-13 07:36alert_routeNo consequential evidence arrived, so the next briefing will not be late; revisit only if an independent benchmark, prefill result, or material methodology disclosure appears.
08-13 07:21sensor_dirtycomment_update
08-13 05:27repriceThe latest comment refresh again adds no independent reproduction, prefill measurement, or methodological clarification. The case remains a cold first-party benchmark claim and should await substantiv
08-13 05:27alert_silentNo consequential evidence arrived, so waiting for the next briefing creates no attention regret; revisit only on an independent benchmark, prefill result, or material methodology disclosure.
08-13 05:27alert_routeNo consequential evidence arrived, so waiting for the next briefing creates no attention regret; revisit only on an independent benchmark, prefill result, or material methodology disclosure.
08-13 05:21sensor_dirtycomment_update
08-13 01:41repriceThe refreshed comments still provide no independent reproduction, prefill measurement, or methodological clarification. This remains a cold first-party benchmark claim; further discussion-only updates
08-13 01:41alert_silentOnly the Reddit discussion changed, with no consequential evidence that would make the next briefing late; wait for an independent benchmark, prefill result, or substantive methodology disclosure.
08-13 01:41alert_routeOnly the Reddit discussion changed, with no consequential evidence that would make the next briefing late; wait for an independent benchmark, prefill result, or substantive methodology disclosure.
08-13 00:22sensor_dirtycomment_update
08-12 21:41repriceThe refreshed discussion adds no independent reproduction, prefill measurement, or methodological clarification. Repeated amplification is no longer changing the case’s meaning; it remains a first-par
08-12 21:41alert_silentNo consequential new evidence arrived, so the next briefing will not be late; revisit only when an independent benchmark, prefill result, or material methodology disclosure appears.
08-12 21:41alert_routeNo consequential new evidence arrived, so the next briefing will not be late; revisit only when an independent benchmark, prefill result, or material methodology disclosure appears.
08-12 21:21sensor_dirtycomment_update
08-12 20:21sensor_dirtyengagement_update
08-12 19:26repriceThe latest comment refresh still adds no independent reproduction, prefill measurement, or methodological clarification. Repetitive discussion no longer changes the case’s meaning; it remains a first-
08-12 19:26alert_silentNo consequential evidence arrived, so another discussion-only update can wait for the next briefing; revisit sooner only if an independent benchmark, prefill result, or material methodology disclosure
08-12 19:26alert_routeNo consequential evidence arrived, so another discussion-only update can wait for the next briefing; revisit sooner only if an independent benchmark, prefill result, or material methodology disclosure
08-12 19:21sensor_dirtycomment_update
08-12 18:39repriceTwo additional comments add no independent reproduction, prefill measurement, or methodological clarification. The discussion remains repetitive amplification of the first-party benchmark, so the case
08-12 18:39alert_silentNo consequential delta occurred; the next briefing would not be late unless an independent benchmark, prefill result, or material methodology update appears.
08-12 18:39alert_routeNo consequential delta occurred; the next briefing would not be late unless an independent benchmark, prefill result, or material methodology update appears.
08-12 18:21sensor_dirtycomment_update
08-12 17:31repriceThe comment refresh adds no independent reproduction, prefill benchmark, or methodological clarification. Repeated amplification has stopped changing the case’s meaning, so it should remain cold until