DiggTop ⌕

September 2026: What 160 Runs Changed

September added 160 rows to the board, 138 of them verified — the same number of verified rows August produced from 235 submissions. The month also moved hardware prices in a direction few builders expected: up, on both new and used cards.

As of 2026-09-28. All benchmark figures are imported rows that upstream sources verified; we did not run these tests. Prices are our own dated snapshots and carry their source links at the end.

The month in numbers

August 2026 September 2026
Rows added 235 160
Verified 138 138
Distinct models 96 49
Distinct GPUs 45 34
Verified rate 59% 86%
New price snapshots — 3

August was a bulk-import month: wide, shallow, and only three in five rows survived methodology review. September was narrower and deeper — fewer models, but far more of what arrived was usable. The X collection pipeline contributed 84 of the month's 160 rows, more than the localmaxxing import for the first time.

What got measured

Two models were measured properly for the first time in September:

Model First measured Verified rows Best tok/s
DeepSeek-V4.1-Flash Sep 10 34 231.6
Qwen 3.8 Flash Sep 21 8 86.8

DeepSeek-V4.1-Flash is the month's real addition: 34 verified rows across DGX Spark configurations up to four boxes, which is what made our 552B-on-one-box analysis possible. Qwen 3.8 Flash arrived later and is still thin.

The single fastest row recorded all month was an RTX 5090 running LFM2-350M at 2,231.4 tok/s (NVFP4, SGLang, 704-token context). It is a real number and a useless buying signal — a 350M model is a benchmark of the card, not of anything you would deploy. The same caveat applies to the RTX PRO 6000 Blackwell sweep: 1,271.4 tok/s on MobileLLM-125M, 1,060 on Qwen2.5-0.5B, 980.8 on MobileLLM-350M. Small models measure memory bandwidth; the interesting rows this month were the 27B-and-up ones.

The price story is the story

Card Sep 4 Sep 10 Sep 28
RTX 5090 (new) $5,000 $4,899.99 $9,000
RTX 3090 (used) $1,050 — $1,300
RTX 3060 (new) $477 $454.62 $495
RTX 5060 Ti 16GB (new) $800 $779.99 $800

A new RTX 5090 roughly doubled in three weeks — US retail stock dried up and third-party listings ran from $7,000 into five figures, with a street average around $9,000 in the 2026-09-28 snapshot. Used RTX 3090s went from $1,050 to $1,300, a 24% move on a five-year-old card, on reports of memory-supply pressure pushing buyers into the used market.

Two things follow from that, and they pull in opposite directions:

  1. The value argument for old silicon got stronger. A used 3090 at $1,300 against a 5090 at $9,000 is a 6.9× price gap. Our 27B analysis found the 3090 reaching 164.6 tok/s against the 5090's 213.0 on the best non-exotic row — a 1.3× throughput gap for a 6.9× price gap.
  2. The ceiling argument for new cards got weaker. Paying $9,000 for 32GB is hard to justify when the same money buys several used 24GB cards or a $949.99 Arc Pro B70, which our board has at 255.9 tok/s on some models in the right format.

We added two cards to the price tracking this month for exactly this reason: the CMP 170HX (a repurposed mining card, around $1,050 used after its 64–80GB memory unlock) and the Intel Arc Pro B70 at $949.99. Both are now compared on the board rather than dismissed on paper.

How the data got here

The board is not hand-curated. This month the X collection pipeline ran 200 collection passes, extracted 1,158 candidate runs from 1,868 discovered posts, and ingested 40 posts' worth of runs through the methodology gate — the rule that a row needs engine, version, quantisation, context length and hardware state before it may appear on the board. The daily digest re-rendered 31 times across the 16 dates from Sep 12 to Sep 27, keeping partial or unmethodical posts off the board but on the record.

That gate is why the verified rate went up, not down: 86% of September's rows carried the minimum needed to mean something. Rows without it stay in the digest.

What we are watching in October

The price reports carry every snapshot with its sources, and the board has all 452 rows with per-row engine, quant and context. If you ran something we have not measured, submit it — 84 of this month's rows came from people posting their numbers in public.

FAQ

How many benchmark rows did DiggTop verify in September 2026?

138 verified rows out of 160 added — an 86% verified rate, up from 59% in August. Verified means the row carries hardware, quantisation, engine version and context length from its source.

Why did the RTX 5090 price double in September 2026?

US retail stock dried up during the month: our snapshots went from $4,899.99 on 2026-09-10 to roughly $9,000 on 2026-09-28, with third-party listings reported between $7,000 and $9,500. It looks like a supply shock rather than a demand-only move, so October's snapshot will show whether it holds.

Is a used RTX 3090 still worth it at $1,300?

Against a $9,000 RTX 5090 the ratio still favours the 3090 for local LLM work: our 27B comparison shows 164.6 tok/s on a 3090 versus 213.0 on the best non-exotic 5090 row. You give up 32GB of VRAM, NVFP4 support and a warranty, and you get roughly 7× the card count per dollar.

What counts as a verified benchmark row?

A row needs engine, engine version, quantisation format, context length and hardware state — the minimum for a measurement to be reproducible. Rows missing any of those stay out of the board and appear only in the daily digest. The rules are published in our methodology.

Sources

↑