Gilmore ConceptsKanban
10/10. Passed on the first attempt, 7.0 min
The Lab / earlier run, Oct 28 2026
qwen38-27b, 2×RTX 3060: 4 attempts, 69 min
This is an earlier run. See the latest results
Every attempt, in order
Bar length is time. Numbers are minutes.
qwen38-27b, 2×RTX 3060, Oct 28
Every attempt as a table
| Setup | App | Attempt | Result | Minutes | Failed checks |
|---|---|---|---|---|---|
| qwen38-27b, 2×RTX 3060 | Kanban | 1 | Passed | 7.0 | none |
| qwen38-27b, 2×RTX 3060 | Sand sim | 1 | Passed | 9.1 | none |
| qwen38-27b, 2×RTX 3060 | Dungeon | 1 | Failed | 35.5 | moves, walls_block, exit_reachable, enemies, damage, game_over |
| qwen38-27b, 2×RTX 3060 | Dungeon | 2 | Passed | 17.4 | none |
- Passed
- Failed, tried again
Watch it
These are clips of the actual apps the model wrote.
qwen38-27b, 2×RTX 3060, Oct 28
Sand sim
11/11. Passed on the first attempt, 9.1 min
Dungeon
11/11. Passed on the second attempt, 52.9 min
The numbers
| Setup | qwen38-27b, 2×RTX 30602×RTX 3060 12 GB, Ryzen 7 5800X, 48 GB DDR4 Full specs |
|---|---|
| Score | 32/32 |
| Attempts | 4 |
| Total time | 69 min |
| Generation speed | 16.2 tok/s |
| Prompt speed | 268 tok/s |
| Tokens generated | 64,921 |
| Motherboard | B450M DS3H WIFI |
Speeds are token-weighted averages from the server's timing lines in each run.
What this doesn't prove
Each setup has one run, and neither set a fixed seed. Randomness is the likely reason for the extra attempts, but that's not proven.
Flash Next ran through Strata, which doesn't log speeds, so its speed rows say not recorded. It is a 2-bit build with most of its weights in system RAM.
The same model took 66 minutes on Nov 1 and 25 on Nov 7. Retries and context size swing a lot from run to run, so one night's order isn't a ranking.
The same six Dungeon checks failed every time it failed. I haven't dug into why yet. It could be a repeated model mistake or a problem in my checker.
The FP16 unlock
On Nov 5 the 3090 got the FP16-accumulate unlock: gpu-unlock, an open-source kernel module by Zagitta (v2.0.0). GeForce cards run some tensor-core math at a reduced rate, and the module lifts that.
These are quick synthetic tests, run once each (n=1), on the same card before and after. They are not the app-building bench.
Show the unlock results
Matrix math, TFLOPS (n=1)
| Precision | Locked (Nov 4) | Unlocked (Nov 5) | Change |
|---|---|---|---|
| FP16, FP32 accumulate | 77.5 | 125.3 | +62% |
| FP16, FP16 accumulate | 139.0 | 129.2 | −7% |
| BF16 | 77.8 | 126.5 | +63% |
| TF32 | 40.2 | 67.2 | +67% |
| FP32 | 28.2 | 26.3 | −7% |
- The two rows that went down 7% both ran at the card's power limit (420 W; plain FP32 drew 412 W). Why they dropped hasn't been tested.
- Image generation (SDXL, 1024×1024, 25 steps): 5.3 to 5.6 seconds with BF16, the same as FP16 with FP16 accumulate. There is no locked number to compare with. n=1.
- Text generation (qwen38-27b, GGUF): 45.5 tok/s. No gain was expected, since that work is limited by memory speed, not math. n=1.
App-building bench, unlocked vs locked (n=1)
On the night of Nov 5 the unlocked card ran the agentic part of the bench once. The server command and binary, sampling and test code were the same as the Nov 1 run. The card drew about 417 W against its 420 W limit, at 1800 MHz.
| 3090 at 420 W (Nov 1) | 3090 unlocked (Nov 5) | |
|---|---|---|
| Score | 32/32 | 32/32 |
| Attempts (Kanban, sand-sim, dungeon) | 1, 2, 3 | 1, 1, 2 |
| Total time | 66.0 min | 47.0 min |
| Kanban | 3.3 min, 5 turns | 13.4 min, 19 turns |
| Sand-sim | 9.3 min, 15 turns | 6.4 min, 10 turns |
| Dungeon | 53.1 min, 61 turns | 25.3 min, 37 turns |
| Generation speed | 33.5 tok/s | 38.1 tok/s |
| Prompt speed | 703 tok/s | 710 tok/s |
| Tokens generated | 124,851 | 98,378 |
| Peak context (dungeon) | 133,184 | 72,559 |
- Most of the 19 minutes saved came from needing 4 attempts instead of 6, not from a faster card.
- Prompt speed didn't change. Generation was 14% faster, but this run's contexts were about half as deep, and generation slows as context grows. So the gain can't be credited to the unlock. The synthetic tests above showed no text-generation gain either.
- One run per setup (n=1). Kanban took 4 times longer this time (19 turns vs 5) on its first attempt, with the same settings. Runs vary that much, so one run each can't separate a setup change from chance.
- The first dungeon attempt failed the same six checks again: 4 of 4 failed dungeon attempts, across three runs, failed exactly the same ones.
Energy and cost per token (bench load)
At $0.1227 per kWh. The GPU figures come from the card's own energy counters. The circuit figures come from a power meter on the whole circuit, one reading a minute. Costs are per million generated tokens.
| 3090 at 420 W (Nov 1) | 3090 unlocked (Nov 5) | |
|---|---|---|
| GPU energy, whole run | 432.6 Wh (avg 399 W) | 301.0 Wh (avg 393 W) |
| GPU cost, whole run | $0.053 | $0.037 |
| GPU, per 1M generated tokens | 3,465 Wh, $0.43 | 3,059 Wh, $0.38 |
| Whole circuit, average during the run | 699 W | 615 W |
| Whole circuit, per 1M generated tokens | $0.76 | $0.60 |
- These are bench-load numbers: full draw for the whole run. Normal chat use spends more time idle, so it costs more per token.
- Energy per token was 12% lower on the unlocked run at about the same draw, but most of that is the same catch as above: shallower contexts give more tokens per second.
- The circuit wasn't idle before the Nov 1 run, so the two circuit figures aren't a clean comparison.
Tweak log
- Nov 7, 3090 unlockedSame-night comparison. qwen38-27b 32/32 in 25 min, Qwen3.8 Flash Next 32/32 in 31 min.
- Nov 5, 3090 unlockedFP16-accumulate unlock (gpu-unlock) on the 3090. Synthetic results and one bench run are up.
- Nov 1, 3090 at 420 WValid run. 32/32 after 6 attempts, in 66 minutes.
- Nov 1, 3090 at 150 WThrown out. An old boot setting capped the card at 210 MHz.
- Oct 28, 2×3060Baseline. 32/32 after 4 attempts, in 69 minutes.