Datasets:
set stringclasses 2
values | arm stringclasses 4
values | precision stringclasses 4
values | role stringclasses 4
values | runtime_weight_memory_gib float64 5.44 15.3 | runtime_weight_memory_bytes int64 5.84B 16.4B | memory_source stringclasses 2
values | memory_reduction_vs_bf16_pct float64 60.9 64.4 β | effective_bpw float64 5.65 6.25 β | accuracy_delta_vs_bf16_pp float64 -1.78 0.98 β | tier_retention float64 0.89 0.93 β | korean_retention float64 0.86 0.94 β | ttft_ms_p50 float64 15.3 43.8 | ttft_ms_p95 float64 16 47.3 | tpot_ms_median float64 2.56 3.97 | tokens_per_s_batch1 float64 252 390 | tpot_ratio_vs_bf16 float64 0.65 1 | kernels stringclasses 4
values | kernel_mismatches float64 0 0 β | product_gate stringclasses 3
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
SEL-82 | BF16 | BF16 weights and activations | reference | 15.268 | 16,394,145,792 | D80/D81 measurement, same engine and GPU (not re-measured in D82) | null | null | null | null | null | 15.499 | 21.935 | 3.9696 | 251.9 | 1 | BF16 dense=144 | null | reference |
SEL-82 | A82 | NVFP4A16 on all 252 decoder linear modules (FP4 E2M1, group 16, 16-bit activations) + INT8 lm_head (sym, group 128) | candidate (D82 H0) | 5.438 | 5,838,677,504 | D82 measurement | 64.4 | 5.6497 | 0.98 | 0.9317 | 0.9136 | 15.302 | 16.038 | 3.6393 | 274.8 | 0.9168 | MarlinNvFp4LinearKernel=144 | 0 | PASS |
SEL-82 | A81 | NVFP4A16 body + 46 INT8 weight-only fused units (41 attention, 5 MLP) + INT8 lm_head | control (public D81 checkpoint) | 5.917 | 6,353,372,672 | D82 measurement | 61.2 | 6.1732 | -0.51 | 0.9255 | 0.9259 | 16.07 | 18.07 | 3.6746 | 272.1 | 0.9257 | MarlinLinearKernel=46;MarlinNvFp4LinearKernel=98 | 0 | PASS |
SEL-82 | NVFP4 W4A4 | NVFP4 weights and NVFP4 activations | report-only baseline | 5.977 | 6,417,801,216 | D82 measurement | 60.9 | 6.2481 | -1.78 | 0.889 | 0.8616 | 43.817 | 47.318 | 2.5828 | 387.2 | 0.6506 | FlashInferCuteDslNvFp4LinearKernel=144 | 0 | FAIL |
FINAL-82 | BF16 | BF16 weights and activations | reference | 15.268 | 16,394,145,792 | D80/D81 measurement, same engine and GPU (not re-measured in D82) | null | null | null | null | null | 16.235 | 18.975 | 3.956 | 252.8 | 1 | BF16 dense=144 | null | reference |
FINAL-82 | A82 | NVFP4A16 on all 252 decoder linear modules (FP4 E2M1, group 16, 16-bit activations) + INT8 lm_head (sym, group 128) | candidate (D82 H0) | 5.438 | 5,838,677,504 | D82 measurement | 64.4 | 5.6497 | 0.76 | 0.9317 | 0.9136 | 15.603 | 17.592 | 3.6359 | 275 | 0.9191 | MarlinNvFp4LinearKernel=144 | 0 | PASS |
FINAL-82 | A81 | NVFP4A16 body + 46 INT8 weight-only fused units (41 attention, 5 MLP) + INT8 lm_head | control (public D81 checkpoint) | 5.917 | 6,353,372,672 | D82 measurement | 61.2 | 6.1732 | 0.06 | 0.9317 | 0.9366 | 16.091 | 17.317 | 3.5076 | 285.1 | 0.8867 | MarlinLinearKernel=46;MarlinNvFp4LinearKernel=98 | 0 | PASS |
FINAL-82 | NVFP4 W4A4 | NVFP4 weights and NVFP4 activations | report-only baseline | 5.977 | 6,417,801,216 | D82 measurement | 60.9 | 6.2481 | -1.58 | 0.8879 | 0.8592 | 42.109 | 45.705 | 2.5627 | 390.2 | 0.6478 | FlashInferCuteDslNvFp4LinearKernel=144 | 0 | FAIL |
B200 Low-Bit Quantization Benchmarks: D82 aggregated results
Aggregated results of experiment D82, which tested whether
ThakiCloud/Qwen3-8B-NVFP4A16-INT8Head-B200 (A82: NVFP4A16 body + INT8 lm_head, no INT8 body units)
is non-inferior to the public A81 checkpoint ThakiCloud/Qwen3-8B-NVFP4A16-INT8-Mixed-B200 on Qwen3-8B,
one NVIDIA B200 and vLLM 0.29.0. D81 results stay unchanged in ThakiCloud/B200-LowBit-Quantization-Benchmarks.
This repository contains aggregates only. No evaluation items, per-item predictions, prompts or calibration text are
included. The evaluation sets are identified by their SHA-256 in data/run_metadata.json.
Verdict
One preregistered hypothesis (k = 1): A82 β A81 paired accuracy over all gated items, one-sided Ξ± = 0.025, margin β1.5 pp.
| Set | Paired n | A82 β A81 (pp) | 95% CI (pp) | Verdict |
|---|---|---|---|---|
| SEL-82 | 10198 | +1.49 | [+0.88, +2.10] | H0_SELECTED |
| FINAL-82 | 10210 | +0.71 | [+0.10, +1.31] | FINAL82_H0_PASS |
No superiority test was preregistered, so these results do not show that A82 is more accurate than A81.
Per-axis differences, A82 β A81 (report-only)
The numeric and reasoning axes were declared UNDERPOWERED before any result was seen. No axis-level claim is made.
| Axis | SEL-82 diff (pp) | SEL-82 95% CI | FINAL-82 diff (pp) | FINAL-82 95% CI |
|---|---|---|---|---|
| ko | +0.38 | [-0.53, +1.29] | -0.81 | [-1.71, +0.08] |
| en | -0.10 | [-0.89, +0.68] | -0.31 | [-1.17, +0.55] |
| number (underpowered) | +5.50 | [+4.26, +6.74] | +3.40 | [+2.19, +4.62] |
| reasoning (underpowered) | -3.47 | [-5.66, -1.28] | -0.40 | [-2.62, +1.82] |
On SEL-82 A82 was significantly behind A81 on reasoning; on FINAL-82 the interval includes zero. Neither result shows equivalence on reasoning.
FINAL-82 arms vs BF16
Memory is vLLM weight-load memory in bytes / 2^30. The BF16 memory value comes from D80/D81 (same engine and GPU) and was not re-measured in D82. Latency is batch 1, 32 prompts, 256 output tokens.
| Arm | Weight memory (GiB) | TTFT p50 (ms) | TPOT median (ms) | tok/s (batch 1) | Korean retention | Tier retention | Gate |
|---|---|---|---|---|---|---|---|
| BF16 | 15.268 | 16.235 | 3.956 | 252.8 | - | - | reference |
| A82 | 5.438 | 15.603 | 3.6359 | 275.0 | 0.9136 | 0.9317 | PASS |
| A81 | 5.917 | 16.091 | 3.5076 | 285.1 | 0.9366 | 0.9317 | PASS |
| NVFP4 W4A4 | 5.977 | 42.109 | 2.5627 | 390.2 | 0.8592 | 0.8879 | FAIL |
Concurrency (FINAL-82, report-only)
Aggregate tokens/s by number of concurrent requests (offline engine, no HTTP server; median of 3 runs).
| Arm | 1 | 8 | 32 | Request latency at 32 (ms) |
|---|---|---|---|---|
| BF16 | 252.3 | 2061.1 | 7939.6 | 1031.8 |
| A81 | 271.4 | 2105.6 | 6666.7 | 1228.8 |
| A82 | 274.9 | 2133.2 | 6749.7 | 1213.7 |
| NVFP4 W4A4 | 418.6 | 2948.2 | 10967.6 | 746.9 |
At 32 concurrent requests BF16 had the highest aggregate throughput of BF16, A81 and A82.
Files
data/arms.csv: per set and arm memory, latency, kernels, retention and gate status.data/axes.csv: per set, arm and axis accuracy for BF16 and the quantized arm.data/paired_comparisons.csv: every paired comparison with n, difference, 95% CI and status.data/concurrency.csv: aggregate throughput and request latency at 1, 8 and 32 concurrent requests.data/kernel_verification.json: kernel counts and float-fallback check per arm (FINAL-82).data/run_metadata.json: set hashes, preregistration hash, gates, environment, sources.recipes/a82_recipe.json: the A82 quantization recipe and checkpoint identity.REPRODUCE.md: how the numbers were produced and what cannot be reproduced from this repository alone.
Sources and license
Evaluation items came from KMMLU, KoBEST, MMLU, HellaSwag, BBH, BIG-bench, HRM8K and a programmatic numeric generator. No AI-Hub data was used. The aggregated numbers in this repository are released under CC BY 4.0; the underlying benchmarks keep their own licenses and are not redistributed here.
- Downloads last month
- 52