Experiments

Every experiment with its verdict, so none of them get run twice without new evidence.

rejected is as useful as accepted. Re-running anything listed here needs new evidence, and the entry has to say what changed.

Date Title Verdict Categories
2026-09-08 MTU 9000 moves the fabric roofline to 24.5 GB/s one-way, 48.9 duplex accepted experiment, fabric, hardware
2026-09-08 A reduction barrier makes Qwen GDN graph replay bit exact accepted incident, kernels, attention, numerics
2026-09-08 Moving the KDA state tile out of shared memory is 4.6x at M=8 accepted experiment, kernels, attention, memory
2026-09-08 Radix-paged KV holds 93 streams at 262k where 8 fit unshared accepted experiment, cache, memory, attention
2026-09-08 Qwen3.8 linear-attention FP8 passes 12 of 12 quality cases accepted experiment, numerics, attention, telemetry, baseline
2026-09-08 Qwen3.8 base linear attention passes all 12 quality cases in NVFP4 accepted experiment, numerics, attention, telemetry, baseline
2026-09-08 Qwen3.8 base full attention passes all 12 quality cases in NVFP4 accepted experiment, numerics, attention, memory, telemetry, baseline
2026-09-08 Qwen3.8 base routers and PLE pass all 12 quality cases in NVFP4 accepted experiment, numerics, moe, attention, memory, telemetry, baseline
2026-09-08 Whole-family linear-attention NVFP4 reaches 102.76 tok/s at c16 accepted experiment, baseline, numerics, attention, runtime
2026-09-08 Head’s 2.3x slowdown is a 728 MHz GPU clock ceiling, not kswapd accepted experiment, hardware, memory, baseline
2026-09-08 Display-power renegotiation does not lift the head’s 728 MHz cap rejected incident, hardware, baseline
2026-09-08 A cold power cycle lifts the head from 728 MHz to 2418 MHz accepted experiment, hardware, baseline
2026-09-08 Removing display power restores the head’s normal GPU clock range open incident, hardware, baseline
2026-09-08 K1 wins the matched Qwen3.8 K1 to K7 general-engine sweep rejected experiment, runtime, scheduler, hardware, baseline
2026-09-08 Local EXT4 puts Qwen3.8 first token at 468.83 seconds accepted experiment, runtime, memory, hardware, baseline
2026-09-08 Static MTP1 wins the Qwen3.8 general-engine depth sweep rejected experiment, runtime, baseline, hardware
2026-09-08 The 728 MHz power state inflated Qwen3.8 loading by 5.1x accepted experiment, runtime, hardware, baseline
2026-09-07 Rank-local E256 MoE misses the Qwen K1 and K4 step budgets rejected experiment, moe, kernels, numerics, telemetry
2026-09-07 The expert split decodes the same tokens; the fabric is idle on one rank and the wall on the other accepted experiment, fabric, moe, runtime, numerics
2026-09-07 The pair is 6.1x one booster, not 2.9x behind it, and the expert cache is most of the ladder accepted experiment, fabric, cache, memory, moe, runtime, baseline
2026-09-07 Exact-N640 retained tail wins only at Qwen c4 accepted experiment, kernels, numerics, telemetry
2026-09-07 A TP2 vocab GEMV cuts Qwen3.8 M1 head latency by 31% accepted experiment, kernels, numerics, telemetry
2026-09-07 Exact-N640 dynamic MoE cuts Qwen c16 latency by 17.7% accepted experiment, kernels, numerics, telemetry
2026-09-07 The Qwen controller publishes one authenticated TP2 boundary accepted experiment, cache, runtime
2026-09-07 Layer 3 completes the exact Qwen attention state transition accepted experiment, attention, runtime, kernels, fabric, telemetry
2026-09-07 Qwen3.8 linear-attention NVFP4 removes 2.791 GiB per c16 step accepted experiment, numerics, attention, telemetry, tooling
2026-09-07 Unmodified Qwen3.8 passes the 12-case quantization gate accepted experiment, numerics, baseline
2026-09-07 The fully touched Qwen3.8 c16 state allocation fits beside the 62.46 GiB model accepted experiment, memory, cache, attention
2026-09-07 The Qwen3.8 state plan counts 2,088,755,200 CUDA bytes per rank-stream accepted research, memory, cache, attention
2026-09-07 Qwen3.8 reaches 163.40 tok/s at c16 without speculation on 64 KiB pages accepted experiment, baseline, runtime
2026-09-07 One-tensor staging loads Qwen3.8 on 64 KiB pages accepted experiment, runtime, memory
2026-09-07 One booster decodes GLM-5.3-Flash end to end, and a token reads 21.05 GiB, not the recorded 9.6 accepted experiment, runtime, moe, memory, attention, numerics, baseline
2026-09-07 The CUTLASS grouped GEMM beats the scalar GEMV loop at every M, M=1 included, so the crossover is 1 accepted experiment, runtime, moe, kernels, numerics
2026-09-07 Tensor dispatch and duplicate passes dominate Qwen3.8 startup accepted experiment, runtime, memory
2026-09-06 Two idle NIC ports held 444 MiB; unbinding them costs nothing accepted experiment, fabric, memory, hardware
2026-09-06 Deleting the key transpose puts the DSA indexer scan on its streaming floor accepted kernels, experiment, attention, memory
2026-09-06 Fusing the MoE layer’s two grouped GEMMs is worth 4%, and a CUDA graph already collects it rejected experiment, kernels, moe, memory, numerics
2026-09-06 CUTLASS NVFP4 grouped GEMM runs on sm_121 unpatched and holds 80% of roofline at MoE decode shapes rejected experiment, kernels, numerics, baseline
2026-09-06 The indexer caches per-token gates too: 7 unshared 262k streams per booster accepted research, cache, attention, memory
2026-09-06 At the 262k cap the DSA indexer costs 12x the sparse MLA it selects for accepted experiment, attention, kernels, cache
2026-09-06 Real NVFP4 scales need the sm_100 swizzle, and a booster loads in 14 s accepted experiment, numerics, memory, kernels
2026-09-06 vLLM reaches 207 tok/s at c8, while the c16 weight roofline is 282 tok/s accepted experiment, baseline, memory, numerics
2026-09-06 The idle GPU driver holds 0.7 GiB; the 64K slab tax holds 1.4 accepted experiment, hardware, memory
2026-09-06 A GB10 node reads at 238 GB/s and the pair’s fabric moves 23 GB/s accepted experiment, hardware, memory, fabric
2026-09-06 3.5 GiB per node is firmware-owned; asking NVIDIA for a headless option open research, hardware, memory
2026-09-06 Stripping the OS returns 217 MiB; the rest is carveout and driver floor rejected experiment, hardware, memory
2026-09-06 The 98 GiB CUDA ceiling was page cache, not a limit accepted experiment, hardware, memory, runtime
2026-09-06 Dropping random kmalloc caches returns 0.9 GiB; mlx5 owns more than the GPU driver accepted experiment, hardware, memory
2026-09-05 gh and quarto install clean on arm64 accepted infrastructure, tooling
2026-09-05 Specializing to one fuel and one GPU is worth 1.2x to 1.5x at batch 1 accepted research, kernels, runtime
2026-09-05 cudaHostRegister needs serialization and row-aligned chunks accepted research, memory, hardware
2026-09-05 General engines reach 132.8 tok/s at c32 on this pair accepted baseline, runtime
2026-09-05 Both sparks are identical: GB10 sm_121, 64 KiB pages, 123.73 GiB accepted experiment, hardware, fabric
2026-09-05 64 KiB pages return 2.10 GiB per node accepted experiment, hardware, memory
2026-09-05 Three GB10 limits that decide how kernels get written accepted research, hardware, kernels
2026-09-05 TensorRT-LLM is not the runtime for rocket rejected research, runtime, kernels
2026-09-05 TensorRT-LLM: rejected as a runtime, not as a kernel source rejected research, runtime, kernels
No matching items