Kernel measurements
GLM-5.3-Flash-NVFP4 on NVIDIA GB10 (DGX Spark)
62 per-kernel times, measured with nsys at engine commit c5b9a80e6 in the regime each row names, which is not the plan arm the blueprint page draws. 24 rows time a kernel this plan launches; the other 38 time kernels the legacy engine launches outside the plan. Each row names its source line and record hash.
Kernels this plan launches
| Kernel | Regime and shape | Time (µs) | Floor (µs) | Of floor | Bound | Source · record |
|---|---|---|---|---|---|---|
glm5next_ffn::glm5next_moe_combine | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k rows=2 top_k=8 H=4096 | 103.33 | 0.66 | 0.64% | memory | docs/kernel-perf/measurements.toml:3016-3029 (file at c8175076d)82b48821e8d5 |
w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k DSA o_absorb M=2 N=4096 K=11264 | 201.98 | 185.51 | 91.85% | memory | docs/kernel-perf/measurements.toml:3046-3059 (file at c8175076d)065db9f9b9b1 |
glm5next_mhc_c1::glm5next_hc_pre_norm_c1 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 10.62 | 3.85 | 36.24% | memory | docs/kernel-perf/measurements.toml:3091-3104 (file at c8175076d)8fa57ad78499 |
w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k DSA q_absorb M=2 N=11264 K=1536 | 82.43 | 69.86 | 84.75% | memory | docs/kernel-perf/measurements.toml:3121-3134 (file at c8175076d)c3a1be313177 |
glm5next_ffn::glm5next_swiglu_clamp | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k shared expert n=2x768 | 8.67 | 0.04 | 0.43% | memory | docs/kernel-perf/measurements.toml:3181-3194 (file at c8175076d)d310d9893753 |
kda_layer_ops::kda_o_norm_gated_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 15.36 | 0.18 | 1.18% | memory | docs/kernel-perf/measurements.toml:3211-3224 (file at c8175076d)dda13cbce773 |
w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8_split4 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 45.15 | 38.01 | 84.18% | memory | docs/kernel-perf/measurements.toml:3226-3239 (file at c8175076d)59a9ef414212 |
w4a4_gemv_mx::w4a4_gemv_mx8 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 52.42 | 37.98 | 72.47% | memory | docs/kernel-perf/measurements.toml:3241-3254 (file at c8175076d)1e5eccd40ef3 |
glm5next_ffn::glm5next_router_topk_rank | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k rows=2 experts=288 top_k=8 | 7.87 | 0.01 | 0.18% | memory | docs/kernel-perf/measurements.toml:3256-3269 (file at c8175076d)a555a317557b |
w4a4_gemv_mx_moe::w4a4_quant_rows_static | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 5.44 | 0.08 | 1.55% | memory | docs/kernel-perf/measurements.toml:3301-3314 (file at c8175076d)33222668aad5 |
kda_gates_w8a8::kda_gates_w8a8_k128_rows | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 7.78 | 3.22 | 41.38% | memory | docs/kernel-perf/measurements.toml:3316-3329 (file at c8175076d)58f52bf4ec48 |
kda_layer_ops::kda_pack_qkv_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 6.98 | 0.27 | 3.89% | memory | docs/kernel-perf/measurements.toml:3346-3359 (file at c8175076d)9662573fd1a9 |
glm5next_ffn::glm5next_moe_combine | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k rows=2 top_k=8 H=4096 | 103.74 | 0.66 | 0.63% | memory | docs/kernel-perf/measurements.toml:3465-3478 (file at c8175076d)f743326ffa78 |
w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k DSA o_absorb M=2 N=4096 K=11264 | 199.65 | 185.51 | 92.92% | memory | docs/kernel-perf/measurements.toml:3495-3508 (file at c8175076d)c248b17cadf4 |
glm5next_mhc_c1::glm5next_hc_pre_norm_c1 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 10.69 | 3.85 | 36.03% | memory | docs/kernel-perf/measurements.toml:3555-3568 (file at c8175076d)0b7bc1dd5569 |
w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k DSA q_absorb M=2 N=11264 K=1536 | 81.66 | 69.86 | 85.54% | memory | docs/kernel-perf/measurements.toml:3585-3598 (file at c8175076d)a78bae20d133 |
glm5next_ffn::glm5next_swiglu_clamp | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k shared expert n=2x768 | 8.7 | 0.04 | 0.43% | memory | docs/kernel-perf/measurements.toml:3645-3658 (file at c8175076d)e6d49b5751bb |
kda_layer_ops::kda_o_norm_gated_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 15.81 | 0.18 | 1.15% | memory | docs/kernel-perf/measurements.toml:3675-3688 (file at c8175076d)5a6f4e539eee |
w8a8_gemv::w8a8_gemv_rowscale_mb1_ku8_split4 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 44.96 | 38.01 | 84.54% | memory | docs/kernel-perf/measurements.toml:3690-3703 (file at c8175076d)c69efef2f0f3 |
w4a4_gemv_mx::w4a4_gemv_mx8 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 52.38 | 37.98 | 72.51% | memory | docs/kernel-perf/measurements.toml:3705-3718 (file at c8175076d)82ba5b0c58c8 |
glm5next_ffn::glm5next_router_topk_rank | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k rows=2 experts=288 top_k=8 | 8.77 | 0.01 | 0.16% | memory | docs/kernel-perf/measurements.toml:3720-3733 (file at c8175076d)d7d174c6b43e |
w4a4_gemv_mx_moe::w4a4_quant_rows_static | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 5.41 | 0.08 | 1.56% | memory | docs/kernel-perf/measurements.toml:3780-3793 (file at c8175076d)29cafbd36684 |
kda_gates_w8a8::kda_gates_w8a8_k128_rows | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 7.81 | 3.22 | 41.21% | memory | docs/kernel-perf/measurements.toml:3795-3808 (file at c8175076d)b34361bb0a86 |
kda_layer_ops::kda_pack_qkv_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 6.94 | 0.27 | 3.91% | memory | docs/kernel-perf/measurements.toml:3810-3823 (file at c8175076d)204b4a8231c0 |
Kernels outside this plan
| Kernel | Regime and shape | Time (µs) | Floor (µs) | Of floor | Bound | Source · record |
|---|---|---|---|---|---|---|
w4a16_gemv_pipe::w4a16_gemv_pipe_m2 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k shared expert down M=2 N=4096 K=768 | 191.07 | 7.18 | 3.76% | memory | docs/kernel-perf/measurements.toml:3001-3014 (file at c8175076d)8feeba49fb41 |
w4a16_gemv_pipe::w4a16_gemv_pipe_m2 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k KDA in-proj q|k|v|f_a|b|g_a M=2 N=8726 K=4096 | 90.37 | 80.95 | 89.58% | memory | docs/kernel-perf/measurements.toml:3031-3044 (file at c8175076d)f8ff990c4ad3 |
w4a16_gemv_pipe::w4a16_gemv_pipe_m2 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k KDA o_proj M=2 N=4096 K=2816 | 34.62 | 26.17 | 75.58% | memory | docs/kernel-perf/measurements.toml:3061-3074 (file at c8175076d)51e5fc3100bd |
w4a16_gemv_pipe::w4a16_gemv_pipe_m2 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k shared expert gate|up M=2 N=768+768 K=4096 | 25.22 | 14.3 | 56.72% | memory | docs/kernel-perf/measurements.toml:3076-3089 (file at c8175076d)d52b61a0bbae |
screen_head::screen_head_scan | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 938.54 | 877.87 | 93.54% | memory | docs/kernel-perf/measurements.toml:3106-3119 (file at c8175076d)242141bc0b2c |
dense_gemv_bf16_batchm_few::dense_gemv_bf16_batchm_few_fp32out | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 20.74 | 9.55 | 46.06% | memory | docs/kernel-perf/measurements.toml:3136-3149 (file at c8175076d)ab6edcc1be42 |
dense_gemv_fp8w::dense_gemv_fp8w | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 855.78 | 850.4 | 99.37% | memory | docs/kernel-perf/measurements.toml:3151-3164 (file at c8175076d)022e3b54f2ce |
dense_gemv_bf16_batchm_few::dense_gemv_bf16_batchm_few_fp32out_pair | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 60.93 | 51.81 | 85.03% | memory | docs/kernel-perf/measurements.toml:3166-3179 (file at c8175076d)1579c117f9f6 |
kda_recurrent::kda_recurrent_decode_bf16_seq_reg_snap | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 15.39 | 17.69 | 114.92% | memory | docs/kernel-perf/measurements.toml:3196-3209 (file at c8175076d)91ed5ebca753 |
dense_gemv_bf16::dense_gemv_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k MTP o_absorb M=1 N=4096 K=11264 | 408.13 | 370.7 | 90.83% | memory | docs/kernel-perf/measurements.toml:3271-3284 (file at c8175076d)ca2e67025687 |
kda_conv_tokens::causal_conv1d_window_advance_to | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 9.38 | 1.02 | 10.86% | memory | docs/kernel-perf/measurements.toml:3286-3299 (file at c8175076d)beab3ec949fb |
glm5next_dsa_mla_decode_hb::glm5next_dsa_mla_decode_hb_fp8 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k rows=2 heads=22 latent=512 splits=8 | 21.86 | 3.81 | 17.44% | memory | docs/kernel-perf/measurements.toml:3331-3344 (file at c8175076d)5c74482a59a3 |
glm5next_hc_post_sum::glm5next_hc_post_rank_sum | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 2.21 | 1.32 | 59.63% | memory | docs/kernel-perf/measurements.toml:3361-3374 (file at c8175076d)6321267ea7ce |
dense_gemv_bf16::dense_gemv_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k MTP catch-up eh_proj M=1 N=4096 K=8192 | 294.7 | 269.61 | 91.49% | memory | docs/kernel-perf/measurements.toml:3376-3389 (file at c8175076d)d75c51909baf |
kda_conv_tokens::causal_conv1d_update_l2norm_tokens | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 4.77 | 1.09 | 22.77% | memory | docs/kernel-perf/measurements.toml:3391-3404 (file at c8175076d)5a4ee11d7d99 |
w4a16_gemv::glm5next_moe_row_union | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 3.46 | 0 | 0.03% | memory | docs/kernel-perf/measurements.toml:3406-3419 (file at c8175076d)d77d77c67077 |
dense_gemv_bf16::dense_gemv_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k MTP q_absorb M=1 N=11264 K=1536 | 146.21 | 139.07 | 95.12% | memory | docs/kernel-perf/measurements.toml:3421-3434 (file at c8175076d)3b57068a194a |
w4a16_gemv::w4a16_gemv_sw_moe | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k MTP routed gate or up slots=8 N=704 K=4096 | 65.78 | 52.19 | 79.35% | memory | docs/kernel-perf/measurements.toml:3436-3449 (file at c8175076d)3d1b5306c52e |
w4a16_gemv_pipe::w4a16_gemv_pipe_m2 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k shared expert down M=2 N=4096 K=768 | 190.69 | 7.18 | 3.77% | memory | docs/kernel-perf/measurements.toml:3450-3463 (file at c8175076d)605a0dbce007 |
w4a16_gemv_pipe::w4a16_gemv_pipe_m2 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k KDA in-proj q|k|v|f_a|b|g_a M=2 N=8726 K=4096 | 90.21 | 80.95 | 89.73% | memory | docs/kernel-perf/measurements.toml:3480-3493 (file at c8175076d)071818dfd82a |
dsa_scores_heads_window::dsa_index_scores_heads_window | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 130.56 | 67.56 | 51.75% | memory | docs/kernel-perf/measurements.toml:3510-3523 (file at c8175076d)8a13d8cff3f9 |
w4a16_gemv_pipe::w4a16_gemv_pipe_m2 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k KDA o_proj M=2 N=4096 K=2816 | 34.43 | 26.17 | 76% | memory | docs/kernel-perf/measurements.toml:3525-3538 (file at c8175076d)ad43ebab2b62 |
w4a16_gemv_pipe::w4a16_gemv_pipe_m2 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k shared expert gate|up M=2 N=768+768 K=4096 | 24.26 | 14.3 | 58.97% | memory | docs/kernel-perf/measurements.toml:3540-3553 (file at c8175076d)0b1bd2175d74 |
screen_head::screen_head_scan | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 939.07 | 877.87 | 93.48% | memory | docs/kernel-perf/measurements.toml:3570-3583 (file at c8175076d)8ea02ef3023c |
dense_gemv_fp8w::dense_gemv_fp8w | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 871.39 | 850.4 | 97.59% | memory | docs/kernel-perf/measurements.toml:3600-3613 (file at c8175076d)0534f44bf7d5 |
dense_gemv_bf16_batchm_few::dense_gemv_bf16_batchm_few_fp32out | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 20.19 | 9.55 | 47.3% | memory | docs/kernel-perf/measurements.toml:3615-3628 (file at c8175076d)9524f9b27027 |
dense_gemv_bf16_batchm_few::dense_gemv_bf16_batchm_few_fp32out_pair | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 60.19 | 51.81 | 86.07% | memory | docs/kernel-perf/measurements.toml:3630-3643 (file at c8175076d)7f61f6cdf965 |
kda_recurrent::kda_recurrent_decode_bf16_seq_reg_snap | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 15.84 | 17.69 | 111.67% | memory | docs/kernel-perf/measurements.toml:3660-3673 (file at c8175076d)5089f92ad978 |
dense_gemv_bf16::dense_gemv_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k MTP o_absorb M=1 N=4096 K=11264 | 407.39 | 370.7 | 90.99% | memory | docs/kernel-perf/measurements.toml:3735-3748 (file at c8175076d)1a484cd9381e |
glm5next_dsa_mla_decode_hb::glm5next_dsa_mla_decode_hb_fp8 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k rows=2 heads=22 latent=512 splits=8 | 32.48 | 4.64 | 14.3% | memory | docs/kernel-perf/measurements.toml:3750-3763 (file at c8175076d)c68817d06ed0 |
kda_conv_tokens::causal_conv1d_window_advance_to | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 9.15 | 1.02 | 11.12% | memory | docs/kernel-perf/measurements.toml:3765-3778 (file at c8175076d)47871f726509 |
dense_gemv_bf16::dense_gemv_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k MTP catch-up eh_proj M=1 N=4096 K=8192 | 296.16 | 269.61 | 91.04% | memory | docs/kernel-perf/measurements.toml:3825-3838 (file at c8175076d)e8d7e9d566fd |
glm5next_hc_post_sum::glm5next_hc_post_rank_sum | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 2.46 | 1.32 | 53.43% | memory | docs/kernel-perf/measurements.toml:3840-3853 (file at c8175076d)446835e6b59a |
w4a16_gemv::glm5next_moe_row_union | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 3.46 | 0 | 0.03% | memory | docs/kernel-perf/measurements.toml:3855-3868 (file at c8175076d)950c76bbcb26 |
kda_conv_tokens::causal_conv1d_update_l2norm_tokens | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k | 4.67 | 1.09 | 23.24% | memory | docs/kernel-perf/measurements.toml:3870-3883 (file at c8175076d)5d1c985e6cc7 |
dense_gemv_bf16::dense_gemv_bf16 | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈128k MTP q_absorb M=1 N=11264 K=1536 | 145.54 | 139.07 | 95.56% | memory | docs/kernel-perf/measurements.toml:3885-3898 (file at c8175076d)cbdd139ade2e |
w4a4_gemv_mx_moe::w4a4_gemv_mx8_moe_union_sweep_sl_cs | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 224.13 | 173.17 | 77.26% | memory | docs/kernel-perf/measurements.toml:3900-3913 (file at c8175076d)11a901435a8e |
w4a4_gemv_mx_moe::w4a4_gemv_mx8_moe_union_sweep_k64_cs | decode C=1 (R=2, MTP k=1, TP=3 rank 0) KV≈1k | 104 | 87.61 | 84.24% | memory | docs/kernel-perf/measurements.toml:3915-3928 (file at c8175076d)a32ff83ead39 |