Datasets:

agent-perf-bench
/

AgentPerfBench

source stringclasses 13 values	gpu stringclasses 4 values	model stringclasses 13 values	kernel_family stringclasses 7 values	kernel_name stringclasses 328 values	dtype stringclasses 2 values	held_out bool 2 classes	M float64 1 16.4k ⌀	N float64 8 248k ⌀	K float64 2.05k 29.6k ⌀	bs float64 1 1 ⌀	seq float64 1 16.4k ⌀	n_heads float64 8 64 ⌀	head_dim float64 64 256 ⌀	kv_heads float64 1 64 ⌀	numel float64 64 531M ⌀	op_type stringclasses 177 values	gpu_time_duration_ms float64 0 703	launch_block_size float64 16 512	launch_grid_size float64 1 1.04M	dram_bytes_sum float64 1.27 94.3B ⌀	launch_registers_per_thread float64 16 255
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	void gemmk1_kernel<int, float, 256, 5, 0, 0, 0, 0, cublasGemvTensorStridedBatched<const float>, cublasGemvTensorStridedBatched<const float>, cublasGemvTensorStridedBatched<float>, float, 0>(cublasGemmk1Params<T2, T9, T10, T11, T12, biasType<T11::value_type, T12>::type>)	bf16	true	null	null	null	1	128	null	null	null	null	gemv	0.00458	256	8	6,910	20
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13526	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02554	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02528	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13965	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58064	256	112	475,470,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.59558	256	112	475,510,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38464	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13949	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.0256	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02534	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.14077	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58614	256	112	475,850,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.58227	256	112	476,440,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.3881	256	96	477,150,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13898	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02554	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02518	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13338	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.60003	256	112	475,280,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.59392	256	112	474,630,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38301	256	96	477,190,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.14259	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.0256	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02528	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13763	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.59203	256	112	474,990,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.5912	256	112	475,320,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38675	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13395	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02509	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02525	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13795	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58624	256	112	475,870,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.59174	256	112	475,710,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38342	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13373	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02579	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02509	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.1384	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58982	256	112	475,880,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.58886	256	112	475,790,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.37894	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13363	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02573	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02525	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13872	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58163	256	112	475,630,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.58531	256	112	475,660,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38106	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13936	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02506	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02522	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13869	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58349	256	112	476,720,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.59296	256	112	475,500,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.388	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.14176	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02554	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02512	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13955	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58605	256	112	475,830,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.59654	256	112	475,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38694	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13158	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02515	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02493	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13603	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58077	256	112	474,930,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.59629	256	112	475,840,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38144	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13498	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02547	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02512	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.1369	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.59056	256	112	475,900,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.58938	256	112	475,640,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38128	256	96	477,140,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13917	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02515	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.0249	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13974	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58096	256	112	475,640,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.58496	256	112	475,360,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38742	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.14	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.02541	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02496	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.13894	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58774	256	112	476,280,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.59514	256	112	475,470,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.38282	256	96	477,130,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13574	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	k_proj	0.0255	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_128x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	1,024	8,192	1	128	null	null	null	null	v_proj	0.02493	128	72	18,900,000	246
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	o_proj	0.14115	256	96	136,340,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	gate	0.58477	256	112	475,870,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	28,672	8,192	1	128	null	null	null	null	up	0.58832	256	112	475,040,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	28,672	1	128	null	null	null	null	down_proj	0.39146	256	96	477,140,000	238
Llama-3.1-70B-Instruct_prefill	A100	Llama-70B	gemm	ampere_bf16_s16816gemm_bf16_256x128_ldg8_f2f_stages_64x3_tn	bf16	true	128	8,192	8,192	1	128	null	null	null	null	q_proj	0.13014	256	96	136,350,000	238

End of preview. Expand in Data Studio

AgentPerfBench

LLM inference benchmark: 3,197 main sweep rows and 37 per-layer kernel validation rows, plus 148,077 per-kernel NCU profiles, across 9 models, 14 GPU configurations, and 2 serving engines (vLLM 0.19.0, SGLang 0.5.9). All models served in BF16 except gpt-oss, which uses mxfp4 for projection weights.

Dataset configurations

trace_replay (2,932 rows)

Replays exact ISL/OSL sequences from recorded agent sessions (SWE-Bench, TerminalBench, OSWorld, ShareGPT). 77 unique (model, hardware, engine) combinations across 17 profiles.

17 profiles: chat-medium, chat-multiturn-long, chat-multiturn-medium, chat-multiturn-short, chat-short, chat-singleturn, coding-singleturn, decode-heavy, osworld-multiturn-long, osworld-multiturn-medium, osworld-multiturn-short, prefill-heavy, random-1k, swebench-multiturn-medium, swebench-multiturn-short, terminalbench-multiturn-medium, terminalbench-multiturn-short

synthetic_distributional (265 rows)

ISL/OSL sampled from lognormal fits to real workload statistics. 38 unique (model, hardware, engine) combinations across 5 profiles.

5 profiles: chat-multiturn-synth, chat-singleturn-synth, osworld-multiturn-synth, swebench-multiturn-synth, terminalbench-multiturn-synth

per_layer_kernel (37 rows)

Per-component operational intensity decomposition and Nsight Compute kernel profiles for Llama-3.1-8B on H100 (prefill phase). Analytical rows provide computed FLOPs, bytes, and OI per model component at batch sizes 1 and 80. NCU rows report measured SM and memory throughput per kernel from an 8-layer forward pass. Record types: analytical_total, analytical_component, ncu_kernel.

kernels_labeled (148,077 rows)

Per-kernel Nsight Compute (ncu) profiles across 4 GPUs (A100, H100, RTX 3090, RTX 2080 Ti) and 13 model/sweep sources.

mse_validation (28 rows)

Curated H100 / Llama-3.1-8B / vLLM validation table for the distributional synthetic replay generator. Paired synthetic and real trace replay runs; supplementary rows preserve no-replacement and high-concurrency debug runs. Raw JSON artifacts referenced through R2 URI columns. Per-run successful/failed request counts retained.

Quality filtering

Configurations where fewer than 75% of requests completed successfully are excluded. Summary metrics are computed from successful requests only.

Config	Rows
trace_replay	2,932
synthetic_distributional	265
per_layer_kernel	37
kernels_labeled	148,077
mse_validation	28

Coverage

Hardware

All benchmarks collected on PyTorch 2.10.0.

GPU	VRAM	HBM bandwidth	Peak half-precision TFLOPS
NVIDIA H100 SXM	80 GB	3.35 TB/s	989
NVIDIA A100 SXM4	40 GB	1.56 TB/s	312
NVIDIA RTX 3090	24 GB	936 GB/s	71
NVIDIA RTX 2080 Ti	11 GB	616 GB/s	27

Multi-GPU configurations: 1, 2, 4, or 8 GPUs with tensor parallelism.

Models

All models served in BF16 unless noted.

Model	Family	Parameters	Architecture	Notes
Llama-3.1-8B	Llama	8B	Dense
Llama-3.1-70B	Llama	70B	Dense
Llama-3.3-70B	Llama	70B	Dense
Qwen2.5-72B	Qwen	72B	Dense
Qwen3.5-9B	Qwen	9B	Dense
Qwen3.5-27B	Qwen	27B	Dense
Mixtral-8x7B	Mixtral	46.7B (12.9B active)	MoE
gpt-oss-20b	GPT-OSS	21B (3.6B active)	MoE	mxfp4 projections
gpt-oss-120b	GPT-OSS	117B (5.1B active)	MoE	mxfp4 projections

Engines

vLLM 0.19.0
SGLang 0.5.9

Schema

Each row in summary.parquet (trace_replay and synthetic_distributional):

Column	Type	Description
run_id	string	Deterministic hash of run parameters
model	string	Model short name
model_family	string	Model family (llama, qwen, gpt-oss, mixtral)
hardware	string	GPU configuration (e.g., H100x4)
engine	string	Serving engine (vllm, sglang)
tensor_parallelism	int	TP degree
profile	string	Workload profile name
concurrency	int	Concurrent request level
num_requests	int	Total requests in run
duration_s	float	Total run duration
request_throughput	float	Requests/second
input_token_throughput	float	Input tokens/second
output_token_throughput	float	Output tokens/second
total_token_throughput	float	Total tokens/second
mean/median/p90/p99_ttft_ms	float	Time to first token
mean/median/p90/p99_tpot_ms	float	Time per output token
mean/median/p90/p99_itl_ms	float	Inter-token latency
mean/median/p90/p99_e2el_ms	float	End-to-end latency

Loading

from datasets import load_dataset

ds = load_dataset("agent-perf-bench/AgentPerfBench", "trace_replay")
# or "synthetic_distributional", "per_layer_kernel", "kernels_labeled", "mse_validation"

Benchmark methodology

Closed-loop concurrency with semaphore control.
3-request warmup before each configuration.
Metrics: TTFT, TPOT, ITL, E2EL, request throughput, token throughput (mean, median, p90, p99).
Metrics computed over successful requests only.
Collection period: March 2026 onwards.

Limitations

Distributional profiles are fitted approximations, not direct production replays.
Closed-loop concurrency only; no open-loop (Poisson) arrivals.

Ethical considerations

No PII. Trace-replay profiles derive from open benchmarks (SWE-Bench MIT, TerminalBench, OSWorld). Synthetic profiles use random tokens.

License

Benchmark data released under Apache-2.0. Source datasets retain their original licenses.

Source datasets

Future releases

Additional hardware configurations and model families.
Open-loop (Poisson) arrival mode.
Additional per-kernel roofline profiles.

Citation

@inproceedings{agentperfbench2026,
  title={AgentPerfBench: A Benchmarking and Evaluation Suite for Inference Performance of Agentic LLMs},
  author={Anonymous},
  booktitle={NeurIPS 2026 Evaluations and Datasets Track},
  year={2026}
}

Downloads last month: 190