boltbeam.fyi

Qwen3.8-27B W4A16-AWQ on L40S

safetensors INT4_G128_ZP+BF16

Engines

vLLM

0.31.0 ↗measured 2026-11-09 19:53 UTCnewer release at the time: 0.33.0

43.6 tok/s

89% of the roofline48.9 tok/s possible

now, a token
22.95 ms
floor
20.46 ms
read, a token
15.59 GB
More
memory, measured
762 GB/s

What was run

python vllm_decode_drive.py --model Qwen3.8-27B-W4A16-AWQ --contexts 512 --batches 1 --tokens 64 --reps 1 --once 64 --max-num-batched-tokens 584 --model-context-limit 584 --capture-pass paired

Speed at context

128 tokens
vLLM43.3 tok/s
512 tokens
vLLM43.6 tok/s of 48.9
The pink tick is the roofline at that context.

Where the time goes

Now, a token
22.95 ms
Roofline floor, a token
20.46 ms
Gap to the floor
2.49 ms
Kernels a token
838

Per role ms a token. floor (bytes ÷ measured memory speed) above the floor

ffn_gate_up INT4_G128_ZP8.38 / 7.79 ms 93%
ffn_down INT4_G128_ZP4.18 / 3.89 ms 93%
ssm_out INT4_G128_ZP1.24 / 1.03 ms 83%
attn_gate + attn_qkv INT4_G128_ZP2.93 / 2.75 ms 94%
lm_head BF163.51 / 3.34 ms 95%
ssm_alpha_beta BF160.21 / 0.06 ms 29%
attention0.12 / 0.05 ms 38%
recurrent0.48 / 0.41 ms 84%
attn_qo INT4_G128_ZP0.41 / 0.34 ms 83%
attn_kv + attn_qo INT4_G128_ZP0.86 / 0.80 ms 94%
triton_red_fused__to_copy_add_fused_add…0.29 ms · no floor
triton_per_fused__to_copy_add_mean_mul_…0.17 ms · no floor
kernel gaps0.10 ms · no floor
triton_poi_fused_mul_silu_slice_20.08 ms · no floor
flash::flash_fwd_splitkv_combine_kernel0.07 ms · no floor
25 other small kernels0.22 ms · no floor
Per kernel (34): ms now / ms at the roofline, launches a token
Marlin18.01 / 16.61 ms ×256
gemm · lm_head BF163.51 / 3.34 ms ×1
_causal_conv1d_update_kernel + fused_recurrent_gated_delta_rule_packed_decode_kernel0.48 / 0.41 ms ×96
triton_red_fused__to_copy_add_fused_add…0.29 ms · no floor ×128
gemm · ssm_alpha_beta BF160.21 / 0.06 ms ×48
triton_per_fused__to_copy_add_mean_mul_…0.17 ms · no floor ×48
flash_fwd_splitkv_kernel0.12 / 0.05 ms ×16
triton_poi_fused_mul_silu_slice_20.08 ms · no floor ×64
flash::flash_fwd_splitkv_combine_kernel0.07 ms · no floor ×16
triton_poi_fused_40.04 ms · no floor ×47
triton_poi_fused_zeros_50.04 ms · no floor ×47
_fused_qk_rmsnorm_rope_gate_kernel0.03 ms · no floor ×16
vllm::reshape_and_cache_flash_kernel0.03 ms · no floor ×16
triton_poi_fused_mul_sigmoid_view_00.02 ms · no floor ×16
at::native::index_elementwise_kernel0.01 ms · no floor ×3
_gumbel_sample_kernel0.01 ms · no floor ×1
_bias_kernel0.01 ms · no floor ×1
_gather_block_tables_kernel0.00 ms · no floor ×1
at::native::reduce_kernel0.00 ms · no floor ×1
_prepare_rope_positions_kernel0.00 ms · no floor ×1
triton_red_fused__to_copy_add_embedding…0.00 ms · no floor ×1
_combine_sampled_and_draft_tokens_kernel0.00 ms · no floor ×1
_post_update_kernel0.00 ms · no floor ×1
_get_num_sampled_and_rejected_kernel0.00 ms · no floor ×1
at::native::vectorized_elementwise_kern…0.00 ms · no floor ×3
at::native::unrolled_elementwise_kernel0.00 ms · no floor ×1
_compute_slot_mappings_kernel0.00 ms · no floor ×1
at::native::_scatter_gather_elementwise…0.00 ms · no floor ×1
triton_poi_fused_10.00 ms · no floor ×1
_prepare_pos_seq_lens_kernel0.00 ms · no floor ×1

No floor = the kernel reads no weights (norms, conversions, copies): all of its time is gap.

Run l40s-qwen3.8-27b-awq-vllm

Tested on

GPU
L40S · 45 GB
System RAM
2016 GB
CPU threads
192
Memory, measured
762 GB/s
Temp, max
58 °C
Power
133 W mean of 350 W
SM clock
2509 MHz, throttled
Driver
570.211.01
Measured
2026-11-09 19:53 UTC · context 512 tokens · measured, in-model (nsys): split by role from the engine's own labels, 9 weight roles

Qwen3.8-27B on other GPUs