boltbeam.fyi

Qwen3.8-27B W4A16-AWQ on RTX 5090

safetensors INT4_G128_ZP+BF16

Engines

vLLM0.31.0

87.5 tok/s

80% of the roofline109.7 tok/s possible

now, a token
11.43 ms
floor
9.11 ms
read, a token
15.59 GB

Speed at context

128 tokens
vLLM90.1 tok/s
512 tokens
vLLM87.5 tok/s of 109.7
Measured at short contexts only. The pink tick is the roofline at that context.

Where the time goes

Now, a token
11.43 ms
Roofline floor, a token
9.11 ms
Gap to the floor
2.31 ms
Kernels a token
790

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

ffn_gate_up INT4_G128_ZP3.85 / 3.47 ms 90%
ffn_down INT4_G128_ZP1.94 / 1.73 ms 89%
attn_gate + attn_qkv INT4_G128_ZP1.39 / 1.22 ms 88%
recurrent0.34 / 0.18 ms 53%
ssm_out INT4_G128_ZP0.62 / 0.46 ms 74%
ssm_alpha_beta BF160.19 / 0.03 ms 15%
lm_head BF161.63 / 1.49 ms 92%
attention0.11 / 0.02 ms 20%
attn_kv + attn_qo INT4_G128_ZP0.43 / 0.36 ms 83%
attn_qo INT4_G128_ZP0.21 / 0.15 ms 75%
triton_red_fused__to_copy_add_fused_add…0.36 ms · no floor
kernel gaps0.19 ms · no floor
flash::flash_fwd_splitkv_combine_kernel0.06 ms · no floor
layer_norm_fwd_kernel0.05 ms · no floor
_fused_qk_rmsnorm_rope_gate_kernel0.03 ms · no floor
24 other small kernels0.15 ms · no floor
Per kernel (33): ms now / ms at the roofline, launches a token
Marlin8.44 / 7.40 ms ×256
gemm · lm_head BF161.63 / 1.49 ms ×1
triton_red_fused__to_copy_add_fused_add…0.36 ms · no floor ×128
_causal_conv1d_update_kernel + fused_recurrent_gated_delta_rule_packed_decode_kernel0.34 / 0.18 ms ×96
gemm · ssm_alpha_beta BF160.19 / 0.03 ms ×48
flash_fwd_splitkv_kernel0.11 / 0.02 ms ×16
flash::flash_fwd_splitkv_combine_kernel0.06 ms · no floor ×16
layer_norm_fwd_kernel0.05 ms · no floor ×48
_fused_qk_rmsnorm_rope_gate_kernel0.03 ms · no floor ×16
triton_poi_fused_mul_silu_slice_10.03 ms · no floor ×48
vllm::reshape_and_cache_flash_kernel0.02 ms · no floor ×16
_gumbel_sample_kernel0.02 ms · no floor ×1
triton_poi_fused_zeros_30.01 ms · no floor ×32
triton_poi_fused_mul_sigmoid_view_00.01 ms · no floor ×16
triton_poi_fused_mul_silu_slice_20.01 ms · no floor ×16
at::native::index_elementwise_kernel0.01 ms · no floor ×3
triton_poi_fused_zeros_40.01 ms · no floor ×15
triton_red_fused__to_copy_add_embedding…0.01 ms · no floor ×1
_bias_kernel0.00 ms · no floor ×1
at::native::reduce_kernel0.00 ms · no floor ×1
_gather_block_tables_kernel0.00 ms · no floor ×1
_post_update_kernel0.00 ms · no floor ×1
_compute_slot_mappings_kernel0.00 ms · no floor ×1
at::native::unrolled_elementwise_kernel0.00 ms · no floor ×1
_combine_sampled_and_draft_tokens_kernel0.00 ms · no floor ×1
_prepare_rope_positions_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::_scatter_gather_elementwise…0.00 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 rtx5090-27b-vllm-batch-b

Tested on

GPU
RTX 5090 · 32 GB
System RAM
31 GB
CPU threads
16
Memory, measured
1710 GB/s
Temp, max
77 °C
Power
178 W mean of 600 W
SM clock
2211 MHz, throttled
Driver
595.99.02
Measured
2026-11-09 · 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