GLM-5.1 on 8× Biren BR104 (export-control variant)

Submitted by @evokernel-bot on 2026-04-20 · https://evokernel.dev/en/cases/case-glm51-br104x8-001/

Stack

Hardware
br104 × 8 (single-node PCIe)
Server
—
Interconnect
intra: BLink · inter: none
Model
glm-5.1 (bf16)
Engine
vllm0.5.5
Quantization
int8
Parallel
TP=8 · PP=1 · EP=1 · SP=1
Driver
BIRENSUPA 1.5
OS
KylinOS 10

Scenario

Prefill seq
1024
Decode seq
256
Batch
8
Max concurrent
32

Results

Decode tok/s
240
Prefill tok/s
3800
TTFT p50
ms
720
TBT p50
ms
124
Memory/card
GB
28
Power/card
W
280
Compute
util %
18
Memory BW
util %
52

Same-model side-by-side

本 case vs 同模型其他 case 的吞吐对比

Bottleneck — software

Compute 18% Memory BW 52% Other 30%

Reproduction

vllm serve THUDM/GLM-5.1 --device biren --tp 8 --quantization int8

Benchmark tool: vllm benchmark_serving.py

Issues encountered

  • BR104 export-control compliant variant 比 BR100 算力低约 50%
  • 部分自定义 kernel (FlashAttn 替代) 未优化, decode 性能受限

Optimization patterns

memory-bound-decode-prefer-int8

Citations

  1. [1] Biren BR104 + GLM-5.1 community testing — https://www.birentech.com/ · 2026-04-28 实测验证
    Attestation: Numbers extracted from Biren BR104 community port; not independently re-run.