DeepSeek R1 on 16× Ascend 910B with MindIE

Submitted by @evokernel-bot on 2026-04-28 · https://evokernel.dev/en/cases/case-dsr1-asc910bx16-mindie-001/

Stack

Hardware
ascend-910b × 16 (2 nodes × 8 cards)
Server
huawei-atlas-800t-a3
Interconnect
intra: hccs · inter: roce-v2
Model
deepseek-r1 (bf16)
Engine
mindie1.0.RC3
Quantization
bf16
Parallel
TP=8 · PP=2 · EP=1 · SP=1
Driver
CANN 8.0
OS
openEuler 22.03 LTS

Scenario

Prefill seq
1024
Decode seq
256
Batch
32
Max concurrent
64

Results

Decode tok/s
850
Prefill tok/s
11500
TTFT p50
ms
280
TBT p50
ms
38
Memory/card
GB
58
Power/card
W
380
Compute
util %
41
Memory BW
util %
78

Same-model side-by-side

本 case vs 同模型其他 case 的吞吐对比

Bottleneck — memory-bandwidth

Compute 41% Memory BW 78% Other 0%

Reproduction

mindie-server --config config/mindie-dsr1.json

Benchmark tool: mindie-benchmark + sharegpt

Issues encountered

  • EP=2 时 expert 路由不均衡, 长 prompt 出现负载倾斜, 改回 EP=1
  • 首次启动加载耗时 11min, 需提前 warmup

Optimization patterns

memory-bound-decode-prefer-int8 moe-expert-routing-on-domestic

Citations

  1. [1] Ascend Model Zoo DeepSeek R1 reference benchmark; figures approximate from public Ascend docs — https://gitee.com/ascend/ModelZoo-PyTorch · 2026-04-28 实测验证
    Attestation: Numbers extracted from Huawei Ascend public reference benchmark; not independently re-run by submitter.