Performance

벤치마크 카탈로그

단일 노드, 클러스터, 이기종 GPU 구성 전반에 걸친 투명하고 재현 가능한 벤치마크.

Single Node

단일 노드 추론 — Moreh vLLM

Moreh vLLM 기반 서버 단위 추론 성능.

InferenceMAX DeepSeek R1 0528

8× AMD MI300X

Throughput geomean

Moreh vLLM1.47×
SGLang1.0×
Blog Post

InferenceMAX DeepSeek R1 0528

8× AMD MI300X

E2E Latency geomean (낮을수록 좋음)

Moreh vLLM0.68×
SGLang1.0×
Blog Post

DeepSeek R1 671B

8× AMD MI300X

Output TPS geomean

Moreh vLLM1.68×
ROCm vLLM1.0×
Technical Report

DeepSeek R1 671B

8× AMD MI300X

TTFT (낮을수록 좋음)

Moreh vLLM0.57×
ROCm vLLM1.0×
Technical Report

Llama 3.3 70B

2× AMD MI300X

Output TPS geomean

Moreh vLLM1.74×
ROCm vLLM1.0×
Technical Report

Llama 3.3 70B

2× AMD MI300X

TTFT (낮을수록 좋음)

Moreh vLLM0.50×
ROCm vLLM1.0×
Technical Report

Step3 321B

8× AMD MI308X

Decode TPS

Moreh vLLM4,082
NVIDIA H20 baseline3,147
Customer Case

Telco 7.8B LLM

1× AMD MI300X

Output TPS

Moreh vLLM (MI300X)186.75
vLLM (H100)143.39
Customer Case

Telco 7.8B LLM

1× AMD MI300X

SLO-compliant max concurrency

Moreh vLLM (MI300X)880
vLLM (H100)636
Customer Case
Cluster

클러스터 추론 — MoAI Inference Framework

PD disaggregation, 지능형 라우팅 등 클러스터 규모 최적화.

DeepSeek R1 671B

5× AMD MI300X nodes

Output tok/s per decode node

PD disagg + EP22,000+
Docs

DeepSeek R1 671B

5× AMD MI300X nodes

End-to-end latency (낮을수록 좋음)

PD disaggregation0.74×
Non-disaggregated1.0×
Technical Report

DeepSeek R1 671B

2× vs 5× AMD MI300X nodes

Throughput

Cache-aware (2 nodes)2.2×
Naive routing (5 nodes)1.0×
Docs

DeepSeek R1 671B

2× vs 5× AMD MI300X nodes

TTFT (낮을수록 좋음)

Cache-aware (2 nodes)0.03–0.05×
Naive routing (5 nodes)1.0×
Docs
Heterogeneous

이기종 GPU 통합

벤더와 세대를 넘나드는 GPU 오케스트레이션으로 더 높은 처리량과 낮은 지연 시간 달성.

GPT-OSS 120B

H100 + AMD MI300X

Throughput

Cross-vendor PD disagg1.7×
Same-vendor PD disagg1.0×
Technical Report

GPT-OSS 120B

H100 + AMD MI300X

End-to-end latency (낮을수록 좋음)

Cross-vendor PD disagg0.57×
Same-vendor PD disagg1.0×
Technical Report

DeepSeek R1 671B

AMD MI300X + MI308X

Throughput

PD disaggregation1.53×
Load-balanced1.0×
Blog Post

GPT-OSS 120B

H100 + AMD MI250

Throughput

Speculative decoding1.17×
All-inference baseline1.0×
Technical Report

GPT-OSS 120B

4× AMD MI250 nodes

TTFT at 100K context (낮을수록 좋음)

Multi-node prefill engine<2s
Single-node baseline~9s
Blog Post