VLSI KoreaEight H100 GPUs; system score is not inherited by the GH100 die.
NVIDIA · 근거VLSI MUSEUM / ENGINEERING RECORD2024
NVIDIAAI benchmark systems
system
칩의 기록은 여기서 시작합니다.
이 기록의 신뢰할 수 있는 공개 다이 사진은 아직 확인하지 못했습니다. 신원과 확인된 기술 근거는 계속 탐색할 수 있습니다.
관련 기술 탐색
사양
8 SOURCE-LINKED FIELDS| Benchmark | Manufacturer / primary sourceNVIDIA Scope: Eight-GPU benchmark system Conditions: MLPerf Inference v4.1; Closed/Data Center/Offline; Mixtral 8x7B; eight GPUs; TensorRT-LLM, extensive FP8 use. MoE 8 experts/2 active; not structured 2:4 sparsity. Submitted 4.1-0043. Result date 2024-08-28. Source date: 2024-08-28 |
|---|---|
| Workload | Manufacturer / primary sourceNVIDIA Scope: Eight-GPU benchmark system Conditions: MLPerf Inference v4.1; Closed/Data Center/Offline; Mixtral 8x7B; eight GPUs; TensorRT-LLM, extensive FP8 use. MoE 8 experts/2 active; not structured 2:4 sparsity. Submitted 4.1-0043. Result date 2024-08-28. Source date: 2024-08-28 |
| Scenario | Manufacturer / primary sourceNVIDIA Scope: Eight-GPU benchmark system Conditions: MLPerf Inference v4.1; Closed/Data Center/Offline; Mixtral 8x7B; eight GPUs; TensorRT-LLM, extensive FP8 use. MoE 8 experts/2 active; not structured 2:4 sparsity. Submitted 4.1-0043. Result date 2024-08-28. Source date: 2024-08-28 |
| Accelerator | Manufacturer / primary sourceNVIDIA Scope: Eight-GPU benchmark system Conditions: MLPerf Inference v4.1; Closed/Data Center/Offline; Mixtral 8x7B; eight GPUs; TensorRT-LLM, extensive FP8 use. MoE 8 experts/2 active; not structured 2:4 sparsity. Submitted 4.1-0043. Result date 2024-08-28. Source date: 2024-08-28 |
| Accelerator count | Manufacturer / primary sourceNVIDIA Scope: Eight-GPU benchmark system Conditions: MLPerf Inference v4.1; Closed/Data Center/Offline; Mixtral 8x7B; eight GPUs; TensorRT-LLM, extensive FP8 use. MoE 8 experts/2 active; not structured 2:4 sparsity. Submitted 4.1-0043. Result date 2024-08-28. Source date: 2024-08-28 |
| Measured throughput | Manufacturer / primary sourceNVIDIA Scope: Eight-GPU benchmark system Conditions: MLPerf Inference v4.1; Closed/Data Center/Offline; Mixtral 8x7B; eight GPUs; TensorRT-LLM, extensive FP8 use. MoE 8 experts/2 active; not structured 2:4 sparsity. Submitted 4.1-0043. Result date 2024-08-28. Source date: 2024-08-28 |
| Numerical format | Manufacturer / primary sourceNVIDIA Scope: Eight-GPU benchmark system Conditions: MLPerf Inference v4.1; Closed/Data Center/Offline; Mixtral 8x7B; eight GPUs; TensorRT-LLM, extensive FP8 use. MoE 8 experts/2 active; not structured 2:4 sparsity. Submitted 4.1-0043. Result date 2024-08-28. Source date: 2024-08-28 |
| Software | Manufacturer / primary sourceNVIDIA Scope: Eight-GPU benchmark system Conditions: MLPerf Inference v4.1; Closed/Data Center/Offline; Mixtral 8x7B; eight GPUs; TensorRT-LLM, extensive FP8 use. MoE 8 experts/2 active; not structured 2:4 sparsity. Submitted 4.1-0043. Result date 2024-08-28. Source date: 2024-08-28 |
THE VISUAL EVIDENCE
0 개의 확인된 출처시각 자료
이 칩의 근거를 더해 주세요.
이 기록의 신뢰할 수 있는 공개 시각 자료는 아직 확인하지 못했습니다. 다이 사진·패키지 이미지·논문·제조사 도해를 제안할 수 있습니다.
근거가 있는 연결
1 SOURCE-LINKED RELATIONSHIPS기록의 신원과 범위
system · NVIDIA 8 × H100 — MLPerf v4.1 Mixtral