Intel Arc Pro B50 x4, MLPerf v6.0 Offline: Llama-3.1-8B INT8 aggregate tok/s
1721.84 output tok/s AGGREGATE across 4x Arc Pro B50 (Offline; ~430 tok/s per card)
Fields
id
benchmarks/b50-4x-mlperf-v6-0-llama31-8b-offline-tokensMetric
tps_aggregate
Value
1,721.84
Unit
output tok/s AGGREGATE across 4x Arc Pro B50 (Offline; ~430 tok/s per card)
Model
Llama-3.1-8B (INT8)
Methodology
VENDOR-MEASURED (MLPerf Inference v6.0 closed submission, Intel). System 1-node-4x-BMG-B50: 4x [MAXSUN] MS-Intel Arc Pro B50 16G (Battlemage, 16 GB GDDR6, 224 GB/s, PCIe Gen5 x8 electrical) on one Intel Xeon Platinum 8480C host, Ubuntu 25.04, PyTorch vLLM, Llama-3.1-8B INT8 weights. Offline scenario. 'Tokens per second' = 1721.84 verbatim from the submission summary (run_1). AGGREGATE across all 4 accelerators, NOT per-request and NOT per card. Prompt and output lengths are the MLPerf llama3.1-8b dataset, not a fixed llama-bench length; accuracy log reports GEN_LEN 8149276 over the run (ROUGE1 38.7265).
Measured by
vendor
Reproducible
yes
notes
Batch size is the LoadGen Offline schedule, not batch 1. Dividing by 4 gives about 430 tok/s per card only as an average; the submission does not publish a per-card split. Compare with the same submitter's 4x Arc Pro B60 Offline row (3347.43 tok/s, FP8) rather than with single-card llama-bench numbers.
Notes
Batch size is the LoadGen Offline schedule, not batch 1. Dividing by 4 gives about 430 tok/s per card only as an average; the submission does not publish a per-card split. Compare with the same submitter's 4x Arc Pro B60 Offline row (3347.43 tok/s, FP8) rather than with single-card llama-bench numbers.