inference/research
verifiedhigh · 0.95updated 2026-10-04

CoreWeave GB300 NVL72, MLPerf Inference v6.1 Offline: Llama 2 70B across a full 72-GPU rack (18 nodes), the largest single v6.1 LLM result

1136100 output tokens/s (aggregate across all 72 GPUs; harness prints 1.1361e+06 in scientific notation)

Fields

id
benchmarks/bench-mlperf-v6-1-coreweave-gb300-nvl72-llama2-70b-offline
Metric
tps_aggregate
Value
1,136,100
Unit
output tokens/s (aggregate across all 72 GPUs; harness prints 1.1361e+06 in scientific notation)
Model
Llama 2 70B (MLPerf v6.1 submission, FP4 weights)
Methodology
AGGREGATE. MLPerf Inference v6.1 Closed division, Datacenter suite, Offline scenario at RACK scale, GB300 generation. LOADS: batch size not a fixed published number; min_duration 600000 ms, min_query_count 2000, early stopping satisfied: Yes. Model shape TOKENS_PER_SAMPLE 267.0. Hardware: 72x NVIDIA GB300-288GB aarch64, 18 nodes, 4 per node. Framework per summary.csv: 'TensorRT-LLM feat/1.2-mlpinf, NVIDIA Dynamo mlperf-v6.0-dynamo-v0.8.0, vLLM CentML:mlperf-inf-mm-q3vl-v6.0'. Samples per second 4160.83. POWER: has_power=False, design_power_watts=null, no declared cap, no cooling configuration. OMITTED BY SOURCE: replica/TP layout, fabric, KV dtype, concurrency, clock state.
Measured by
vendor
Reproducible
yes
notes
THE BIGGEST LLAMA NUMBER IN THE CURRENT ROUND, and the number most likely to be quoted without its qualifiers. Three of them are load-bearing. (1) It is 72 GPUs, not a node: 15779 tok/s per GPU, which is LOWER than an 8-GPU B300 node's 13660/GPU... no - it is higher, but only by 15%, while using nine times the silicon. (2) No watts: has_power is false, so this figure cannot support any cost-per-token claim, and a 72-GPU GB300 NVL72 rack is a very different power envelope from an 8-GPU node. (3) The accuracy target is the -99.9 track, so this is a 99.9%-quality Llama 2 70B, not a vanilla one. NVIDIA's own 8-GPU slice of the very same GB300 NVL72 reaches 131464 tok/s, which is the number that actually tells you what to buy. INTERCONNECT IDS FILLED 2026-10-04: interconnect_ids was empty; it now carries the interconnect list of the accelerator record(s) this run names (nvidia-b300). BASIS AND ITS LIMITS, stated because this field is easy to over-read: it records the FABRIC PRESENT IN THE HARDWARE CONFIGURATION, which is how every previously-filled record in this repo reads it, not traffic measured over that fabric. The clearest evidence for that convention is that single-GPU records already carry a scale-up fabric here - trtllm-llama31-8b-fp8-h100-1k1k-tps is 1x H100 at TP=1 and carries ['nvlink-4','nvswitch-4'] - and that host-link-only records carry a PCIe value (arc-pro-b70-mlperf-v6-1-llama2-70b-offline-tokens carries ['pcie-gen5'] on 4x Arc Pro B70, which has no GPU-to-GPU fabric at all). So a value here does NOT mean the interconnect was exercised by this particular workload, and must not be cited as evidence that it was. Whether the run actually crossed that fabric is a question about TP/EP/PP size, which this repo records in methodology, not here. NOT INVENTED: nothing was filled where the accelerator record itself carries no interconnect, so this field stays empty for any part whose own record has none - which is the honest state for a single-socket Apple sOC and for parts whose vendor publishes no host interface.

Notes

THE BIGGEST LLAMA NUMBER IN THE CURRENT ROUND, and the number most likely to be quoted without its qualifiers. Three of them are load-bearing. (1) It is 72 GPUs, not a node: 15779 tok/s per GPU, which is LOWER than an 8-GPU B300 node's 13660/GPU... no - it is higher, but only by 15%, while using nine times the silicon. (2) No watts: has_power is false, so this figure cannot support any cost-per-token claim, and a 72-GPU GB300 NVL72 rack is a very different power envelope from an 8-GPU node. (3) The accuracy target is the -99.9 track, so this is a 99.9%-quality Llama 2 70B, not a vanilla one. NVIDIA's own 8-GPU slice of the very same GB300 NVL72 reaches 131464 tok/s, which is the number that actually tells you what to buy. INTERCONNECT IDS FILLED 2026-10-04: interconnect_ids was empty; it now carries the interconnect list of the accelerator record(s) this run names (nvidia-b300). BASIS AND ITS LIMITS, stated because this field is easy to over-read: it records the FABRIC PRESENT IN THE HARDWARE CONFIGURATION, which is how every previously-filled record in this repo reads it, not traffic measured over that fabric. The clearest evidence for that convention is that single-GPU records already carry a scale-up fabric here - trtllm-llama31-8b-fp8-h100-1k1k-tps is 1x H100 at TP=1 and carries ['nvlink-4','nvswitch-4'] - and that host-link-only records carry a PCIe value (arc-pro-b70-mlperf-v6-1-llama2-70b-offline-tokens carries ['pcie-gen5'] on 4x Arc Pro B70, which has no GPU-to-GPU fabric at all). So a value here does NOT mean the interconnect was exercised by this particular workload, and must not be cited as evidence that it was. Whether the run actually crossed that fabric is a question about TP/EP/PP size, which this repo records in methodology, not here. NOT INVENTED: nothing was filled where the accelerator record itself carries no interconnect, so this field stays empty for any part whose own record has none - which is the honest state for a single-socket Apple sOC and for parts whose vendor publishes no host interface.

Cross-references

Accelerators (1)

Engine (1)

Format (1)

Interconnect (2)