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

NVIDIA DGX B300, MLPerf Inference v6.1 Server: Llama 2 70B on 8x B300-SXM-270GB, aggregate output throughput under the Server latency constraint

106949.36 output tokens/s (aggregate, harness 'Completed tokens per second')

Fields

id
benchmarks/bench-mlperf-v6-1-nvidia-b300-llama2-70b-server
Metric
tps_aggregate
Value
106,949.36
Unit
output tokens/s (aggregate, harness 'Completed tokens per second')
Model
Llama 2 70B (MLPerf v6.1 submission, FP4 weights)
Methodology
AGGREGATE. MLPerf Inference v6.1 Closed division, Datacenter suite, SERVER scenario, which is a materially different measurement from Offline on identical hardware. LOADS: target_qps 0.5 (llama2-70b/user.conf '*.Server.target_qps = 0.5', Poisson arrivals), min_duration 120000 ms, min_query_count 100, plus the compiled-in per-query TTFT and TPOT 99th-percentile bounds which the log reports as 'Performance constraints satisfied : Yes' and 'TTFT/TPOT Early Stopping Result: Run successful'. Batch size is not a fixed published number - the submitter's scheduler adapts it - but it is bounded by what the latency constraint permits, which is exactly why Server lands below Offline. Completed samples per second 391.79. Model shape TOKENS_PER_SAMPLE 267.0. POWER: has_power=False, design_power_watts=null, no declared cap. OMITTED BY SOURCE: the numeric TTFT/TPOT bounds applied (they are compiled into the harness for this model/scenario and are not written into user.conf or the log), TP size, KV dtype, concurrency.
Measured by
vendor
Reproducible
yes
notes
This is the number to use when someone asks 'how fast can this node serve Llama 2 70B', because it is the only one carrying a latency guarantee. The 2.2% discount versus Offline (109277) is the measured price of holding a tail-latency bound. Note the direction of the claim: this is an AGGREGATE throughput with a per-query latency bound satisfied, NOT a per-user tok/s figure - the harness does not expose a tok/s-per-user metric for the Server scenario in these logs. 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

This is the number to use when someone asks 'how fast can this node serve Llama 2 70B', because it is the only one carrying a latency guarantee. The 2.2% discount versus Offline (109277) is the measured price of holding a tail-latency bound. Note the direction of the claim: this is an AGGREGATE throughput with a per-query latency bound satisfied, NOT a per-user tok/s figure - the harness does not expose a tok/s-per-user metric for the Server scenario in these logs. 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)