Engine comparison
vLLM vs SGLang: a measured engine comparison
We measured Llama 3.1 8B Instruct on NVIDIA H100 80GB and NVIDIA L40S 48GB at BF16 and FP8, using vLLM 0.23.0 and SGLang 0.5.13, as of June 20, 2026. Every claim below stays inside those conditions.
Our measurements split the lead by condition, making workload shape more useful than a universal ranking.
Head-to-head evidence
We compare measured throughput and TTFT p50 at every published concurrency, then show derived cost cells at each published configuration. Every ratio sits beside both source values.
| Metric and condition | vLLM | SGLang | Conditional verdict |
|---|---|---|---|
| Saturation throughput on NVIDIA H100 80GB at BF16NVIDIA H100 80GB, BF16, concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026output tokens per second, mean of three timed repeats after warmup, unique prompts with prefix caching off | 5,333 output tokens per second | 5,235 output tokens per second | We measured higher saturation output throughput for vLLM. vLLM recorded 5,333 output tokens per second versus SGLang at 5,235 output tokens per second, under NVIDIA H100 80GB, BF16, concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026. The render-derived ratio is 1.02x.derived at render time from the two measured throughput values at concurrency 256, output tokens per second, mean of three timed repeats after warmup, unique prompts with prefix caching off |
| TTFT p50 at concurrency 1NVIDIA H100 80GB, BF16, concurrency 1, Llama 3.1 8B Instruct, as of June 20, 2026p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column | 35 milliseconds | 41 milliseconds | We measured lower TTFT p50 for vLLM. vLLM recorded 35 milliseconds versus SGLang at 41 milliseconds, under NVIDIA H100 80GB, BF16, concurrency 1, Llama 3.1 8B Instruct, as of June 20, 2026. The render-derived ratio is 1.17x.derived at render time from the two measured latency values at concurrency 1, p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column |
| TTFT p50 at concurrency 8NVIDIA H100 80GB, BF16, concurrency 8, Llama 3.1 8B Instruct, as of June 20, 2026p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column | 177 milliseconds | 221 milliseconds | We measured lower TTFT p50 for vLLM. vLLM recorded 177 milliseconds versus SGLang at 221 milliseconds, under NVIDIA H100 80GB, BF16, concurrency 8, Llama 3.1 8B Instruct, as of June 20, 2026. The render-derived ratio is 1.25x.derived at render time from the two measured latency values at concurrency 8, p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column |
| TTFT p50 at concurrency 32NVIDIA H100 80GB, BF16, concurrency 32, Llama 3.1 8B Instruct, as of June 20, 2026p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column | 514 milliseconds | 597 milliseconds | We measured lower TTFT p50 for vLLM. vLLM recorded 514 milliseconds versus SGLang at 597 milliseconds, under NVIDIA H100 80GB, BF16, concurrency 32, Llama 3.1 8B Instruct, as of June 20, 2026. The render-derived ratio is 1.16x.derived at render time from the two measured latency values at concurrency 32, p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column |
| TTFT p50 at concurrency 64NVIDIA H100 80GB, BF16, concurrency 64, Llama 3.1 8B Instruct, as of June 20, 2026p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column | 790 milliseconds | 997 milliseconds | We measured lower TTFT p50 for vLLM. vLLM recorded 790 milliseconds versus SGLang at 997 milliseconds, under NVIDIA H100 80GB, BF16, concurrency 64, Llama 3.1 8B Instruct, as of June 20, 2026. The render-derived ratio is 1.26x.derived at render time from the two measured latency values at concurrency 64, p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column |
| TTFT p50 at concurrency 128NVIDIA H100 80GB, BF16, concurrency 128, Llama 3.1 8B Instruct, as of June 20, 2026p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column | 1,658 milliseconds | 1,761 milliseconds | We measured lower TTFT p50 for vLLM. vLLM recorded 1,658 milliseconds versus SGLang at 1,761 milliseconds, under NVIDIA H100 80GB, BF16, concurrency 128, Llama 3.1 8B Instruct, as of June 20, 2026. The render-derived ratio is 1.06x.derived at render time from the two measured latency values at concurrency 128, p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column |
| TTFT p50 at concurrency 256NVIDIA H100 80GB, BF16, concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column | 2,221 milliseconds | 2,543 milliseconds | We measured lower TTFT p50 for vLLM. vLLM recorded 2,221 milliseconds versus SGLang at 2,543 milliseconds, under NVIDIA H100 80GB, BF16, concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026. The render-derived ratio is 1.14x.derived at render time from the two measured latency values at concurrency 256, p50 time to the first streamed token, charted as the mean of three timed repeats after warmup, same request stream as the throughput column |
| USD per 1M output tokens, derived from the recorded GPU rateNVIDIA H100 80GB, BF16, source saturation concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026derived from the GPU hourly price and the measured saturation throughput for that configuration, not measured | $0.206 BF16, derived | $0.21 BF16, derived | We derived lower cost for vLLM. vLLM $0.206 versus SGLang $0.21, under NVIDIA H100 80GB, BF16, source saturation concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026. |
| USD per 1M output tokens, derived from the recorded GPU rateNVIDIA H100 80GB, FP8, source saturation concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026derived from the GPU hourly price and the measured saturation throughput for that configuration, not measured | $0.158 FP8, derived | $0.17 FP8, derived | We derived lower cost for vLLM. vLLM $0.158 versus SGLang $0.17, under NVIDIA H100 80GB, FP8, source saturation concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026. |
| USD per 1M output tokens, derived from the recorded GPU rateNVIDIA L40S 48GB, BF16, source saturation concurrency 128, Llama 3.1 8B Instruct, as of June 20, 2026derived from the GPU hourly price and the measured saturation throughput for that configuration, not measured | $0.425 BF16, derived | $0.438 BF16, derived | We derived lower cost for vLLM. vLLM $0.425 versus SGLang $0.438, under NVIDIA L40S 48GB, BF16, source saturation concurrency 128, Llama 3.1 8B Instruct, as of June 20, 2026. |
Throughput at saturation
We keep measured throughput on its own scale. Derived cost never shares this chart.
vLLM5,333 output tokens per second
SGLang5,235 output tokens per second
Prefix caching depends on workload
We separate one shared prefix from many distinct prefixes because the source scopes those workloads differently.
One shared prefix
A single shared prefix is the easy case that any block-level cache handles well. It is not the workload SGLang's RadixAttention is built for.
We measured higher cache-on throughput for vLLM. vLLM, cache on, 90 percent hit rate recorded 2,855 output tokens per second versus vLLM, cache on, zero hit rate at 999 output tokens per second, under One shared prefix of 2,048 tokens sent to a varied fraction of requests, from 0 to 90 percent, with the engine's prefix cache on and off, at concurrency 32. output tokens per second, mean of three timed repeats after warmup, cache state per row. As of June 20, 2026.. The render-derived ratio is 2.86x.derived at render time for vLLM from its own cache-on throughput at a 90 percent hit rate against its cache-on throughput at a zero hit rate, output tokens per second, mean of three timed repeats after warmup, cache state per row
We measured higher cache-on throughput for SGLang. SGLang, cache on, 90 percent hit rate recorded 2,488 output tokens per second versus SGLang, cache on, zero hit rate at 937 output tokens per second, under One shared prefix of 2,048 tokens sent to a varied fraction of requests, from 0 to 90 percent, with the engine's prefix cache on and off, at concurrency 32. output tokens per second, mean of three timed repeats after warmup, cache state per row. As of June 20, 2026.. The render-derived ratio is 2.66x.derived at render time for SGLang from its own cache-on throughput at a 90 percent hit rate against its cache-on throughput at a zero hit rate, output tokens per second, mean of three timed repeats after warmup, cache state per row
Many distinct prefixes
RadixAttention may pull ahead with longer prefixes, deeper trees, or heavier eviction pressure than we tested. On our test it did not, and we are not going to claim otherwise.
We measured higher measured throughput on the many-prefix workload. vLLM recorded 2,677 output tokens per second versus SGLang at 2,493 output tokens per second, under 256 requests spread across a growing number of distinct 2,048-token prefixes, 8 then 32 then 128 of them, both engines with caching on, at concurrency 32. output tokens per second, mean of three timed repeats after warmup, both engines with prefix caching on. Distinct prefixes: 8. NVIDIA H100 80GB. Precision was not published. As of June 20, 2026.. The render-derived ratio is 1.07x.derived at render time at 8 distinct prefixes, output tokens per second, mean of three timed repeats after warmup, both engines with prefix caching on
We measured higher measured throughput on the many-prefix workload. vLLM recorded 2,277 output tokens per second versus SGLang at 2,086 output tokens per second, under 256 requests spread across a growing number of distinct 2,048-token prefixes, 8 then 32 then 128 of them, both engines with caching on, at concurrency 32. output tokens per second, mean of three timed repeats after warmup, both engines with prefix caching on. Distinct prefixes: 32. NVIDIA H100 80GB. Precision was not published. As of June 20, 2026.. The render-derived ratio is 1.09x.derived at render time at 32 distinct prefixes, output tokens per second, mean of three timed repeats after warmup, both engines with prefix caching on
We measured higher measured throughput on the many-prefix workload. vLLM recorded 1,467 output tokens per second versus SGLang at 1,376 output tokens per second, under 256 requests spread across a growing number of distinct 2,048-token prefixes, 8 then 32 then 128 of them, both engines with caching on, at concurrency 32. output tokens per second, mean of three timed repeats after warmup, both engines with prefix caching on. Distinct prefixes: 128. NVIDIA H100 80GB. Precision was not published. As of June 20, 2026.. The render-derived ratio is 1.07x.derived at render time at 128 distinct prefixes, output tokens per second, mean of three timed repeats after warmup, both engines with prefix caching on
The limitations bound every result
- We benchmarked TensorRT-LLM on its PyTorch backend, which has no ahead-of-time engine compile, so its peak throughput here is not the ceiling a compiled TensorRT engine would reach. And our single-prefix cache test is not the many-distinct-prefix case SGLang's RadixAttention is designed for, so we ran that separately too. We did not let either feed the headline.
- One H100 80GB and one L40S 48GB, single GPU, tensor-parallel size 1.
- Same Llama-3.1-8B-Instruct weights and same request stream across engines. Three timed repeats; the 95 percent confidence intervals are tight and live in the committed raw data.
- Latency here is TTFT and per-request percentiles, not goodput at a fixed SLO. Whether 514 ms first-token at high load is acceptable depends on your use case; an SLO-goodput sweep is future work.
- Numbers are self-reported on our harness. We do not cross-calibrate against an external suite like MLPerf, so the open harness is the check: re-run it and compare.
- TensorRT-LLM ran the PyTorch backend (no engine compile), so there is no compile time to report and its peak number may differ from a built TensorRT engine.
- So the 4,813 tok/s we measured is the PyTorch-backend peak, not TensorRT-LLM's ceiling, and we got no compile-time number.
- The L40S sweep stopped at concurrency 128 (its saturation), the H100 at 256.
- Engines move weekly. These numbers are vLLM 0.23.0, SGLang 0.5.13, TensorRT-LLM 1.2.1, as of June 20 2026.
- A single shared prefix is the easy case that any block-level cache handles well. It is not the workload SGLang's RadixAttention is built for.
- RadixAttention may pull ahead with longer prefixes, deeper trees, or heavier eviction pressure than we tested. On our test it did not, and we are not going to claim otherwise.
What is not measured
We did not measure models beyond Llama 3.1 8B Instruct, GPUs beyond NVIDIA H100 80GB and NVIDIA L40S 48GB, or engine versions newer than vLLM 0.23.0 and SGLang 0.5.13 on the June 20, 2026 as-of date.
- No fp8 throughput and no fp8 latency. The source publishes fp8 only as a derived cost per 1M output tokens at saturation, so no fp8 tokens-per-second, first-token latency, or speedup may be rendered from this dataset.
- No L40S throughput and no L40S latency. The L40S appears only as a derived cost cell. Every sweep row here is the H100 at bf16.
- No TensorRT-LLM prefix-cache result. Both prefix experiments ran vLLM and SGLang only.
- No compiled TensorRT engine. Every TensorRT-LLM number here is its PyTorch backend, which the source states is not TensorRT-LLM's throughput ceiling, and no engine compile time was recorded.
- No prefix-cache latency values. The source states one engine had lower first-token latency on the many-prefix workload but publishes no latency numbers for either prefix experiment, so that comparison stays qualitative.
- No hardware or precision label for the single-shared-prefix sweep. Its caption names only the concurrency, so those rows must not be presented under a GPU or a precision the source does not state.
- No goodput at a fixed service-level objective, and no accuracy or output-quality comparison between engines. This dataset is throughput, first-token latency, and derived cost only.
- No cross-engine or cross-precision ratio is stored. Pages derive every multiplier and percentage from the absolute pair at render time and print both absolute values beside it.
- Not cross-calibrated against an external benchmark suite. These are self-reported numbers from an open harness, and the harness is the check.
- No claim that these results still hold. The dataset is pinned to one as-of date and one engine version triple, and engines change on a weekly cadence.
Questions answered from the measurement
- Which engine had higher measured saturation throughput, vLLM or SGLang?
- We measured higher saturation throughput for vLLM. vLLM recorded 5,333 output tokens per second; SGLang recorded 5,235 output tokens per second. Condition: NVIDIA H100 80GB, BF16, concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026.
- Which engine had lower measured TTFT p50 at the NVIDIA H100 80GB saturation point?
- We measured lower TTFT p50 for vLLM. vLLM recorded 2,221 milliseconds; SGLang recorded 2,543 milliseconds. Condition: NVIDIA H100 80GB, BF16, concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026.
- Which engine had lower derived cost per output token in the published NVIDIA H100 80GB BF16 configuration?
- We derived lower cost for vLLM. vLLM was $0.206; SGLang was $0.21 USD per 1M output tokens. Condition: NVIDIA H100 80GB, BF16, source saturation concurrency 256, Llama 3.1 8B Instruct, as of June 20, 2026. Both values are derived from the recorded GPU rate and measured saturation throughput.
If you need custom optimization for a specific model, describe what you need
End-to-end encryption
Isolated GPU infrastructure
No training on your data
SOC 2 Type II
RunInfraby RightNow© 2026 RunInfra. All rights reserved.
Ask AI about RunInfra
