Advancing the Frontier of AI Inference
An inference engine built to make your infrastructure go further. Our own kernels. Our own execution architecture.
Backboard0.299 s
vLLM13.275 s
Backboard547.3 ms
SGLang764.1 ms
Backboard61.2 s
vLLM504.9 s
Economic impact
Lower GPU cost per output token
≈
31
%
Qwen3.8-27B
≈
54
%
Gemma 4 12B
≈
80
%
GLM-5.3 Full
Latest results.
New benchmarks from the lab. Published with configurations, measurements, and methodology.
November 9, 2026
2.18×
output throughput vs SGLang
Gemma 4 12B at BF16 on one RTX PRO 6000 Blackwell GPU, with 1K-token prompts and 64 concurrent requests.
November 6, 2026
p90 interactivity vs SGLang FP8
Backboard at BF16 versus the published SGLang FP8 baseline on 8×H100. The largest p90 streaming-rate gain occurred at one active session.
12.8×
November 5, 2026
44.4×
44.4× faster than vLLM
faster median time to first token than vLLM
Higher output throughput than vLLM and SGLang across all eight tested combinations of input length and concurrency on one RTX PRO 6000 Blackwell GPU.
Internally measured results across distinct test configurations. See each study for methodology, baseline details, and limitations.
Make more of the infrastructure you operate.
Cloud and inference providers
Evaluate how much more demand your existing GPU fleet can serve within your response-time requirements.
Enterprise AI teams
Measure performance against your own models, applications, and workloads within your environment.
Dedicated and sovereign infrastructure
Assess additional serving capacity on the compute you own or operate, within your deployment requirements.
Bring us your workload.
The most useful benchmark is yours.
Share your models, hardware, and current serving stack. We’ll work with your team to measure the difference.
