Nearly 5 the output from the same eight B200s

4.94×.

That's how much more output our inference engine pulled out of eight B200s than vLLM did. Same model. Same box. Same 30 minutes.

No new hardware. No smaller model. Just a different engine.

What we ran

• Model: GLM-5.3, full BF16
• Hardware: 8× NVIDIA B200
• Baseline: vLLM 0.32
• Load: 32 workflows in flight, 30-minute window, 32K context

Real workflows from our GLM-5.3 benchmark set. Not a synthetic loop.

Both engines got the same first 116 workflows. After that, whoever freed up capacity first got to take more.

Sustained load

This is the test that matters if you run GPUs for a living. Traffic doesn't stop. Requests keep coming.

Metric	Backboard vs vLLM
Output throughput	4.94×
Workflows admitted	3.98×
Median time to first token	110× faster
p90 time to first token	8.25× faster
Per-request generation speed	2.05×

The 110× is the one I keep coming back to.

Time to first token is how long a user, or an agent, sits there waiting before anything happens. Under sustained load that's where most engines fall apart. Requests pile up in the queue. Everyone waits.

We cleared work fast enough to keep the queue moving. So we kept admitting new workflows. Nearly four times as many over the same window.

Burst load

Different test. Drop 96 workflows on the engine all at once and see what happens.

Metric	Backboard vs vLLM
Output throughput	2.61×
Per-request generation speed	3.00×
Median time to first token	10× faster
p90 time to first token	8.25× faster

We finished 95 of the 96. vLLM finished 82.

Why this matters

Everyone is short on GPUs. You can't buy your way out of that fast.

What you can do is get more out of the ones you already have. Nearly 5× the output from the same eight cards is capacity you already paid for. It was just sitting there.

That's the whole idea behind the engine. More useful work per GPU. Better experience for whoever is waiting on the answer.

And it's the full model doing the work. BF16 weights. No quantized serving model. No relaxed token acceptance.

Read the full report

Setup and methodology are in the GLM-5.3 report on our blog.

If you run your own GPUs and want to see what this does on your workload, talk to us. Bring your best-tuned setup. We like a fair fight.