TDCC / engineering note Back to home

Internal benchmark · release note

Two compact laptops. One 27B workload. 80 tokens per second.

We paired two small machines for a Qwen 3.8 27B workload and moved a roughly 50 tok/s run to 80 tok/s—without changing the model weights or the intended output behavior.

TDCC engineeringPerformance noteSeptember 2026
Diagram 01 · shared generation pathTwo devices, one stream
REQUEST Prompt in one model request DEVICE A active work DEVICE B active work QWEN 3.8 27B One workload coordinated path RESULT One stream tokens out COMPACT LAPTOPS COORDINATED GENERATION
The point is not to make two devices look like one large machine. It is to keep a single generation path moving across the compute that is already available.

The short version

More of the same hardware spent time generating.

Distributed AI is often framed as a search for bigger clusters. This benchmark started from the opposite constraint: two compact laptops, an open model workload, and no interest in changing what the model is meant to produce.

On our internal test, the unoptimized path landed near 50 tokens per second. With TDCC’s coordination layer in place, the same Qwen 3.8 27B workload reached 80 tokens per second. That is a 60% improvement in observed generation throughput.

We did not alter model weights, swap in a smaller model, or aim for a different output behavior. The work was in the path around generation: coordinating when work moves, reducing avoidable idle time, and keeping the stream moving across both devices.

Diagram 02 · observed throughput

50 → 80 tok/s, with the model’s intended behavior unchanged.

Internal benchmark result

Same two-device setup. Same Qwen 3.8 27B workload. The comparison is about keeping the generation path productive, not modifying the model itself.

0 30 60 90 TOKENS / SECOND ~50 tok/s BASELINE PATH 80 tok/s TDCC-COORDINATED PATH +60% OBSERVED UPLIFT
Measurement note: performance varies by hardware, context length, prompt, sampling settings, operating-system state, and thermal conditions. Treat these numbers as a point-in-time internal result, not a guaranteed outcome.

Diagram 03 · the high-level work

Three ways to leave less useful time on the table.

01 / COORDINATE ABPATH

Coordinate the work.

Give available devices a shared, orderly role in the generation path instead of treating them as isolated endpoints.

02 / REDUCE IDLE TIME GAPPEDACTIVE

Reduce avoidable waiting.

Keep each handoff purposeful so capable hardware spends less time waiting for the next useful unit of work.

03 / KEEP STREAMING TOKENOUTPUT

Keep generation streaming.

Once output starts, protect continuity. The aim is one coherent stream for the user, not a visible sequence of device boundaries.

What we are testing

Useful compute is not only about having more hardware. It is about helping the hardware you already have spend more time doing useful work.

What is next

A benchmark is a signal, not a finish line.

We will keep testing across more real-world conditions and turn the lessons into a simpler TDCC experience. The product goal stays clear: make distributed compute feel like one dependable place to run AI work.

Explore the TDCC App