Skip to the result

ResearchTechnical note 13

One CPU core against a high-end GPU

The same task was served by a neural network on a current high-end GPU and by METRON’s compiled laws on a single CPU core. The CPU core served about 27 times more requests per second, and made no errors.

Measured
13 November 2026
METRON
One CPU core
Neural network
RTX 5090
Record
§195

At a glance

27×

more requests per second on one CPU core than the neural network on the GPU.

23.7µs

median time to answer a single request.

1.73M/s

requests served per second by one core.

In plain words

We raced two ways of doing the same job. On one side, a neural network on one of the fastest graphics cards you can buy. On the other, our system on a single core of an ordinary processor. Ours answered many times more questions per second and got every one right.

For engineers

METRON’s compiled laws on one core against an 871k-parameter network on an RTX 5090 with CUDA graphs, the same task, 5 repetitions per point across batch sizes. Single request: p50 23.7 µs, p99 30.8 µs. Peak throughput 1.73 M/s at batch 4,096 against 63.5 k/s at batch 64.

Why it matters

GPUs are the scarce, expensive resource of the AI industry. Work that moves from a GPU to a CPU core, and gets faster and exact in the process, is capacity returned and cost removed.

MTRMETRON is AIDON’s exact co-processor. It learns laws from data, proves them and runs them, alongside a neural model or on its own.

The bottleneck

AI inference is priced in GPUs.

Serving a neural model takes an accelerator, and accelerators are the bottleneck of the industry: expensive, power-hungry and in short supply.

Even a small network pays a fixed overhead on every call, and what it returns is an approximation.

Result

More requests, on less hardware, with no errors.

Requests served per second, best batch size for each

METRON, one CPU core1.73 million
Neural network, RTX 509063.5 thousand

Fig. 1 The same task served two ways. One core of a desktop CPU running METRON’s compiled laws serves about 27 times more requests per second than the neural network on a current high-end GPU, and its answers are exact.

Table 1 AIDON on one CPU core, by batch size.
Requests per batchMedian time per batchRequests per second
123.7 µs41,319
830.7 µs256,618
6466.8 µs947,108
512355.0 µs1,434,547
4,0962,350.4 µs1,730,887

Accuracy on the task: METRON 100 %, the neural network 99.57 %.

Method

How the comparison was run

  1. Same taskidentical requests for both systems
  2. METRONcompiled laws on one CPU core
  3. Neural networkRTX 5090, best configuration measured
  1. Both systems serve the same requests, across batch sizes from 1 to several thousand.

  2. Each point is repeated five times and the median is reported.

  3. Each system is credited with its best batch size. The GPU is not handicapped.

Controls

What was checked

Best case for the GPU

The network was measured with CUDA graphs and at every batch size; the figure used is its best.

Thresholds set in advance

The pass marks were written down before the measurement was run.

Repetitions

Five repetitions per point, 100 of 100 points completed.

Accuracy alongside speed

Speed was recorded together with accuracy: 100 % for METRON, 99.57 % for the network.

Open questions

What this opens

  • The same comparison with the laws on a dedicated chip.
  • Larger tasks, and networks served with further compiler optimisations.

Fields of application

Where it fits

  • Data centres, where every request moved off a GPU frees scarce capacity.
  • Edge devices that have a CPU and no accelerator.
  • High-frequency workloads with a latency budget in microseconds.

Record

How this result is kept

Record
§195, measured
Run
13 November 2026, 1 h 17 min
Sample
100 measurement points, 5 repetitions each
Measurement file
SHA-256 dbe519373c394742a2938f6c4b0743730bc7cfd56f5690d9c9fefa0b8fce7e26

Every figure on this page was read by a program from the raw measurement file, not typed. The fingerprint identifies that file: when it is disclosed, anyone can check that it is the same one.

Contact

Talk to the founders

Get in touch. We would be glad to share our ideas and walk you through the results, the proof and the measurements.