AEVON Technologies · French DeepTech

Faster neural-network layers, on the CPUs you already have.

PBC placement replaces the linear layers of a neural network. A standard layer costs n³ operations; PBC placement costs n². It is not a cheap copy of the exact product: it builds a space that is easier for the network to learn from. On raw, noisy signals the network ends up markedly more accurate than with exact layers.

See the measurements How to evaluate it

up to 24×faster training step, per layer

up to 49×faster inference, per layer

+18 ptsaccuracy over the best-tuned exact layer, on raw radio signals

Our product

PBC placement

The exact matrix product is the naive way to mix information in a layer: nothing in it is designed for learning. PBC placement is a different layer, built for two properties: smoothness (close inputs give close outputs, so training is easier) and injectivity (different inputs stay distinguishable). That is why it can match, and on raw signals clearly beat, the accuracy of exact layers, while the cost of each layer drops from n³ to n².

What it does

It takes the place of nn.Linear, in both training (forward and backward) and inference. Weights are stored in exactly the same format, so existing checkpoints load as they are.

How you use it

One call converts an existing PyTorch model: pbc_torch.convert(model). The rest of your code (optimizer, loss, data loading) does not change.

Where it runs

On standard x86-64 processors (AVX2), with all cores. No GPU or special hardware needed. The gain grows with the width of the layer.

Measured results

Faster — and, in our tests, more accurate.

Speed was measured on our evaluation server (AMD EPYC, 16 cores, AVX2, float32), the same machine our clients use. Accuracy does not depend on the machine. All figures are indicative: the reference is always your own model on your own data.

Speed — one layer, exact vs PBC placement

Training step = forward + backward pass. Speed-up = exact time ÷ PBC time. Best of 5 runs, version 2, 23 October 2026.

Layer (batch × in → out)	Train, exact	Train, PBC	Speed-up	Inference, exact	Inference, PBC	Speed-up
256 × 256 → 256	0.26 ms	0.54 ms	0.5×	0.08 ms	0.11 ms	0.7×
1024 × 1024 → 1024	6.68 ms	1.43 ms	4.7×	1.68 ms	0.21 ms	8×
1024 × 2048 → 2048	20.1 ms	2.45 ms	8×	6.30 ms	0.33 ms	19×
2048 × 4096 → 4096	154 ms	7.09 ms	22×	49.8 ms	1.09 ms	46×
2048 × 4096 → 11008	394 ms	16.1 ms	24×	128 ms	2.63 ms	49×

Small layers (around 256 and below) are not faster: fixed costs dominate. The gain starts around width 1024.

Accuracy — exact layers vs PBC placement

Each test is run on 5 independent seeds (data, initial weights, batch order), identical for both methods. The exact model gets its best chance: learning rate and output scale (the same lever PBC uses) are tuned for it on a separate validation set. PBC placement uses its default version-2 settings. The test set is never used for any choice.

Test (5 seeds, mean ± sd)	Best-tuned exact	PBC placement v2	Difference	Result
Radio signals (raw I/Q), PBC as input layer	36.4 % ± 0.7	54.5 % ± 1.1	+18.2 pts	better, 5/5 seeds
Radio signals (raw I/Q), full model, 3 PBC layers	36.6 % ± 0.4	45.2 % ± 2.5	+8.6 pts	better, 5/5 seeds
Smooth regression (tanh teacher), width 256	R² 0.878 ± 0.007	R² 0.908 ± 0.007	+0.029	better, 5/5 seeds
Smooth regression (tanh teacher), width 2048	R² 0.911 ± 0.003	R² 0.945 ± 0.012	+0.034	better, 5/5 seeds
Smooth regression, width 256, long training (150 epochs)	R² 0.984 ± 0.002	R² 0.956 ± 0.010	−0.028	exact better
Same long training, wider PBC layer (1024, “single wide layer” setting)*	R² 0.984 ± 0.002	R² 0.987 ± 0.002	+0.003	better, 4/5 seeds

Radio task: 8 modulations, random phase, frequency offset and noise (0–20 dB), dense network reading the raw samples; chance level 12.5 %. Controls: adding dropout (0.2 or 0.4) to the exact model does not change its accuracy, so PBC's advantage is not a simple regularisation effect; PBC predictions are identical whether samples are processed one at a time or in batches. * The wider PBC layer still costs about 6× fewer operations per sample than the exact width-256 layer. “Single wide layer” setting: pbc_placement.reglages(gain=0.05, echelle=2), learning rate 10⁻³; with it, the width-2048 regression above also reaches R² 0.985.

How to read these results

PBC placement is strongest where the data is raw and noisy, especially as the first layer of the network, but it also beats the best-tuned exact layer on smooth regressions.
One case where exact stays ahead at equal width: very long training on a perfectly smooth, noise-free function. A single wider PBC layer, still far cheaper, wins it back.
The exact baseline is a dense network on raw signals. A model with signal-specific front-ends (convolutions, Fourier features) starts from a higher baseline; your own model is the reference.
PBC placement is designed to be trained with. It prefers moderate learning rates (3·10⁻⁴ to 10⁻³); at 3·10⁻³ and above it can diverge where exact layers do not. To convert a model already trained with exact layers, a one-line setting restores version-1 behaviour.
The speed-ups are per layer. The whole model gains less, because the optimizer, loss, data loading and any layer left exact run at the same speed.
Current limits: CPU only (x86-64 with AVX2), real-valued tensors (complex signals are fed as real and imaginary parts).

Evaluation

Test it on your own model.

Industrial teams in AI and telecoms are evaluating PBC placement today. We give your team an account on a dedicated server in France, ready to use.

1 · Access

Send us the public SSH key of the person who will run the tests. We create a personal account for your team, isolated from other users.

2 · Check and convert

A self-test confirms the installation in a minute. Then convert your PyTorch model with one call and train it as usual.

3 · Measure

A benchmark script times exact and PBC layers on the same inputs, same cores. You measure accuracy on your data, in your own conditions.

Research

Exact matrix multiplication.

Alongside the product, we do fundamental research on the cost of the exact matrix product.

Experimental · not a product

A decoding algorithm with a sub-cubic operation count

On binary matrices, we have an experimental algorithm that recovers the exact product with a number of operations growing more slowly than n³, verified in tests up to n = 128.

This is a result in operation count only. The algorithm reads memory in a scattered way, so on today's processors and GPUs it is not faster in practice. It would only pay off on hardware designed specifically for it. We publish no performance claim on it.

Long-term industrial R&D

Thermodynamic engine

AEVON Technologies is developing a closed-loop, combustion-free nitrogen cycle designed to produce both useful energy and industrial cooling.

Target applications

Data centers, heavy industry and maritime infrastructure, which need both energy and cooling.

Industrial path

Patent filed in 2025, international PCT extension in preparation. The project is structured around licensing to industrial partners.

Founders

The team

AEVON Technologies is a French DeepTech company. PBC placement is our commercial product; it funds our fundamental research and the long-term engine program.

Robin Besson

Co-founder & CTO

Researcher and engineer working at the intersection of mathematics, algorithms and prototyping. After studies in mathematics and a degree in architecture (ENSAPVS), he conducted six years of independent research in applied thermodynamics with researchers from the ENS, then joined the École Polytechnique entrepreneurship program and the MATRICE incubator.

At AEVON he leads research and development: PBC placement, the research on exact matrix multiplication, and the thermodynamic engine.

Valentin Faure

Co-founder & CEO

Valentin built his early experience in real estate, then as Sales Director in the sports events industry, where he developed commercial operations, B2B partnerships and complex project execution.

At AEVON he leads strategy, financing, partnerships and the commercial deployment of PBC placement and of the engine licensing model.

Contact

Evaluate PBC placement on your model.

Tell us about your model and your layer sizes. We open an evaluation account for your team.

contact@aevontechnologies.com