Research and results
We publish what we measure, how we measured it and where the method stops. The reports are kept with the code they describe.
Software extension of GPU memory
V²RAM manages the residency of tensor blocks one level below the framework executor, inside the buffer object. Blocks move between the card, pinned host memory and a memory-mapped store under a budget, guided only by the sequence of accesses that the run reveals. No network, layer or training loop changes.
| Configuration | Budget | Resident | With V²RAM | Ratio |
|---|---|---|---|---|
| Inference, working set of 200 MiB | 128 MiB | 13.2 ms | 35.2 ms | 2.7× |
| Training, working set of 1,186 MiB | 128 MiB | 256.6 ms | 353.7 ms | 1.4× |
Milliseconds per step, median of three runs. The budget is set below the working set on purpose.
| Batch | Working set | Peak on the card | Outcome |
|---|---|---|---|
| Without V²RAM | 6,028 MiB | – | Did not start: host out of memory |
| 1 | 6,028 MiB | 10,748 MiB | Ran |
| 2 | 7,709 MiB | 10,170 MiB | Ran |
| 4 | 10,984 MiB | 9,568 MiB | Ran |
| 8 | 17,506 MiB | 9,178 MiB | Ran |
| 16 | 30,600 MiB (estimated) | 9,092 MiB | Stopped by the system |
A model of 270 million parameters with grouped-query attention and four experts per layer.
Detection and replacement of personal data
Four devices share the detection: scanners that verify check digits, word tables in five languages, a structural SVM tagger and a byte-level transformer of 116,791 parameters. Each value is replaced by a tag that carries its encryption under FF1 and AES-256.
Version 1.0.11 reaches an F1 of 0.946 on the test set, 0.919 with six times the noise, 0.953 on whole documents and 0.951 on places abroad; the tables give the detailed measures of version 1.0.5. On hand-written texts in the five languages, the F1 lies between 0.886 and 0.907. On sixteen files written for the test in the five languages (Word, PowerPoint, PDF, CSV, source code, images), version 1.0.11 leaves readable 2 of the 1,223 values to protect, against 109 for the previous version. On a private sample of real documents, precision is 0.859 and recall 0.912: about one value in eleven stays readable, which is why a protected document should be read again before it is shared. The detector reads about 41,000 characters per second on two cores.
| Test set | Precision | Recall | F1 |
|---|---|---|---|
| Test, five languages | 0.955 | 0.935 | 0.945 |
| Values never seen in training | 0.951 | 0.935 | 0.943 |
| Six times the noise | 0.941 | 0.897 | 0.918 |
| Whole documents | 0.961 | 0.943 | 0.952 |
| Places abroad only | 0.960 | 0.939 | 0.949 |
| By language | Precision | Recall | F1 |
|---|---|---|---|
| French | 0.975 | 0.965 | 0.970 |
| English | 0.971 | 0.964 | 0.967 |
| Italian | 0.950 | 0.926 | 0.938 |
| German | 0.949 | 0.926 | 0.937 |
| Spanish | 0.928 | 0.898 | 0.913 |
| System | F1 on the test set |
|---|---|
| Rules alone | 0.840 |
| SVM tagger alone | 0.749 |
| Byte-level transformer alone | 0.903 |
| Whole detector | 0.945 |
Inference on quantized weights
The engines of Albireo keep the weights in the memory mapping of the GGUF file and multiply them in their quantized form. Integer kernels, a cache of keys and values in 8 bits, drafted tokens and layers computed on the card are each measured against a reference.
- Keys and values in 8 bits give the same next token as half precision on the 210 positions of a text, with a mean divergence of 6.5e-5.
- Drafted tokens rewrite a function at 25.5 tokens per second against 19.7 on two cores, and keep the distribution of the sampling.
- Four answers written together take 32 to 40% less time per step than four separate ones.
- On the 24 threads of an Intel Core Ultra 9 275HX, the processor engine reads a prompt at 216 tokens per second and writes at 19.5 with Qwen3-VL 4B; with Qwen3 0.6B it reads at more than 1,300 tokens per second and writes at more than 80.
| Engine | Tokens per second |
|---|---|
| NVIDIA CUDA, GeForce RTX 5070 Ti Laptop GPU | 56.8 |
| Processor, 24 threads | 23.6 |
| Kernels | Integers | Floats |
|---|---|---|
| AVX-VNNI | 230 tokens/s | 126 tokens/s |
| AVX2 | 154 tokens/s | 125 tokens/s |
Routing of operations to the NPU and the integrated GPU
Built with the OpenVINO runtime, the library offers each tensor operation to the inference devices of the processor. The host and each device are timed on a few calls, and the fastest keeps that shape. A machine without these devices runs exactly as before.
| Configuration | Time per step |
|---|---|
| Host alone, with Intel oneMKL | 97.4 ms |
| Automatic routing | 63.3 ms |
The routing hands softmax, activations, cross entropy and convolutions to the integrated GPU, which computes them 2.2 to 6.9 times faster than the host. The NPU stays a candidate for every operation: at the sizes of this benchmark its fixed cost per call still outweighs its arithmetic, and the routing selects it as soon as a measurement finds it the fastest.
Prediction as compression
A trained network is a probability model of what comes next, and an arithmetic coder turns it into a lossless code. One example builds the whole chain: a tokenizer and a decoder trained in Dlib, weights quantized and compiled into the program, and arithmetic fixed operation by operation so that a decoder on another machine reproduces every probability.
| Archive | 117,169,624 bytes |
|---|---|
| Bits per byte | 0.937 |
| Decompressor, tokenizer and network included | 7,063,888 bytes |
| Time per direction, one core | About 25 hours |
| Compression with the kernels chosen by the processor | 24 h 48 min instead of 30 h 18 min |
Hierarchical reasoning and adaptive computation
Two example programs train hierarchical reasoning models on Sudoku-Extreme and on ARC puzzles. They bring to the library a halting head trained jointly with the network, the Adam-atan2 solver, a stablemax loss and an average of the weights, each available to the other models of the repository.
Images: self-supervision, colorization and synthesis
Our earlier work on images produced methods and models that remain published.
| Work | Result |
|---|---|
| BAREL, self-supervised learning followed by supervised metric learning | 96.1% on the LFW benchmark for face recognition, with a training set nearly 8 times smaller than that of the first ResNet model of Dlib |
| Barlow Twins on CIFAR-10 with a ResNet-50 | 88.25% test accuracy after self-supervised training |
| Colorify, high-resolution colorization | A network of 223 layers and about 41.7 million parameters trained on about 900,000 images |
| FFgan, synthetic faces | Color faces of 162 pixels from a DCGAN trained on over 150,000 faces |
Contributions to Dlib
Dlib is one of the reference open-source C++ libraries for machine learning and computer vision. Before publishing our own extensions, we contributed code and trained models to the library itself, first for faces (recognition, gender, age, synthesis), then for transformers.
In the library
Fourteen contributions were merged into Dlib between August 2024 and October 2025.
- Layers for transformers: RMS normalization, embeddings, linear, lower triangular mask, transposition, reshaping, plane-wise matrix product
- An adaptive computation time layer and a dropout layer with a rate set at compile time
- A byte-pair encoding tokenizer
- Example programs: a transformer network, then one with rotary position embeddings and a mixture of experts
- A loader for the ARC-AGI dataset
- A fix to the reorg layer, on processor and on CUDA
| Model | Task | Result |
|---|---|---|
face_recognition_densenet_model_v1 | Face recognition, DenseNet trained with the BAREL method | 96.1 % on LFW, with a training set nearly 8 times smaller than that of the first ResNet model of Dlib |
dnn_gender_classifier_v1 | Gender classification from a face | About 97.3 % on the LFW evaluation, trained on about 200,000 faces |
dnn_age_predictor_v1 | Age estimation from a face, ResNet-10 | Trained on about 110,000 labelled images |
dcgan_162x162_synth_faces | Synthetic faces, DCGAN | Colour faces of 162 pixels |
highres_colorify | Automatic colorization, ResNet with a U-Net structure | 223 layers and about 41.7 million parameters, usable with an 8 GB graphics card |
res50_self_supervised_cifar_10 | Self-supervised learning, Barlow Twins, ResNet-50 | 88.25 % test accuracy on CIFAR-10 |
resnet34_stable_imagenet_1k | Image classification, ResNet-34 | 99.98 % top-1 on the test split of Stable ImageNet-1K |
vit-s-16_stable_imagenet_1k | Image classification, Vision Transformer | 99.76 % top-1 on the test split of Stable ImageNet-1K |
The gender and age models are published by Cydral Technology under the Creative Commons Zero 1.0 licence.
Publications
- Buffer-Level Residency Management for GPU Memory Extension
- Aldric Pierrain, Cydral Technology. Research report, September 2026
- Technical application architecture of Pseudalis Community
- Aldric Pierrain, Cydral Technology. Research report, November 2026
- Onward to the Hutter Prize: a token-level transformer compresses enwik9 to 117 MB on one CPU core
- Aldric Pierrain, Cydral Technology. Research report, November 2026
- BAREL: Backpropagation And Reinforcement-based Environment for Learning
- Aldric Pierrain, Cydral Technology. Publication, April 2023
- A Small Language Model Stack, Built in C++ on One Machine
- Aldric Pierrain, Cydral Technology. Article
Measurement method
Each figure names its hardware, its model and its protocol. Results are cross-checked against reference implementations before they are stated, and the limits of a method are written in the same place as its results.
This work pursues the objectives of P16, the France 2030 programme led by Inria for sovereign AI libraries.