English edition
LeCompute is a French publication that dissects AI compute infrastructure: GPUs and memory, inference runtimes like vLLM and llama.cpp, quantization, and the real economics of serving LLMs. A selection of our deep dives is available in English below. The full catalogue lives on the French site.
8 articles in English
Runtimes
On July 8, 2026, OpenAI disavowed SWE-Bench Pro: ~30% broken tasks. It is the third reference instrument declared dead in three generations. Anatomy of a metrology crisis: noise ceiling, contamination, and a harness effect worth a whole model generation.
Costs
On July 8, 2026, OpenAI disavowed the reference benchmark for agentic coding. On the 9th, it merged Codex into the ChatGPT app and launched GPT-5.6. Behind the sequence, a comeback that turns less on intelligence than on prices, subscriptions and the harness.
Silicon
The DGX Spark and the RTX 5090 answer two opposite questions: hosting a very large model, or serving a mid-size one fast. The duel in numbers, measurements and mechanisms included.
Runtimes
NVIDIA Dynamo 1.0, SGLang 0.5.12, TensorRT-LLM 1.3, Tenstorrent Blackhole: prefill/decode disaggregation reaches production. Implementations, benchmarks, and counter-arguments.
Runtimes
The KV cache is no longer a VRAM byproduct you just absorb: it has its own quantization (TurboQuant, ~3 bits), its transfer protocol, its tiered storage (KVBM G1→G4) and its own routing. Anatomy of that shift.
Silicon
How much VRAM does a local LLM need? The ~2 GB per billion parameters rule, the weight of the KV cache, what quantization changes, and what actually fits on your card.
Silicon
The RTX Spark (N1X) is not a rebadged DGX Spark: it is the Grace Blackwell SoC for Windows on Arm, with 128 GB of unified memory. Why the ~273 GB/s, not the "petaflop", decide what it can actually do.
Silicon
TorchTPU, PyTorch/XLA, JAX, XLA and vLLM form a stack whose keystone is a compiler. How Google attacks the one asset that keeps developers on NVIDIA: the cost of leaving CUDA.
