ATSInfer Optimizes llama.cpp: 3x Faster Decoding for 10B+ Models on 24G VRAM
TeksEdge · x · 2026-07-19
Introduces ATSInfer, a new technology that significantly accelerates local inference speeds for large models exceeding VRAM capacity by extending llama.cpp.
- Core Mechanism: Instead of crudely placing entire network layers or MoE experts on either the GPU or CPU, ATSInfer schedules at the tensor level. It evaluates each tensor's performance on the CPU and GPU, retains the most cost-effective tensors in VRAM, and dynamically overlaps computation with PCIe transfers.
- Tested Performance: Compared to the original llama.cpp under the same VRAM budget, prefill speed increased by up to 1.94x, and overall decoding speed increased by up to 3.29x (3.12x on RTX 4090), with average GPU SM utilization rising by about 70%.
- Test Hardware & Models: Successfully and efficiently ran large models like Llama 3.1-70B and Qwen3.5-122B on RTX 3060 6GB and RTX 4090 24GB systems.
- Local Deployment Impact: This means that through smarter software scheduling, everyday developers can use cheap system memory combined with limited VRAM to run massive models far exceeding their VRAM capacity. The code is not yet publicly available.
Related event: ATSInfer Boosts Local LLM Inference Speed(2 posts)→
More from Infra
- A new series tests which data-science workflows can run on GPUs today — pandeyparul · 2026-07-21
- Former AWS operator says Bedrock margins can beat SageMaker as agentic AI lifts CPU demand — RihardJarc · 2026-07-21
- Engram shows how agent memory can keep, rewrite, or delete facts asynchronously — philipvollet · 2026-07-21
- Lightning AI’s LitLogger captures training metrics, artifacts, commands, and environment data — LightningAI · 2026-07-21
- Moonshot pauses Kimi K3 signups five days after launch as GPU demand surges — eyishazyer · 2026-07-21
- Microsoft expands Mistral models across Azure, Foundry, Copilot Studio and Azure Local — arthurmensch · 2026-07-21