>The method, called KV Cache Transform Coding (KVTC), applies ideas from media compression formats like JPEG to shrink the key-value cache behind multi-turn AI systems, lowering GPU memory demands and speeding up time-to-first-token by up to 8x.
Google AI introduces STATIC, a sparse matrix framework that accelerates constrained decoding for LLM-based generative retrieval. It addresses the inefficiency of traditional trie implementations on hardware accelerators by flattening the trie into a static Compressed Sparse Row (CSR) matrix, achieving up to 948x speedup and demonstrating improvements in YouTube video recommendations.
A user is experiencing slow performance with Qwen3-Coder-Next on their local system despite having a capable setup. They are using a tensor-split configuration with two GPUs (RTX 5060 Ti and RTX 3060) and are seeing speeds between 2-15 tokens/second, with high swap usage. The post details their hardware, parameters used, and seeks advice on troubleshooting the issue.
The RTX 3090 offers a compelling combination of performance and 24GB of VRAM, making it a better choice for local LLM and AI workloads than newer Nvidia Blackwell GPUs like the RTX 5070 and even the RTX 5080, due to VRAM limitations and pricing.
A user shares their optimal settings for running the gpt-oss-120b model on a system with dual RTX 3090 GPUs and 128GB of RAM, aiming for a balance between performance and quality.
M5 delivers over 4x the peak GPU compute performance for AI compared to M4, featuring a next-generation GPU with a Neural Accelerator in each core, a more powerful CPU, a faster Neural Engine, and higher unified memory bandwidth.
Nvidia's DGX Spark is a relatively affordable AI workstation that prioritizes capacity over raw speed, enabling it to run models that consumer GPUs cannot. It features 128GB of memory and is based on the Blackwell architecture.
Nvidia introduces the Rubin CPX GPU, designed to accelerate AI inference by decoupling the context and generation phases. It utilizes GDDR7 memory for lower cost and power consumption, aiming to redefine AI infrastructure.
Nvidia has expanded its Jetson lineup with the Jetson AGX Thor Developer Kit, a compact platform that carries the new Jetson T5000 system-on-module. Marketed as a developer system, the dimensions and form factor place it firmly in the realm of a mini PC, although its design and purpose align more with edge AI deployment than home computing.
A 120 billion parameter OpenAI model can now run on consumer hardware thanks to the Mixture of Experts (MoE) technique, which significantly reduces memory requirements and allows processing on CPUs while offloading key parts to modest GPUs.
