Tag: #unified-memory12 results
A user completed a 16x DGX Spark cluster hitting line rate on a 200Gbps fabric switch. Setup involved racking, scripting SSH/jumbo frames/IPs, smoother than expected. Plans test 434GB models like GLM-5.1-NVFP4 and future prefill/decode split with M5 Ultra.
Ollama now supports MLX, delivering faster performance for running local AI models on Apple Silicon Macs. The update optimizes unified memory usage for significant speedups in inference.
NVIDIA's DGX Spark now supports clustering up to 4 units, creating a compact desktop data center. A single unit delivers 1 petaFLOP AI compute with 128GB LPDDR5X memory; four units scale to 4 petaFLOPS and 512GB.
Perplexity chose Mac, particularly Mac mini, as the optimal platform for local deployment of its agentic AI Personal Computer. Apple CFO Kevan Parekh noted in earnings call that more AI platforms use Mac as base, spotlighting Perplexity's use of Apple Silicon and unified memory.
Apple removed 256GB RAM config for M3 Ultra Mac Studio; 512GB gone earlier. Now max 96GB RAM amid supply constraints. Hurts local large model runners.
Second-hand M3 Ultra Mac Studio prices have skyrocketed, with 512GB unified memory versions reaching $25,000 (170,000 RMB). This far exceeds official pricing, such as $12,000 for a 256GB model with 16TB SSD. The secondary market is out of control due to surging demand.
A user unboxes NVIDIA DGX Spark for on-premises LLM inference using vLLM, PyTorch, and Hugging Face models in an education app. They seek advice on optimal models, vLLM tuning for unified memory, and real-world throughput. This marks a shift from cloud GPUs to local setups.
The article offers an in-depth observation of China's compute-storage integration efforts. It examines companies determinedly advancing this architecture amid competitive pressures. Key players and strategies in this space are highlighted.
A Reddit user claims a former employer plans to destroy more than 100 Apple M4 MacBook Pro units after closing an office. The machines reportedly include 48GB of unified memory and could be securely erased and redeployed instead.