ASUS's first Arc card in two generations: full Big Battlemage, 32 Xe cores, 256 XMX engines, 367 TOPS, blower-style dual-slot, 310W TBP — and it scales to four cards for 128GB of pooled VRAM (~120B-param models). Software list reads like your stack: llama.cpp via SYCL/Vulkan, Ollama, vLLM (XPU), PyTorch IPEX, OpenVINO. Estimated $1,300–2,000. Intel's discrete-GPU story is now officially a local-AI story — and two of these would be 64GB for the price of one used 4090. Tempting is a strong word. Very strong.
via alextech.ai / Wccftech · Oct 17
Xe3P architecture, 480GB of memory in a 350W PCIe form factor, aimed at LLMs up to 300B parameters on a single card. Sampling to customers now, no price/date. Read it as Intel's two-front strategy: datacenter inference silicon over here, Arc Pro for the home lab over there. 480GB on one card is the kind of number that makes the 32GB ceiling feel quaint.
via thejoai.com · Oct 14
The GB10-class box is now a shipping product: unified memory, local LLMs and agents, 8K video/3D workloads, and NVIDIA PAIR letting it serve inference to other machines on your home network. This is the "Mac Studio for Windows" moment — and the template for where home-lab inference is going: one beefy node serving the whole house. You're already living this with the :8081 pair; the industry just caught up.
via ITMatterss / NVIDIA · Oct 17
Out of the $20B Groq IP licensing deal comes the Groq 3 LPU: 128GB of on-chip SRAM, 640 TBps scale-up bandwidth, liquid-cooled LPX racks aimed at extreme-low-latency inference. NVIDIA's own framing: GPUs for flops, LPUs for bandwidth — combine both for the "multi-agent future." Trade-off is brutal economics (tokens/sec per chip is low), but for agentic latency this is the silicon to watch. Racks land alongside Vera Rubin.
via Data Center Dynamics · Oct 15
