Linus Tech Tips Discussion

l packaging, they could do weird things with tiles and split the memory bus up on chiplets with interconnect between them, so it's not going to be easy to estimate from that.   From power should be easier. This is rumored pcb gas a single 16 pin for 600W delivery. An Arc B390 has 12 Xe3LPG cores for 80W. Discounting the many differences Xe3P would have, just to get an estimate, this could have in the ballpark of 600/80*12=90 Xe3P cores. Could be a 6.5X jump. For reference Nvidia 5060 (3,840 CUDA) to 5090 (21,760 CUDA) is a 5.6X jump.   It's a really interesting rumored product.   There are many things we don't know that impact viability. Cost, compute, memory configuration and software stack (Intel is behind Nvidia's CUDA, but ahead of AMD's ROCm).   Using DDR does push the memory bus to really incredible sizes for an add in PCI-E card. In terms of €/GB it should be incredibly compelling.   This is clearly an LLM inference card thanks to it's potentially huge VRAM. Assuming it could get to 320 GB, this card should be able to run 300B class LLMs without offloading anything off it's VRAM, but what speed it can get is unknown. Diffusion performance is more compute bound, this could have lots of it so potentially it could run even video models.   I'm looking forward to this release, it's rumored to be 2026H2.   Sources https://videocardz.com/newz/intel-crescent-island-pcb-leak-shows-xe3p-gpu-and-room-for-160gb-lpddr5x-memory