Intel Arc Pro B60 24GB

The Arc Pro B60 is Intel's local inference card: the full BMG-G21 die with 20 Xe-cores and 160 XMX matrix engines, but paired with 24 GB of GDDR6 on the same 192-bit bus the 12 GB Arc B580 uses, for 456 GB/s. Intel rates it at 12.28 TFLOPS FP32 and 197 dense INT8 TOPS inside a 120 W to 200 W envelope. Capacity per pound is the entire argument for it, and Intel makes that argument explicitly through Project Battlematrix, a Xeon workstation platform that takes eight of these cards for 192 GB of video memory and models up to about 150 billion parameters. It ships only through board partners, several of whom put two GPUs and 48 GB on a single card, so a four-slot workstation can reach the same 192 GB. Intel supplies SR-IOV and manageability features on this SKU and a Linux multi-GPU serving stack alongside it, which is the part that historically let Intel down and the part worth checking before buying.