Huawei Atlas 300I Pro inference card (24 GB, 310-series)

THE HIERARCHY POINT: recorded separately from the 310P die because the CARD is the thing with a datasheet, and the datasheet is what anyone can actually verify. Huawei enterprise documentation is freely readable here with no login and states every physical and memory parameter exactly; it simply never states a compute figure. That asymmetry is the honest shape of Ascend's inference line: you can build the roofline from the memory numbers and you cannot get the numerator. Notable: HHHL (full-height, full-height-length) card at 72 W and 280 g, PCIe Gen4 x16 down-compatible to x8/x4/x2, PCI ID 0x19E5:0xD500, and the same vendor CPU as the NPU package (8 Huawei-developed cores at 1.9 GHz) rather than an off-the-shelf Arm core. Codec capacity is a real differentiator versus peers at this power: 128-channel 1080p30 H.264/H.265 decode, 24-channel encode, 4K JPEG at 512 FPS. Splits into up to 7 virtual NPUs. DO NOT COMPARE ITS 24 GB TO THE ATLAS 300V PRO's 48 GB AS IF THEY WERE DIFFERENT CLASSES OF HARDWARE - the two cards have byte-identical memory subsystems (384-bit 4266 Mbit/s LPDDR4X, 204.8 GB/s) and differ only in capacity. That means the 300V Pro is strictly the better deal for any workload that fits, and the two should be treated as one part with two capacities. Attractively for this repo's purposes, llama.cpp has a first-class CANN backend and third parties have run Llama on this card, so it is a realistic non-CUDA inference target despite the bandwidth ceiling.