AMD Radeon RX 6950 XT 16GB

The Radeon RX 6950 XT is the top of AMD's RDNA 2 desktop stack, a mid-2022 refresh of the RX 6900 XT on the same 80 compute unit Navi 21 die with 18 Gbps memory in place of 16 Gbps and a higher power budget. That lifts peak FP32 from 23.04 to 23.65 TFLOPS and memory bandwidth from 512 to 576 GB/s, at 335 W rather than 300 W. For AI work the important structural fact is what RDNA 2 does not have: there are no matrix cores anywhere in the generation, so the 47.31 TFLOPS AMD quotes for FP16 is simply packed vector math running at twice the FP32 rate, and there is no BF16, FP8, INT8 or INT4 throughput figure to quote at all. A 16 GB frame buffer at this price still makes the card a reasonable local inference box, but the throughput ceiling is a vector ceiling, roughly an order of magnitude below what a same-era tensor-core part reaches on low precision.