Chips and Cheese measures Adreno X2 and finds no hardware FMA and 1MB of 21MB usable for compute

Sources: Chips and Cheese analysis, HN discussion
Summary: The measurements cover the Adreno X2 GPU in Qualcomm's Snapdragon X2 Elite. There is no native fused multiply-add, so OpenCL fma() collapses in throughput where AMD, Intel and Nvidia GPUs pay nothing for it. Only about 1MB of the 21MB High Performance Memory can be allocated as local memory, because that block is sized to hold a whole screen as a single rasterization tile rather than to serve compute.
Why it matters: Those two measured limits decide whether existing GPU compute code ports to the Snapdragon X2 Elite without being rewritten.