There's a huge gap between wimpy and GPU, no? I'd guess this chip was right-sized to run inference on a 10 Gbit/s feed, optimized for perf/W.