FPGA corner: when Kria, Versal or MPSoC beats a GPU or NPU

In our like-for-like groups, an AMD FPGA/adaptive SoC has the best perf per watt in 1 of 3 groups where both sides publish power, and the best raw throughput in 1 of 4 groups.

Scope: Kria (K26), Zynq UltraScale+ MPSoC (ZCU102/ZCU104), Versal AI Core/Edge (VCK190, VEK280) and Alveo V70, against every GPU/NPU/CPU in the same model-size class and batch. FPGA figures are mostly AMD Vitis AI Model Zoo tables (grade D, 2023) and mostly image classification. Power is the vendor-stated figure unless marked measured.

What the data says

FPGA leads on perf/W in: image classification (mobile-class CNN, batch not stated).
GPU/NPU leads on perf/W in: object detection (mid-size CNN, batch not stated); image classification (mid-size CNN, batch not stated).
Groups where no power figure exists on one side are listed below but not counted.

Expert notes — Michael Pechorsky

Empty slot. Michael's notes on when an FPGA is the right call (latency determinism, custom pre/post-processing, I/O, long product life, DPU configuration) go here. Nothing has been written yet.

No licensed photo available — placeholder drawing.

Show data — 4 like-for-like groups with FPGA and non-FPGA results

FPGA versus best GPU/NPU per like-for-like group
Workload · group	Fastest FPGA	Fastest GPU/NPU/CPU	Best FPGA perf/W	Best other perf/W
object detection mid-size CNN · batch not stated	AMD Alveo V70 (Versal AI Core, XDNA AIE-ML) 70.16 FPS D Older data (2023)	Intel Xeon Platinum 6972P 969 FPS D undated	AMD Alveo V70 (Versal AI Core, XDNA AIE-ML) 0.935 FPS/W D Older data (2023)spec	NVIDIA Jetson AGX Orin (MaxQ) 3.35 FPS/W D 2026-10measured
object detection unknown · batch not stated	AMD Alveo V70 (Versal AI Core, XDNA AIE-ML) 2174 FPS D Older data (2023)	Intel Xeon Platinum 6972P 5703 FPS D undated	AMD Kria KV260 Vision AI Starter Kit (K26 SOM) 36.76 FPS/W D Older data (2023)spec	n/a
image classification mid-size CNN · batch not stated	AMD Alveo V70 (Versal AI Core, XDNA AIE-ML) 7780 FPS D Older data (2023)	HPE ProLiant Compute DL380 Gen12 (5x NVIDIA RTX PRO 4500 Blackwell) 94651 FPS A undated	AMD Alveo V70 (Versal AI Core, XDNA AIE-ML) 104 FPS/W D Older data (2023)spec	NVIDIA Jetson AGX Orin (MaxQ) 150 FPS/W D 2026-10measured
image classification mobile-class CNN · batch not stated	AMD Alveo V70 (Versal AI Core, XDNA AIE-ML) 63109 FPS D Older data (2023)	Intel Xeon Platinum 8580 39672 FPS D undated	AMD Alveo V70 (Versal AI Core, XDNA AIE-ML) 841 FPS/W D Older data (2023)spec	TRIGKEY Mini PC (AMD Ryzen 5 5560U, CPU) 91.08 FPS/W B 2025 (year only)measured

See also the FPGA / Versal coverage overview.

How to read these numbers

Every number is labelled. Confidence: A MLPerf closed, audited · B documented third-party method · C community / review · D vendor-published, not independently verified · E estimate. Power basis: measured with the result · spec vendor spec/envelope · power n/a not published.

Machine-readable: /find.json (see llms.txt); full tables are in “Show data”. Confidence grades: A audited MLPerf closed · B documented third-party · C community · D vendor-published · E estimate. Unknown values are shown as unknown, never estimated. Generated 2026-10-16 from public sources.