vkpeak (Vulkan Compute) Benchmark Results
Compare raw GPU compute throughput across Single Board Computers using vkpeak: FP32, FP16 and INT8 peak rates in GFLOPS/GIOPS
2 boards tested · scores from 200.6 to 219.6 (FP32 (GFLOPS))
This table shows the three headline rates. For the full results of every vkpeak scenario, including FP64, matrix and copy bandwidth, open the individual board's page.
Full Results
| Compare | # | Board | FP32 (GFLOPS) | FP16 (GFLOPS) | INT8 (GIOPS) |
|---|---|---|---|---|---|
| 1 | Radxa X4L 8GB LPDDR5 Radxa X4L 8GB LPDDR5 | 219.62 | 473.01 | 1,583.74 | |
| 2 | Radxa ROCK 5B+ 16GB LPDDR5 Radxa ROCK 5B+ 16GB LPDDR5 | 200.56 | 529.53 | 1,067.71 |
Frequently Asked Questions
What is the fastest single board computer in vkpeak (Vulkan Compute)?
The Radxa X4L 8GB currently tops our vkpeak (Vulkan Compute) ranking with 219.6 (FP32 (GFLOPS)), measured on real hardware with our standardised test setup.
About vkpeak (Vulkan Compute) testing
vkpeak measures what a board's GPU can do as a compute device rather than as a renderer. It drives the GPU through Vulkan compute shaders and reports the peak arithmetic rate for each data type: FP32 for general-purpose shader work, FP16 for neural-network inference, and INT8 dot products for quantised models. It also measures how fast data moves between system memory and the GPU. It runs headless with no compositor, so it isolates the shader cores and driver from display output. A score of 0 is not a failure: it means the GPU or its driver simply doesn't support that data type, which on many SBCs is the most useful thing to know before planning to run a model on the GPU. Higher is better for every metric.