Benchmarks
Speed depends on the work. These results have limits; they are not a promise for your machine.
Windows: new measurements pending
Windows runs real 3D applications. Speed is being benchmarked again on current master, and the results will be posted here.
No Windows frame-rate figure is published here while that work is pending. For the applications that have run, see What works today.
Linux: large jobs and small launches
These older results come from separate campaigns on rented RTX 3060 hosts. Each host was itself a VM. They used older kayfabe revisions. Large jobs kept pace with the host. Many small GPU launches cost more.
| Workload | Guest / host | Context |
|---|---|---|
| Matrix multiply, N=2048 | 0.997 | Host 786 GFLOPS; guest 784 GFLOPS |
| llama.cpp prompt processing | 0.998 | Qwen2.5-1.5B Q4 |
| llama.cpp token generation | 0.877 | Same model |
| Device-to-device copy, 32 MB | 0.98 | Host 327 GB/s; guest 322 GB/s |
| Eager PyTorch decoding | 0.29 | Many small launches |
| NVENC, 1080p | 0.72 | Older box |
Creating a CUDA context took about eight times longer in the guest. The full 65-program run took 1,193 seconds in the guest and 1,075 on the host. Sixty-one programs passed with equal output. Four need unified-memory paging, which is not supported yet.
The same set has not been run on the author's RTX 4070 without a cloud VM in between. Do not treat these older results as a current hardware comparison.
Reading the numbers
Each ratio compares the guest with its host on the same box. A ratio near one means similar speed for that test. It does not mean that every application is fast.
The performance FAQ explains the workload costs. The Linux guide covers test completion and the four programs that fail. Hardware coverage is on What works today.