Execution experiments
The runtime support matrix covers CPU, Metal, CUDA and optional Vulkan. Separate experiments investigate other execution paths on the same physical hardware. They are not additional devices or new runtime backends.
The four experiments in the source device backlog are complete within their documented scope as of 24 October 2026.
Validated paths
Section titled “Validated paths”| Hardware and path | Result | Runtime availability |
|---|---|---|
| GTX 980 Ti / Vulkan-SPIR-V | Tensor operations, inference, guarded generated kernels and local GPU acceptance | Optional Linux Vulkan backend, validated on this GPU |
| i7-6700K / Mesa lavapipe | 192 guarded float32 cases, two lifecycles, 1,742/1,742 owned resources released, zero validation errors | Standalone CPU software experiment; excluded from GPU discovery |
| GTX 980 Ti / NVIDIA OpenCL | 192 guarded float32 cases, two contexts, 490/490 owned resources released | Standalone GPU experiment |
| i7-6700K / PoCL OpenCL CPU | 192 guarded float32 cases, two contexts, 490/490 owned resources released | Standalone CPU experiment |
The standalone probes exercise real allocation, upload, guarded add/multiply,
readback and cleanup. Their cases include empty and tail sizes, zero/partial/full
active prefixes, separate output and first-input aliasing. Readbacks, untouched
suffixes, padding and preserved inputs match tensor.cx CPU operations, guarded
Kernel.reference execution and NumPy.
The NVIDIA OpenCL run used driver 580.178.04 and OpenCL C 1.2. The PoCL CPU run used PoCL 6.0 with LLVM 18.1.8 and its pthread device. Wrong-device selection, compiler failures and incorrect unguarded results were rejected. Clean CPU wheel installs also ran both OpenCL probes from the source distribution.
Scope and reproduction
Section titled “Scope and reproduction”OpenCL is not a registered tensor.cx backend. These probes do not add
device="opencl", OpenCL DLPack exchange, full operation coverage, multi-device
execution, or a performance claim. PoCL and lavapipe run on the CPU; they are
not evidence of another GPU architecture. The small finite test fixtures do not
establish general float32 special-value or subnormal conformance.
Use the versioned source instructions for the external drivers/tools, exact device selection, commands and acceptance limits:
Other NVIDIA/AMD/Intel GPUs, real two-GPU execution and other operating systems still need their own acceptance. Wider device work and the isolated GPU runner remain in the roadmap.