In a comparison Michael Larabel published for Phoronix on October 12, 2026, Mesa 26.3 Git Rusticl delivered about 66% of ROCm 10 OpenCL’s performance across the tested benchmarks on an AMD Ryzen AI Max+ 395 (“Strix Halo”) in a Framework Desktop. Phoronix’s aggregate result applies to that test suite and platform.
The tested platform and OpenCL versions
Phoronix compared both implementations on the same system. It described Rusticl as supporting OpenCL 3.1 and the tested ROCm driver as providing OpenCL 2.1 compliance. The test setup and version details give the context for the performance results.
Hashcat and GpuOwl outcomes
ROCm 10 OpenCL was faster in every Hashcat algorithm Phoronix tested: MD5, NTLM, SHA1, 7-Zip, bcrypt, SHA-256, SHA-512, LastPass, WPA-PBKDF2-PMKID+EAPOL, and Kerberos 5 + etype 23 + TGS-REP. The reported GpuOwl Mersenne-primality test ran with ROCm OpenCL but crashed under Rusticl with a BUILD_PROGRAM_FAILURE error. Phoronix’s workload results describe these outcomes for the tested applications.
Where Rusticl came closer
Michael Larabel’s Phoronix results for memory and compute workloads and the report’s clpeak summary show a more mixed picture beyond Hashcat and GpuOwl.
| Workload | ROCm 10 OpenCL | Mesa 26.3 Git Rusticl |
| Memory reads | Significantly faster than Rusticl | Behind ROCm |
| Memory copy and write | Quite comparable to Rusticl | Quite comparable to ROCm |
| ProjectPhysX | Led by significant margins in some data types | More competitive in FP32 and INT32 |
| FluidX3D | Ahead, with Rusticl not far behind | Not far behind ROCm |
| clpeak | Lower kernel latency | Generally competitive, but with much higher kernel latency |
The results make the workload a more useful guide than the aggregate alone: Phoronix’s conclusion favors ROCm when performance is the priority, while describing Rusticl as a more convenient, unified Mesa option for users less focused on speed.
