Intel Arc B580 Delivers Promising Linux GPU Compute Potential For Battlemage
The Intel Arc A-Series graphics cards evolved into being competitive for GPU compute purposes as the Intel compute stack matured. With the Intel Arc B580 testing are more indications of great GPU compute performance potential ahead. With SHOC's GEMM benchmark, the Arc B580 came out ahead of not only the Radeon RX 7600 series and GeForce RTX 4060 series, but even faster than the RX 7700 XT and RX 7800 XT graphics cards! More than 2x generational lift for this $249 graphics card!
The Intel Arc B580 easily led here in delivering the best performance per dollar.
But I did also encounter issues with the Intel Compute Runtime stack that the Battlemage support may not be fully-baked... In a subset of workloads, the Intel Arc B580 was performing worse than the A580.
With the FluidX3D CFD software with its FP32-F16S build, the Intel Arc B580 performance was astonishing and even shot well past the GeForce RTX 4070 Ti SUPER and Radeon RX 7900 XTX graphics cards...
A stunning victory for Intel Battlemage with the B580 for the FluidX3D FP32-FP16S performance. (Update: FluidX3D developer Moritz Lehmann has let me know that, unfortunately, the Battlemage results are so fast as ultimately they are incorrect. Battlemage employs a hardware optimization to skip kernel writes if it detects all zeroes to vRAM. The FluidX3D benchmark uses a zero-initialized box and thus this hardware optimization on Battlemage ends up skewing the results. FluidX3D will be updated to address this behavior by using a non-zero-initialized box.)
But not all went well. With the FluidX3D FP32-F16C test, the Arc B580 regressed to last place and well behind the Arc A580... Again, presumably just some software bugs within the Intel Compute Runtime that will hopefully be fixed soon. But the promising results elsewhere leads me to being quite excited about Battlemage for GPU compute workloads once the kinks are worked out.