Looking At The OpenCL Performance Of ATI & NVIDIA On Linux

As highlighted in our NVIDIA GeForce GTX 460 768MB review, under Linux this GF104 "Fermi" graphics card performs quite closely to the ATI Radeon HD 4890 graphics card when it comes to the OpenGL graphics performance. However, the OpenCL computing performance is very much a different story. When beginning with the Phoronix Test Suite's JuliaGPU test profile, the NVIDIA GeForce GTX 460 had completely clobbered the ATI Radeon HD 4890 and Radeon HD 5770 graphics cards. In fact, even the mid-range GeForce GT 240 was outperforming both ATI Radeon HD graphics cards in our initial GPGPU Linux tests.

NVIDIA's wins with this JuliaGPU OpenCL test are due to architectural improvements made to recent NVIDIA graphics cards with the GeForce GT 240 even outperforming the higher-end but older GeForce 9800GTX 512MB graphics card. In fact, the GeForce GT 240 was 2.67x faster than the GeForce 9800GTX in this computing benchmark. The GeForce GTX 460 OpenCL performance in this test was a whopping 7.47x greater than the GeForce 8800GT.

With the MandelGPU test, the GeForce GTX 460 again was the clear winner over the other five graphics cards that we benchmarked under Linux. This time, however, the performance of the Radeon HD 4890 and Radeon HD 5770 were much closer. This test highlights again though that the Fermi architecture is significantly faster than earlier generations of NVIDIA GPUs with the GTX 460 being more than four times faster than the 9800GTX. The GeForce GT 240 continued to run faster than the GeForce 9800GTX with its OpenCL computing performance, which is certainly something you will never see if comparing their OpenGL performance as we have done in earlier articles.

With SmallPT GPU, which is a port of the CPU-based SmallPT benchmark that has already been in the Phoronix Test Suite, the GeForce GTX 460 performance was astonishing compared to the earlier NVIDIA hardware and the ATI competition when using the Cornell rendering scene. The GeForce GTX 460 was twice as fast as the Radeon HD 4890 and Radeon HD 5770 graphics cards and incomparably faster to non-Fermi NVIDIA hardware -- the GT104 was 64x faster than the G92 core.