Following the release of AMD Fusion APUs, system analytics and benchmark maker SiSoft contacted us with the new FAQ regarding the benchmark section. You can read the whole FAQ down below:
"All GPGPU benchmarks (launched in 2009) have been renamed "GP" (general purpose) as they can test GPGPUs, CPUs and APUs. The old name implied they only test GPUs.
Q: What benchmarks can test an APU?
A: All 3 GP(GPU) benchmarks can test APUs (i.e. both the GPU & CPU units simultaneously) using OpenCL, with the workload distributed across all devices depending on their respective processing power:
- GP Arithmetic (equivalent to the CPU Multi-Media benchmark)
- GP Cryptography (equivalent to the CPU Cryptography benchmark)
- GP Memory (equivalent to the Memory Bandwidth benchmark)
Q: Is OpenCL the only interface supported?
A: No, we also support CUDA and DirectX ComputeShader but these only run on GPUs. You can test CUDA/DX CS against OpenCL if you so want, the only differences are the respective compilers. The *very same code* runs on both CPU & GPUs, not just the same workload!
Q: Is this new in the GP(GPU) OpenCL benchmarks?
A: No, Sandra could always test multiple devices simultaneously and with OpenCL it can test both the CPUs and GPUs simultaneously: an APU is nothing special software-wise.
Q: What do you need to test the APU using the GP benchmarks?
A: Just Catalyst as it now includes the OpenCL driver for both CPUs & GPUs.
Q: How can I test the CPU or GPU part of the APU?
A: The combo-box in all GP benchmarks contains all 3 devices, so you can test each individually or together:
- GPU; CPU (a.k.a. whole APU)
- GPU only
- CPU only
Q: How can I compare an APU and traditional discrete CPU & (discrete) GPU?
A: If the GPU is AMD-based, Catalyst will provide OpenCL support for both the CPU (even Intel) and the AMD-based GPU. So you can test the APU against a discrete CPU + discrete GPU both simultaneously.
If the GPU is nV-based, install ForceWare which provides GPU OpenCL and either Intel's CPU OpenCL SDK or AMD's OpenCL CPU driver only. You can test the CPU + discrete GPU together!
If the GPU is Intel (e.g. Sandy Bridge) there is currently no OpenCL support for it, but there is DirectX ComputeShader support - however we believe it actually runs on the CPU (100% utilisation).
Q: Should I always test an APU against a CPU or GPU?
A: We think it is pretty fair to test an APU (CPU+GPU) againt a traditional CPU as future software should take fully advantage of the APU; legacy systems with no GPU-supporting OpenCL (integrated or discrete) will just have to do with the CPU.
So a relatively low-end APU might be faster than a traditional CPU due to the GPU part.
Q: What if the APU has a discrete GPU installed also?
A: No problem, Sandra can test all 3 devices simultaneously: CPU+int GPU+ext GPU."
Here at BSN*, we are going to continue using SiSoft Sandra 2011 and its future iterations in our reviews.