I can't tell if they did use CUDA or not to test the nVidia GPUs. If not, since they used Intel TBB on the Xeon Phi's, the benchmark would be obviously flawed.
Also, perhaps the part where they used newer, more powerful CPUs when testing the Phi's had something to do with the performance jump.
Calculating greeks is not very difficult and bandwidth is going to be the bottleneck due to the high number of inputs necessary to calculate them. So no wonder that CPUs outperforms GPUs. Also, real time calculation of greeks is very easily achievable for the simple products (e.g. listed derivatives), banks are struggling to get real time performance for their more complex OTC products.
6 comments
[ 4.6 ms ] story [ 24.5 ms ] threadAlso, perhaps the part where they used newer, more powerful CPUs when testing the Phi's had something to do with the performance jump.
Stack under test:
- NVIDIA CUDA 5.5
- NVIDIA cuRand and cuBLAS
- CUB library
- Eigen 3 library
- GNU OpenMP (libgomp)
- 2 x NVIDIA K20X @ 2600MHz, 784MHz (with ECC enabled)
- IBM System X iDataPlex dx360 M4 Server
- 2 x Intel Xeon E5-2660 @ 2.20GHz (SandyBridge)
- CentOS Linux 6.4
- 96GB DDR3 RAM