This document summarizes a webinar comparing OpenCL and CUDA for GPU computing. It discusses key factors like programmability, portability, scalability, performance, and community support. Both OpenCL and CUDA can fully utilize hardware but performance depends on details. CUDA only runs on NVIDIA while OpenCL runs on multiple vendors but with less community support. Libraries help improve programmability over low-level kernels. ArrayFire provides a portable interface across CUDA and OpenCL with optimizations for different hardware.

1. ArrayFire Webinar:
OpenCL and CUDA Trade-Offs and Comparisons

2. GPU Software Features
Programmability
Portability Scalability
Performance Community

3. Leading GPU Computing Platforms
Programmability
Portability Scalability
Performance Community

4. Performance
• Both CUDA and OpenCL are fast
• Both can fully utilize the hardware
• Devil in the details:
– Hardware type, algorithm type, code quality

5. Performance
• ArrayFire results at end of webinar

6. Scalability
• Laptops --> Single GPU Machine
– Both CUDA and OpenCL scale, no code change
• Single GPU Machine --> Multi-GPU Machine
– User managed, low-level synchronization
• Multi-GPU Machine --> Cluster
– MPI

7. Scalability
Interesting developments:
• Memory Transfer Optimizations
– CUDA GPUDirect technology
• Mobile GPU Computing
– OpenCL available on ARM, Imgtec, Freescale, …

8. Scalability
• Laptops --> Single GPU Machine
– ArrayFire’s JIT optimizes for GPU type
• Single GPU Machine --> Multi-GPU Machine
– ArrayFire’s deviceset() function is super easy
• Multi-GPU Machine --> Cluster
– MPI

9. Portability
• CUDA is NVIDIA-only
– Open source announcement
– Does not provide CPU fallback
• OpenCL is the open industry standard
– Runs on AMD, Intel, and NVIDIA
– Provides CPU fallback

10. Portability
• ArrayFire is fully portable
– Same ArrayFire code runs on CUDA or OpenCL
– Simply select the right library

11. Community
• NVIDIA CUDA Forums – 26,893 topics
• AMD OpenCL Forums – 4,038 topics
• Stackoverflow CUDA Tag – 1,709 tags
• Stackoverflow OpenCL Tag – 564 tags

12. Community
• AccelerEyes GPU Forums – 1,435 topics
– Largest GPU forums by a software company
• Next largest
– PGI GPU Forums – 485 topics

13. Programmability
• Both CUDA and OpenCL are low-level
– Time consuming kernel development
– Data-parallel algorithm design
• Focus on programmability interfaces

14. Programmability
Faster
SSE or
Slower
AVX
Time-consuming Easy-to-use

15. Programmability
Writing
Faster
Kernels
SSE or
Slower
AVX
Time-consuming Easy-to-use

16. Programmability
Writing
Faster
Kernels
SSE or Compiler
Slower
AVX Directives
Time-consuming Easy-to-use

17. Programmability
Writing Using
Faster
Kernels Libraries
SSE or Compiler
Slower
AVX Directives
Time-consuming Easy-to-use

18. Libraries Make the Difference
Raw math libraries in NVIDIA CUDA
• CUBLAS, CUFFT, CULA, Magma
– Provides all BLAS, LAPACK, and FFT routines
necessary for most dense matrix operations
• CUSPARSE
– A good start for sparse linear algebra

19. Libraries Make the Difference
Raw math libraries in AMD OpenCL
• clAmdBlas, clAmdFft
– Provides most important Blas routines
– Provides radix 2, 3, and 5 FFT routines
– No LAPACK support
– No sparse data support

20. Libraries Make the Difference
Raw math libraries in AMD OpenCL
• clAmdBlas, clAmdFft
– Provides most important Blas routines
– Provides radix 2, 3, and 5 FFT routines
– No LAPACK support
– No sparse data support
– Runs on any OpenCL-compliant device

21. ArrayFire Code and Benchmarks