The AMD Fusion APU

Presentation for the High Performance Computing and Architecture course at KAUST

Transcript

1. The AMD Fusion APU CS 280: High Performance Computing and Architecture Emaad Ahmed Manzoor April 22, 2014
2. Llano AMD Fusion APU: Llano. Branover, Alexander, Denis Foley, and Maurice Steinman. Micro, IEEE '12. On the Efficacy of an APU for Parallel Computing. Daga et al. SAAHPC '11. Trade-Offs The tradeoffs of fused memory hierarchies in heterogeneous computing architectures. Spafford, Kyle L., et al. CF '12. Efficacy Patterns Characterizing the Impact of Memory Access Patterns on AMD Fusion. Lee et al. SC '12.
3. GPUs
4. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Top500.org Performance Energy Efficiency
5. GPUs VS CPUS Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
6. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU On the Efficacy of an APU for Parallel Computing. Daga et al. SAAHPC '11.
7. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Top500.org GPUs: ~50% of theoretical peak CPUs: ~78% of theoretical peak
8. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU On the Efficacy of an APU for Parallel Computing. Daga et al. SAAHPC '11.
9. AMD Fusion APU
10. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU The tradeoffs of fused memory hierarchies in heterogeneous computing architectures. Spafford, Kyle L., et al. CF '12.
11. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Tradeoffs CPU, CPU + Discrete GPU, APU? - Cache coherency VS scalability - Latency VS throughput - Power VS performance
12. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Cache Coherency VS Scalability - GPUs: Flat, simple, incoherent Scales well with number of processor tiles Relaxed consistency for groups of cores Hard to program for - CPUs: Multi-level, high-capacity, coherent Much less scalable - Key tradeoff Couple CPU and GPU caches to enforce coherence, while preserving scalability to a large number of cores
13. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU The tradeoffs of fused memory hierarchies in heterogeneous computing architectures. Spafford, Kyle L., et al. CF '12.
14. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU The tradeoffs of fused memory hierarchies in heterogeneous computing architectures. Spafford, Kyle L., et al. CF '12.
15. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU The tradeoffs of fused memory hierarchies in heterogeneous computing architectures. Spafford, Kyle L., et al. CF '12.
16. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU The tradeoffs of fused memory hierarchies in heterogeneous computing architectures. Spafford, Kyle L., et al. CF '12.
17. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Fusion Memory Hierarchy - CPU-like accesses: Traditional caches - GPU-like accesses: Radeon Memory Bus - Cache coherence: Fusion Compute Link
18. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Latency VS Throughput Allocation of transistors and die space - CPUs: Optimizing latency in single-threaded programs Caches, instruction-level parallelism, branch predictors - GPUs: Simple, many floating point units Schedule thousands of threads onto these cores - Question: How much of the resources should we dedicate to serial VS parallel processing units?
19. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
20. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
21. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU GPU is the computational workhorse
22. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
23. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
24. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Memory hierarchy performance
25. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Capacity VS Bandwidth Type of physical memory used - CPUs: Optimizing latency to high-capacity DDR3 memory - GPUs: Concerned with repeatedly streaming a fixed-size buffer Maximise bandwidth by using GDDR3 Wider memory bus, higher clock speed Lower capacity (same power budget) - Fused: GPU and CPU cores must use the same type of physical memory
26. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Capacity VS Bandwidth Type of physical memory used - CPUs: Optimizing latency to high-capacity DDR3 memory - GPUs: Concerned with repeatedly streaming a fixed-size buffer Maximise bandwidth by using GDDR3 Wider memory bus, higher clock speed Lower capacity (same power budget) - Fused: GPU and CPU cores must use the same type of physical memory
27. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
28. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU GDDR5: High bandwidth at lower clock speed, low capacity
29. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
30. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU HD5670: Only 30% higher clock speed
31. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU HD5670: Only 30% higher clock speed
32. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Bandwidth-limited benchmarks
33. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Performance improvements commensurate or greater than the increase in power consumption
34. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU PCIe bound benchmarks
35. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Performance increase for the Llano, though its slower shader clock and lower memory bandwidth
36. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
37. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU Power VS Performance - Fusion APU: Slower shader clock Tighter integration leading to fine-grained DVFS Lesser data movement across wires
38. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
39. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
40. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU
41. Summary
42. Emaad Ahmed Manzoor April 22, 2014 The AMD Fusion APU - GPUs benefit HPC performance for data-parallel problems - Limited by the PCIe bandwidth - APUs replace the PCIe with a unified northbridge - Tradeoffs - Cache Coherency VS Scalability - Capacity VS Bandwidth - Power VS Performance
43. .