Download to read offline
Uploaded byAlex Moore
Cheap HPC
This document discusses ways to achieve cheap high-performance computing (HPC). It describes how graphics processing units (GPUs) can be used for general purpose computing through APIs like CUDA and OpenCL. GPUs have many parallel compute units that can speed up massively parallel problems. The document also discusses using cloud computing resources and Windows HPC Server to harness unused computing power for HPC tasks in a cost effective manner.
More Related Content
![[Container X mas Party with flexy] Machine Learning Lifecycle with Kubeflow o...](https://cdn.slidesharecdn.com/ss_thumbnails/20181218kubeflowazurepublic-181218060359-thumbnail.jpg?width=640&height=640&fit=bounds)
![Support, Monitoring, Continuous Improvement & Scaling Agentic Automation [3/3]](https://cdn.slidesharecdn.com/ss_thumbnails/agenticcommunityseries-day3-cfd-251120170304-ddef8112-thumbnail.jpg?width=640&height=640&fit=bounds)
Cheap HPC
- 1.
- 2. About Me Consultant forHMB, Inc. .NET / Web / Java Developer Polyglot Glutton - dabble in a lot of other languages. BSCSE, The Ohio State University
- 3. About Me I likesupercomputers
- 4.
- 5. Why Cheap HPC? Becausesometimes we need to go fast(er).
- 6. Why Cheap HPC? Becausesometimes we need to go fast(er). We all love saving money.
- 7. Why Cheap HPC? Becausesometimes we need to go fast(er). We all love saving money. Ways to get extra performance for little to no extra money.
- 8.
- 9.
- 10. GPGPU General Purpose computing on Graphics Processing Units
- 11.
- 12. GPGPU GPUs are greatat simple programs that are massively parallel, such as rendering images.
- 13. GPGPU GPUs are greatat simple programs that are massively parallel, such as rendering images. They can do this because they have many shaders, or “Compute Units”
- 14. GPGPU GPUs are greatat simple programs that are massively parallel, such as rendering images. They can do this because they have many shaders, or “Compute Units” Shaders can be programmed to do most things that a normal CPU can do.
- 15.
- 16. GPGPU! This Laptop (2007): 32 Compute Units across 4 cores @ 940 Mhz 128 MB DDR3
- 17. GPGPU! This Laptop (2007): 32 Compute Units across 4 cores @ 940 Mhz 128 MB DDR3 Tesla (NVidia GPGPU card) 448 Compute Units across 16 cores @ 1.15 Ghz Up to 5GB DDR5
- 18. GPGPU! This laptop’s GPUhas about as much computing power as the world’s fastest supercomputer in 1993 (~91 GigaFLOPS)
- 19. GPGPU! Special desktop computerswith 4 Tesla cards can do about ~4-5 TerraFLOPS About the fastest supercomputer in 1999-2000, ON YOUR DESKTOP
- 20.
- 22.
- 23. GPGPU! So what’s cheapabout it?
- 24. GPGPU! So what’s cheapabout it?
- 25. GPGPU! So what’s cheapabout it? You can probably program your GPU, and it’s just free untapped power
- 26. GPGPU NVidia - CUDA ATI- Stream SDK Microsoft - DirectCompute Apple - OpenCL
- 27. Problem Euler Demo The following iterative sequence is defined for the set of positive integers: n n/2 (n is even) n 3n + 1 (n is odd) Using the rule above and starting with 13, we generate the following sequence: 13 40 20 10 5 16 8 4 2 1
- 28.
- 29.
- 30. Project Euler 14 Ruby~151s C#/Java ~6.1s
- 31. Project Euler 14 Ruby~151s C#/Java ~6.1s C# Multithreaded ~3.2s
- 32. Project Euler 14 Ruby~151s C#/Java ~6.1s C# Multithreaded ~3.2s C (gcc -O3) ~0.5s
- 33. Project Euler 14 Ruby~151s C#/Java ~6.1s C# Multithreaded ~3.2s C (gcc -O3) ~0.5s C / OpenCL ...
- 34. OMG Ponies LiveDemo! Subset of c99 No recursion, no function pointers, no standard headers
- 35. To The Cloud! CloudComputing
- 36. The Cloud If youuse the cloud for your site or as a datastore right now, you can acquisition temporary extra computing power very easily.
- 37. The Cloud TalentRooster neededto re-encode all 1200 of their video resumes. To do this locally, it would have taken about 27 hours, with 60% of that time would have been downloading and uploading.
- 38.
- 39. The Cloud The devteam added an extra EC2 instance, and re-encoded the videos in 3 hours.
- 40. The Cloud The devteam added an extra EC2 instance, and re-encoded the videos in 3 hours. Total Cost:
- 41. The Cloud The devteam added an extra EC2 instance, and re-encoded the videos in 3 hours. Total Cost: $5.46
- 42. The Cloud Amazon alsohas instances with 2 Telsa GPUs on each instance. - Just sayin’ :)
- 43. Stealing Candy FromBabies Windows HPC Server 2008 R2
- 44. HPC Server Instead ofbuying more computing power, just make idle computers do your work.
- 45. HPC Server HPC Server2008 R2 allows you to add Windows 7 workstations as compute nodes. Also allows Azure instances to be added as compute nodes when you need extra horsepower.
- 46. HPC Server About $925for the Suite, including licenses for compute nodes.
- 47.
- 48. Time and Money Time CLOUD THREADING HPC SERVER C GPGPU
- 49. Time and Money Money THREADING HPC SERVER C CLOUD GPGPU
- 50. Thanks! anm@hmbnet.com @alexmoore www.enginerding.com