wafer-scale ai for science and hpc · 2020. 7. 1. · value proposition for hpc+ai for science...
TRANSCRIPT
![Page 1: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/1.jpg)
Cerebras Systems © 2020
Wafer-scale AI for science and HPCCerebras Systems
A HockISC Machine Learning Hardware Workshop 202025 June 2020
![Page 2: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/2.jpg)
Cerebras Systems © 2020
IntroductionOpportunityFrom e.g. fundamental physics to energy, environment, human health
AI has massive potential for science and HPC
![Page 3: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/3.jpg)
Cerebras Systems © 2020
IntroductionOpportunityFrom e.g. fundamental physics to energy, environment, human health
AI has massive potential for science and HPC
ChallengeNN compute is unique, challenging for legacy processors
Training commonly takes days-weeks, even on clusters of GPUThis is inefficient, expensive, constrains innovation
AI for science is compute-limited today
![Page 4: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/4.jpg)
Cerebras Systems © 2020
IntroductionOpportunityFrom e.g. fundamental physics to energy, environment, human health
AI has massive potential for science and HPC
ChallengeNN compute is unique, challenging for legacy processors
Training commonly takes days-weeks, even on clusters of GPUThis is inefficient, expensive, constrains innovation
AI for science is compute-limited today
Need
Massive compute: faster wall clock training and inferenceProgrammability with today’s tools, ability to optimize and extend
Systems that can integrate with HPC and scientific instrument facilitiesWe need a new solution for AI compute
![Page 5: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/5.jpg)
Cerebras Systems © 2020
Enter Cerebras SystemsFounded in 2016
Building systems to accelerate and change the landscape of compute for AI
200+ world-class engineering team- HW, SW, ML research
Cerebras Systems © 2020
![Page 6: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/6.jpg)
Cerebras Systems © 2020
Cerebras’ solution overviewA complete AI compute solution for AI at HPC scale
Unique, massive high performance processor→Training and inference→Orders of magnitude performance gain beyond legacy processors
Software stack to meet users where they are→Programmable as a single node with standard frameworks
System→Replaces racks of equipment with a single system→Straightforward deployment, orchestration
![Page 7: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/7.jpg)
Cerebras Systems © 2020
The Cerebras Wafer-Scale EngineThe world’s largest chip and most powerful processor for AI.
Designed from the ground-up to deliver orders of magnitude performance gain for deep learning.
- 215 x 215 mm, 1.2 trillion transistor chip- 400,000 cores- 18 GB on-chip SRAM- 100 Pb/s interconnect33,000x more bandwidth
![Page 8: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/8.jpg)
Cerebras Systems © 2020
Optimized architecture for DL computeMassive compute: cluster-scale resources on a single chip• Core optimized for neural network primitives• Flexible, programmable core – support evolving range of neural network architectures• Dataflow architecture, sparsity harvesting – designed for sparse compute native to NN
Local memory: all on-chip SRAM – efficient local access model weights & activations; datapathhas full performance from memory.
Fast interconnect: configurable on-chip fabric, communicate layer-to-layer with high bandwidth and low latency
Together unlock, e.g. model parallel execution on-chip, full utilization training and inference down to batch 1.
![Page 9: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/9.jpg)
Cerebras Systems © 2020
The Cerebras Software PlatformOur software stack makes the Wafer-Scale Engine easy to use:
→Programmable with today’s ML frameworks→Flexible and customizable with lower level APIs
![Page 10: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/10.jpg)
Cerebras Systems © 2020
The Cerebras CS-1A full solution in a single system:- Powered by the WSE- Programmable via TF, other frameworks- Install, deploy easily into a standard rack
Readily integrate with existing HPC systems- 1.2 Tbps I/O via 12x 100 GbE
Orchestrate with standard frameworks
Cluster multiple systems for greater acceleration and scale
Cerebras Systems © 2019
![Page 11: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/11.jpg)
Cerebras Systems © 2020
Value proposition for HPC+AI for science Orders of magnitude acceleration in wall-clock training time→More experiments per unit time; larger datasets, greater accuracy→Higher cadence re-training
Orders of magnitude higher throughput, lower latency inference→ Augment physics-based simulation→High performance inference for scientific instrument facilities
Flexible compute engine→Unlock research into new NN architectures, ML methods
![Page 12: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/12.jpg)
Cerebras Systems © 2020
Proud to introduce ISC to the Cerebras CS-1, the world’s most powerful deep learning solution.
Built to accelerate AI+HPC by orders of magnitude and empower researchers like you to do more, faster.
Concluding remarks
![Page 13: Wafer-scale AI for science and HPC · 2020. 7. 1. · Value proposition for HPC+AI for science Orders of magnitude acceleration in wall -clock training time →More experiments per](https://reader036.vdocument.in/reader036/viewer/2022071218/60505e4181e0b1426a1cf543/html5/thumbnails/13.jpg)
Cerebras Systems © 2020
Proud to introduce ISC to the Cerebras CS-1, the world’s most powerful deep learning solution.
Built to accelerate AI+HPC by orders of magnitude and empower researchers like you to do more, faster.
Multiple systems deployed and running customer workloads today, all the way from TF. - Accelerating AI for science, e.g. with DoE and NSF!
Call to action: bring us your big and different problems, system and partnership interests. Can’t wait to work together.
Thank you!
Concluding remarks