展开查看详情
1 .Intel Habana Train Product Introduction
1
2 .Agenda
‐ Habana Product roadmap & introduction
‐ Habana Training Product Applications
‐ Habana Training SW Ecosystem
2
3 .Product RoadMap
Purpose-Built for AI Deep Learning Training
Now, Gaudi2 offers even greater performance, scalability,
and usability.
https://habana.ai/training/
Purpose-Built for AI Inference
Greco will bring advanced performance, lower power and
design efficiency.
Coming soon.
https://habana.ai/inference/
3
4 . Training Processor Architecture
16nm, 32GB HBM, 24MB SRAM, 10x100Gb RoCE RDMA NICs 7nm, 96GB HBM2E, 48MB SRAM, 24x100Gb RoCE RDMA NICs
FP32, BF16, INT32, INT16, INT8, UINT32, UINT16 & UINT8 FP32, TF32, BF16, FP16 & FP8 (both E4M3 & E5M2)
Media (HEVC, H.264, VP9 & JPEG, pre/post-processing)
Matrix Multiplication Engine (MME)
Tensor Processor Core (TPC)
4
6 .Scaling systems with Gaudi2
6
7 .Gaudi2 typical workload performance
https://habana.ai/training/gaudi2/ 7
8 .Gaudi2 typical workload performance
https://habana.ai/training/gaudi2/ 8
9 .Gaudi2 typical workload performance
https://habana.ai/training/gaudi2/ 9
10 .MLPERF
Above metrics published by MLCommons, June 2022, https://mlcommons.org/en/training-normal-20/
https://habana.ai/blog/since-habanas-last-mlperf-submission/ 10
11 .Industries
https://habana.ai/ai-industries/
https://habana.ai/aws-ec2-dl1-instances/ 11
12 .What customers & partners are saying about Gaudi
12
13 .What customers & partners are saying about Gaudi
13
14 .What customers & partners are saying about Gaudi
14
15 .SynapseAI® SW Architecture
15
16 .SW Ecosystem
VMware Tanzu
16
17 .Developer Community
https://developer.habana.ai/
17
18 .Gaudi Documentation
https://docs.habana.ai/en/latest/
18
19 .Habana DL Examples for Training
https://github.com/HabanaAI/Model-References 19
20 .Habana Model Performance
https://developer.habana.ai/resources/habana-models-performance/#performance 20
21 .Developer Forum
https://forum.habana.ai/ 21
热门城市
