Intel Nervana AI Accelerators : Everything You Need To Know!

Intel just introduced their Nervana AI accelerators – the Nervana NNP-T1000 for training, and Nervana NNP-I1000 for inference!

Here is EVERYTHING you need to know about these two new Intel Nervana AI accelerators!

 

Intel Nervana Neural Network Processors

Intel Nervana neural network processors, NNPs for short, are designed to accelerated two key deep learning technologies – training and inference.

Intel Nervana AI Accelerators : Everything You Need To Know!

To target these two different tasks, Intel created two AI accelerator families – Nervana NNP-T that’s optimised for training, and Nervana NNP-I that’s optimised for inference.

They are both paired with a full software stack, developed with open components and deep learning framework integration.

 

Nervana NNP-T For Training

The Intel Nervana NNP-T1000 is not only capable of training even the most complex deep learning models, it is highly scalable – offering near linear scaling and efficiency.

By combining compute, memory and networking capabilities in a single ASIC, it allows for maximum efficiency with flexible and simple scaling.

Recommended : Intel NNP-T1000 PCIe + Mezzanine Cards Revealed!

Each Nervana NNP-T1000 is powered by up to 24 Tensor Processing Clusters (TPCs), and comes with 16 bi-directional Inter-Chip Links (ICL).

Its TPC supports 32-bit floating point (FP32) and brain floating point (bfloat16) formats, allowing for multiple deep learning primitives with maximum processing efficiency.

Its high-speed ICL communication fabric allows for near-linear scaling, directly connecting multiple NNP-T cards within servers, between servers and even inside and across racks.

  • Intel Nervana NNP-T1000
  • Intel Nervana NNP-T1000
  • High compute utilisation using Tensor Processing Clusters (TPC) with bfloat16 numeric format
  • Both on-die SRAM and on-package High-Bandwidth Memory (HBM) keep data local, reducing movement
  • Its Inter-Chip Links (ICL) glueless fabric architecture and fully-programmable router achieves near-linear scaling across multiple cards, systems and PODs
  • Available in PCIe and OCP Open Accelerator Module (OAM) form factors
  • Offers a programmable Tensor-based instruction set architecture (ISA)
  • Supports common open-source deep learning frameworks like TensorFlow, PaddlePaddle and PyTorch
  • Intel Nervana NNP-T1000 Mezzanine card
  • Intel Nervana NNP-T1000 Mezzanine card

 

Intel Nervana NNP-T Accelerator Models

The Intel Nervana NNP-T is currently available in two form factors – a dual-slot PCI Express card, and a OAM Mezzanine Card, with these specifications :

Specifications Intel Nervana NNP-T1300 Intel Nervana NNP-T1400
Form Factor Dual-slot PCIe Card OAM Mezzanine Card
Compliance PCIe CEM OAM 1.0
Compute Cores 22 TPCs 24 TPCs
Frequency 950 MHz 1100 MHz
SRAM 55 MB on-chip, with ECC 60 MB on-chip, with ECC
Memory 32 GB HBM2, with ECC 32 GB HBM2, with ECC
Memory Bandwidth 2.4 Gbps (300 MB/s)
Inter-Chip Link (ICL) 16 x 112 Gbps (448 GB/s)
ICL Topology Ring Ring, Hybrid Cube Mesh,
Fully Connected
Multi-Chassis Scaling Yes Yes
Multi-Rack Scaling Yes Yes
I/O to Host CPU PCIe Gen3 / Gen4 x16
Thermal Solution Passive, Integrated Passive Cooling
TDP 300 W 375 W
Dimensions 265.32 mm x 111.15 mm 165 mm x 102 mm

 

Nervana NNP-I For Inference

The Intel Nervana NNP-I1000, on the other hand, is optimised for multi-modal inferencing of near-real-time, high-volume compute.

Each Nervana NNP-I1000 features 12 Inference Compute Engines (ICE), which are paired with two Intel CPU cores, a large on-die 75 MB SRAM cache and an on-die Network-on-Chip (NoC).

Recommended : Intel NNP-I1000 PCIe + M.2 Cards Revealed!

Intel Nervana NNP-I1000 PCIe + M.2 Cards Revealed!

Intel Nervana NNP-I1000

It offers mixed-precision support, with a special focus on low-precision applications for near-real-time performance.

Like the NNP-T, the NNP-I comes with a full software stack that is built with open components, including direct integration with deep learning frameworks.

Intel Nervana NNP-I-1100 M.2 card

Intel Nervana NNP-I Accelerator Models

The NNP-I1000 comes in a 12 W M.2 form factor, or a 75 W PCI Express card, to accommodate exponentially larger and more complex models, or to run dozens of models and networks in parallel.

Specifications Intel Nervana NNP-I1100 Intel Nervana NNP-I1300
Form Factor M.2 Card PCI Express Card
Compute 1 x Intel Nervana NNP-I1000 2 x Intel Nervana NNP-I1000
SRAM 75 MB 2 x 75 MB
Int8 Performance Up to 50 TOPS Up to 170 TOPS
TDP 12 W 75 W

 

Recommended Reading

Go Back To > Business + Enterprise | Home

 

Support Tech ARP!

If you like our work, you can help support our work by visiting our sponsors, participating in the Tech ARP Forums, or even donating to our fund. Any help you can render is greatly appreciated!


Comments

comments

About The Author

Leave a Reply

%d bloggers like this: