Revision as of 10:12, 1 February 2020

NNP
Developer	Intel
Manufacturer	Intel, TSMC
Type	Neural Processors
Introduction	May 23, 2018 (announced) 2019 (launch)
Process	28 nm 0.028 μm 2.8e-5 mm , 16 nm 0.016 μm 1.6e-5 mm , 10 nm 0.01 μm 1.0e-5 mm
Technology	CMOS
Package	PCIe x16 Gen 3 Card, OCP OAM, M.2

Neural Network Processors (NNP) is a family of neural processors designed by Intel Nervana for both inference and training.

Overview

Neural network processors (NNP) is a family of neural processors designed by Intel for the acceleration of artificial intelligence workloads. The name and originally architecture originated with the Nervana startup prior to its acquisition by Intel in 2016. Although the first product was announced in 2017, it never made it past customer sampling which eventually served as a learning product. Intel eventually productized those chips starting with their second-generation designs in late 2019.

The NNP family comprises two separate series - NNP-I for inference and NNP-T for training. The two series use entirely different architectures. The training chip is a direct descendent of Nervana's original ASIC design. Those chips use the PCIe and OAM form factors that have high TDPs designed for maximum performance at the data center and for workstations. Unlike the NNP-T, NNP-I inference chips are the product of Intel IDC which, architecturally, are very different from the training chips. They use Intel's low-power client SoC has the base SoC and build the AI architecture from there. The inference chips use low-power PCIe, M.2, and ruler form factors designed for servers, workstations, and embedded applications.

On January 31, 2020, Intel announced that it has discontinued the Nervana NNP product line in favor of the unified architecture it has acquired from Habana Labs a month earlier.

Learning (NNP-T)

Lake Crest

Main article: Lake Crest µarch

The first generation of NNPs were based on the Lake Crest microarchitecture. Manufactured on TSMC's 28 nm process, those chips were never productized. Samples were used for customer feedback and the design mostly served as a software development vehicle for their follow-up design.

T-1000 Series (Spring Crest)

NNP T-1000

Main article: Spring Crest µarch

Second-generation NNP-Ts are branded as the NNP T-1000 series and are the first chips to be productized. Fabricated TSMC's 16 nm process based on the Spring Crest microarchitecture, those chips feature a number of enhancements and refinments over the prior generation including a shift from Flexpoint to Bfloat16. Intel claims that these chips have about 3-4x the training performance of first generation. Those chips come with 32 GiB of four HBM2 stacks and are packaged in two forms - PCIe x16 Gen 3 Card and an OCP OAM.

NNP-T 1400 OAM Module.

• Proc 16 nm process
• Mem 32 GiB, HBM2-2400
• TDP 300-400 W (150-250 W typical power)
• Perf 108 TOPS (bfloat16)

Inference (NNP-I)

I-1000 Series (Spring Hill)

NNP I-1000

Main article: Spring Hill µarch

The NNP I-1000 series is Intel's first series of devices designed specifically for the acceleration of inference workloads. Fabricated on Intel's 10 nm process, these chips are based on Spring Hill and incorporate a Sunny Cove core along with twelve specialized inference acceleration engines. The overall SoC design borrows considerable amount of IP from Ice Lake. Those devices come in M.2 and PCIe form factors.

NNP-I Ruler

NNP-I Ruler Chassis.

• Proc 10 nm process
• Mem 4x32b LPDDR4x-4200
• TDP 10-50 W
• Eff 2.0-4.8 TOPs/W
• Perf 48-92 TOPS (Int8)

Intel also announced NNP-I in an EDSFF (ruler) form factor which was designed to provide the highest compute density possible for inference. Intel hasn't announced specific models. The rulers were planned t come with a 10-35W TDP range. 32 NNP-Is in a ruler form factor can be packed in a single 1U rack.

See also

• DL Boost