HL Series
Developer	Habana
Manufacturer	TSMC
Type	NPU
Introduction	2017 (announced) 2018 (launch)
µarch	Goya, Gaudi
Process	16 nm 0.016 μm 1.6e-5 mm
Technology	CMOS

HL-x is a series of neural processors designed by Habana for the data center market.

Overview

HL-x is a series of neural processors designed by Habana for both training and inference in the data center. There are two distinct products based on two separate microarchitectures - one for inference and one for training. Both designs share common unified architecture, simplifying software development.

Members

Habana offers two slightly different products - one of inference and one for training.

Inference

HL-1000

Main article: Goya

HL-100/102 PCIe Card

The HL-1000 is a first-generation neural processor for inference. The HL-1000 is fabricated on TSMC 16-nanometer process based on the Goya microarchitecture. This chip ships in a number of products in a PCIe cards form factor.

• Proc 16 nm
• Mem Dual-channel DDR4-2667
• I/O x16 PCIe Gen 4.0
• TDP 200 W

Model	TDP	Memory	Type
HL-100	200 W	4/8/16 GiB	PCIe Card; Dual-slot, passive cooling
HL-102	200 W	4/8/16 GiB	PCIe Card; Dual-slot, active cooling

Training

HL-2000

Main article: Gaudi

HL-205 OAM board

Habana HLS-1 system with eight HL-205s

The HL-2000 is a first-generation neural processor for training. The HL-1000 is fabricated on TSMC 16-nanometer process based on the Gaudi microarchitecture. This chip ships in a number of products in both a PCIe card and an OCP OAM form factors. Habana also designed the HLS-1 system which incorporates eight HL-205 accelerators and can serve as a ready solution or a reference design

• Proc 16 nm
• Mem HBM2 (1 TB/s)
• I/O x16 PCIe Gen 4.0, 8-10x 10 GbE

Model	TDP	Memory	Scale-Out	Type
HL-200	200 W	32 GiB	8x100 GbE	PCIe Card; passive cooling
HL-202	200 W	32 GiB	8x100 GbE	PCIe Card; active cooling
HL-205	300 W	32 GiB	10x100 GbE	OCP-OAM

See also

• NNP