US start-up puts AI Core in SSDs

Time:  2019-08-05 Source:  New Electronics (UK)

InnoGrit, a US start-up, has developed a set of three controllers for solid-state drives (SSDs), including one for data centres that embeds a neural-network accelerator.

 

The chips are appearing on the market just as NAND flash prices are showing signs of bottoming out. OEMs and data centres are expected to take advantage of the lower prices to continue the shift from hard-disk storage in notebooks and servers that require performance, power, or size advantages.

Commenting on the adoption of SSDs in data centres Zining Wu, who co-founded InnoGrit, said it was "Happening fast. When we talk to data centre customers, all their new designs are flash-based.”

InnoGrit’s Tacoma uses four PCI Gen4 interfaces to support 16 NAND channels, delivering up to 1.5 million I/O operations/second (IOPS) at less than 5-W peak power. It includes a mid-level Arm core as well as an implementation of NDLA, an open-source inference accelerator from Nvidia.

“With this combination, we can do intelligent processing such as data labelling, using the tool chain offered by Nvidia, or some savvy customers will put in their own firmware,” said Wu.

Samsung announced an SSD, last year, that embeds a Xilinx Zynq FPGA to handle processing for a variety of AI, database, and video apps.

The Tacoma uses a 64+8-bit data bus, as well as DDR3/4 and LPDDR3/4 DRAMs, and provides AES-256, SHA3, and ECC security. The mid-tier Rainier controller is a smaller version of the design, supporting eight NAND channels and 32- and 16-bit data buses for low-end servers or high-end notebooks.

Rainier delivers up to 1 million IOPS at less than 3-W peak power. Both Rainier and Tacoma support sequential reads at up to 7 GB/s and sequential writes at 6.1 GB/s.

Wu claims that he has a lead on competitors across several figures of merit. However, rival Phison Electronics of Taiwan recently announced an eight-channel SSD controller using four PCIe Gen 4 links.