» Current | 2020 | 2019 | 2018 | Subscribe

Linley Newsletter

Qualcomm Samples First AI Chip

October 13, 2020

Author: Linley Gwennap

Qualcomm’s first standalone AI accelerator is making its way to the market, and its performance is impressive. The company’s estimated ResNet-50 inference throughput for its Cloud AI 100 design is similar to that of Nvidia’s new Ampere chip, but Qualcomm’s chip consumes only one-fifth the power. Other accelerators, though, offer performance and power efficiency closer to Qualcomm’s. The company has been sampling its new chip since March and expects to reach production in 1H21. Although cloud data centers can use the product, most initial customers plan to deploy it at the network edge.

The AI 100 comes in three versions. The most powerful is a half-height, half-length PCIe card rated at 400 trillion INT8 operations per second (TOPS), just ahead of our original prediction. The company says the flagship model will inference about 25,700 images per second (IPS) on ResNet-50. This card requires only 75W and thus is completely bus powered. It connects to a host processor using eight PCIe Gen4 lanes and is configurable with 16GB or 32GB of LPDDR4X DRAM.

Qualcomm also offers the AI 100 chip in a smaller Dual M.2 module that delivers roughly half as much performance at only 25W, maximizing power efficiency. This module occupies two M.2 slots, giving it twice the power budget and twice the board area of a standard M.2 module. The low-end model, targeting embedded applications, has the same form factor but reduces the power to 15W, causing both performance and efficiency to suffer. The company withheld pricing for these products.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Free Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »


Linley Fall Processor Conference 2020
October 20-22 and 27-29, 2020 (All Times Pacific)
Virtual Event
More Events »