A NVIDIA logo is shown at SIGGRAPH 2017 in Los Angeles, California, U.S. July 31, 2017. REUTERS/Mike Blake
SAN FRANCISCO (Reuters) -Enfabrica, a Silicon Valley-based chip startup working on solving bottlenecks in artificial intelligence data centers, on Tuesday released a chip-and-software system aimed at reining in the cost of memory chips in those centers.
Enfabrica, which has raised $260 million in venture capital to date and is backed by Nvidia, released a system it calls EMFASYS, pronounced like "emphasis."
The system aims to address the fact that a portion of the high cost of flagship AI chips from Nvidia or rivals such as Advanced Micro Devices is not the computing chips themselves, but the expensive high-bandwidth memory (HBM) attached to them that is required to keep those speedy computing chips supplied with data. Those HBM chips are supplied by makers such as SK Hynix and Micron Technology.
The Enfabrica system uses a special networking chip that it has designed to hook the AI computing chips up directly to boxes filled with another kind of memory chip called DDR5 that is slower than its HBM counterpart but much cheaper.
By using special software, also made by Enfabrica, to route data back and forth between AI chips and large amounts of lower-cost memory, Enfabrica is hoping its chip will keep data center speeds up but costs down as tech companies ramp up chatbots and AI agents, said Enfabrica Co-Founder and CEO Rochan Sankar.
Rochan said Enfabrica has three "large AI cloud" customers using the chip but declined to disclose their names.
"It's not replacing" HBM, Sankar told Reuters. "It is capping (costs) where those things would otherwise have to blow through the roof in order to scale to what people are expecting."
(Reporting by Stephen Nellis in San Francisco; Editing by Jamie Freed)
