By Stephen Nellis
SAN FRANCISCO (Reuters) -Nvidia and Menlo Micro on Wednesday said they have used technology from the startup to dramatically speed up the testing of AI chips, easing a significant production bottleneck.
The world's most valuable listed company and the central player in the AI boom has been working to iron out kinks in its processes as it works to feed seemingly insatiable demand for its chips.
It reports earnings after market close on Wednesday, with analysts expecting sales growth
of 56% to $56.9 billion, LSEG data showed. Even so, with valuations of AI companies sky-high, investors are watching for any signs of a bursting bubble.
Nvidia has sold millions of artificial intelligence chips, each of which has to be tested before sale by placing it on a specialized circuit board designed to determine whether it meets design goals such as speed and other functions.
Whereas the AI chips are cutting edge, however, many of the chips in the circuit boards for testing them are decades old. That makes testing the AI chips, which consume huge amounts of power and communicate at some of the fastest speeds in the industry, a challenge.
To address the bottleneck, Nvidia has been working with Menlo Micro, a startup spun out from GE in 2016 and which has raised $227.5 million in funding from Corning and the venture fund of iPhone co-creator Tony Fadell. The result is a set of switching chips that improve the performance of test boards.
Menlo Micro's chips use switches made out of metal, similar to a light switch on a wall but fabricated at the scale of microchips using technology from a field called micro-electromechanical systems.
In a research paper published on Wednesday, engineers from the two firms said testing of Nvidia's graphics processing units (GPU) could be sped up by 30% to 90% depending on the kind of test being performed.
Russ Garcia, Menlo Micro's chief executive, declined to say how much business the startup is doing with Nvidia but said other major chipmakers are adopting its switching chips for testing boards as well.
"The bottom line is, if you don't validate the GPUs before you get into the data center, you're going to have errors and other issues. This is the only way to validate these things at speed," Garcia said in an interview.
(Reporting by Stephen Nellis in San Francisco; Editing by Christopher Cushing)
