Volantis announced an $88 million Series A round on October 1, SiliconANGLE's Maria Deutscher reported. It was led by angel investor Lachy Groom and Abstract Ventures, with more than a half-dozen other participants including Kleiner Perkins chair John Doerr and Naveen Rao, former head of Intel's AI products group. The article says Volantis's engineers previously worked at Nvidia and Broadcom and credits them with the first commercial implementation of CoWoS, an interconnect widely used in GPUs.
The problem it targets
Memory bandwidth largely determines how fast a GPU can run inference, and current designs are limited by wires of about five millimeters that restrict how many memory modules fit on a chip. Volantis is developing a chip architecture for inference built around a custom memory interconnect that uses optical links, which transmit data as light. The light comes from microscopic VCSELs, which the company says are cheaper to manufacture than the lasers typically used in optical networking and can be made from gallium arsenide.
Company claims
Volantis says the interconnect provides more than 30 times the memory bandwidth of current accelerators, and that wires with a range over 200 millimeters allow more than 220 memory chiplets on one AI accelerator. Its first product, the A-1, is a data center inference appliance about one-third the size of a standard server rack. The company claims it will have 10 terabytes of memory and 250 terabits per second of memory bandwidth, and up to 10,000 tokens per second when running a model with 20 trillion parameters, which is an estimate. It plans to start shipping next year.
What the CEO said
Co-founder and CEO Tapa Ghosh wrote in a blog post: “This will enable real-time frontier inference, restart scaling laws & enable entire code bases in context windows.” As a starting point, he said, imagine a coding agent that completes a task in 30 seconds rather than 30 minutes.
Caution
The performance figures have not been independently validated, and the A-1 has not yet shipped.
Source: SiliconANGLE. This is an AI Business Watch summary of the original report.