Urgent.News

What's breaking now, across thousands of outlets.

AI

Altman-backed Volantis reveals plan to vault the memory wall by baking photonics into AI accelerators

Who said you had to use wires to connect memory to compute?

Altman-backed Volantis reveals plan to vault the memory wall by baking photonics into AI accelerators

AI chip startup Volantis, backed by Sam Altman, is set to challenge the memory wall by integrating photonics into their AI accelerators. The San Francisco-based company, founded in 2022, plans to develop an accelerator that can accommodate up to 10 TB of memory at speeds up to 240 TB/s. This ambitious goal is achieved through the use of a photonic interposer, a relatively novel approach that differs from traditional compute architecture or memory technology developments.

Volantis' CEO, Tapa Ghosh, explains that the optical interposer allows for a significant increase in high-bandwidth memory per package by using optical wave guides instead of electrical interconnects. This design enables the memory to be located a few centimeters away from the compute, eliminating the shoreline as a bottleneck and maximizing signal integrity. While HBM (High Bandwidth Memory) is a common example, Volantis' approach aims to push the boundaries even further.

The company is not relying on external lasers for its interposer, but rather using micro vertical-cavity surface-emitting lasers (micro-VCSELs) integrated directly into the interposer. This design choice is intended to improve reliability over time while maintaining high performance. Volantis' CTO, Roy Meade, emphasizes that by adopting a "wide-and-efficient" approach, the company can achieve higher bandwidth without relying solely on increasing link speeds.

Despite the innovative design, Volantis is not planning to license its technology to other chipmakers. Instead, the company aims to develop a fully-fledged inference chip. While the company has raised $88 million in funding to scale from small-scale optical demonstrators to a production-ready chip, the road ahead is still challenging.

The interposer alone is a complex task, compounded by the integration of licensed IP. However, Ghosh remains confident that the company can deliver an MVP product within 12 to 18 months, marking a significant milestone in the pursuit of the memory wall.

Written by urgent.news from The Register's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Also reported by 1 other outlet

Read the original at theregister.com →

More in AI

iphone-use: let an AI agent drive a real iPhone (open source)

Lots of apps have no API: banking, payment, health, chat. However good an AI agent is, it couldn't operate them for you. iphone-use hands a real iPhone to an AI agent: it reads the screen as text…

  • iphone-use connects real iPhone to AI agents via HTTP API
  • Daemon runs on Mac, controls iPhone without interfering screen
  • Performance metrics track each action for optimization

More from Monday 5 October →