Volantis raises $88M to develop photonic inference systems

Volantis Inc. recently secured $88 million in funding to develop advanced photonic inference systems, as announced by the company. The Series A round was led by angel investor Lachy Groom and venture capital firm Abstract Ventures, with additional participation from prominent investors such as Kleiner Perkins chair John Doerr and former Intel AI products group head Naveen Rao.

Volantis' founding team boasts impressive industry experience, with engineers who previously worked at major chipmakers like Nvidia Corp. and Broadcom Inc. Notably, they were involved in the first commercial implementation of CoWoS, a widely used interconnect in graphics processing units.

The company's core technology involves an inference-optimized chip architecture centered around a custom memory interconnect, which promises more than 30 times the memory bandwidth of existing accelerators. Traditional GPU memory bandwidth limitations arise from the short wire lengths (up to 5 millimeters) connecting memory modules to processing cores, restricting the number of memory modules that can be fitted on a chip.

Volantis' chip architecture overcomes this constraint with interconnect wires extending over 200 millimeters, enabling the installation of more than 220 memory chiplets and significantly boosting memory bandwidth.

To achieve this high-bandwidth interconnect, Volantis utilizes an optical design based on vertically-cavity surface-emitting lasers (VCSELs). These lasers convert electrical signals into light using a quantum well and two mirrors, a manufacturing process that is more cost-effective and readily available compared to traditional laser technologies. The use of gallium arsenide, a compound more commonly used for miniature lasers, further reduces production costs.

The company's first product, a data center inference appliance called the A-1, is projected to be about a third the size of a standard server rack. The A-1 features 10 terabytes of memory and 250 terabits per second of memory bandwidth, capable of processing up to 10,000 tokens per second when running a model with 20 trillion parameters.

This real-time inference capability, combined with the ability to handle larger context windows, Volantis claims, will enable faster task completion times, such as a coding agent finishing a task in 30 seconds instead of 30 minutes. The A-1 system is slated for shipment next year.