This startup, backed by Bill Gates, is looking to transform computing as we know it - optical transistors could upend Moore's Law and allow more powerful GPUs

• Neurophos develops the Tulkas T100 optical processor capable of 470 petaFLOPS AI compute
• Optical transistors are 10,000 times smaller than conventional silicon photonics today
• Dual reticle design integrates 768GB of HBM for memory-intensive workloads

Austin-based startup Neurophos has revealed it is hard at work developing an optical processing unit called the Tulkas T100 which promises huge leaps forward in compute.

Funded by Bill Gates’ Gates Frontier Fund, the company claims the chip can deliver 470petaFLOPS of FP4 and INT4 compute while consuming between 1 and 2kW under load.

Its optical tensor core measures approximately 1,000 x 1,000, which is about 15 times larger than the standard 256 x 256 matrices used in current AI GPUs.

Optical transistors and extreme speeds

Neurophos’ optical transistors aim to exceed traditional semiconductor limits by extending Moore’s Law through higher compute density without increasing power consumption or chip size.

Despite its scale, the startup says it requires only a single core per chip, supported by extensive RAM and vector processing units to maintain throughput.

Its optical transistors are roughly 10,000 times smaller than current silicon photonics components, allowing a high-density matrix to fit on a single reticle-sized die.

“The equivalent of the optical transistor that you get from Silicon Photonics factories today is massive. It’s like 2mm long,” said Neurophos CEO Patrick Bowen.

“You just can’t fit enough of them on a chip in order to get a compute density that remotely competes with digital CMOS today.”

The Tulkas T100 operates at 56GHz, far exceeding previous CPU and GPU clock rates.

SRAM feeds the tensor core to maintain efficiency, and SSD storage may assist with moving large datasets during testing and simulation.

The chip uses a dual reticle design with 768GB of HBM to support memory-intensive AI workloads.

Neurophos says the first-generation Tulkas T100 will focus on the prefill stage of AI inference by handling input token processing for large language models.

Bowen envisions pairing racks of Tulkas chips with existing AI GPU racks to accelerate compute.

However, the company does not expect full production until mid-2028, with initial shipments numbering in the thousands.

Engineers are currently testing a proof-of-concept chip to validate the claimed compute density and power consumption.

Competitors such as Nvidia and AMD are also investing heavily in silicon photonics, signaling rising competition in the field.

AI tools and memory bandwidth constraints remain central considerations as optical accelerators seek to complement conventional GPUs.

While the Tulkas T100 shows potential to advance AI computation, its practical impact remains uncertain until the company achieves reliable production.

The optical approach remains experimental and faces challenges related to SRAM requirements, vector processing, and CMOS fabrication integration.

Optical transistors could speed matrix multiplication and reduce energy per operation, but effectiveness depends on memory, SSD storage, and AI integration.

Neurophos asserts its chips are compatible with standard semiconductor fabs, but mass production depends on resolving these engineering challenges.

Via The Register

Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!

And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.