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

What happened

Sam Altman-backed AI chip startup Volantis has announced its plan to address the memory wall in AI accelerators by incorporating photonics. The company is developing an AI accelerator called A-1, which uses a photonic interposer to increase memory bandwidth and capacity. This approach allows for up to 10 TB of memory at 240 TB/s speeds, enabling the support of large-scale AI models. Volantis is focusing on optical waveguides and micro-VCSELs to enhance reliability and performance.

Why it matters

Overcoming the memory wall is crucial for advancing AI capabilities, as current limitations in memory bandwidth and capacity hinder the performance of large-scale models. Volantis' approach could enable more efficient and powerful AI accelerators, benefiting developers and businesses reliant on high-performance computing.

• Volantis is leveraging photonics to overcome the memory wall in AI accelerators.
• The use of optical waveguides and micro-VCSELs improves reliability and performance.
• The company's focus on a fully fledged inference chip sets it apart from competitors.

The bigger picture

Volantis' innovation in photonic interposers could lead to more efficient AI accelerators, addressing current limitations in memory bandwidth and capacity. This could enable the development of larger and more complex AI models, enhancing performance for various applications.

Sam Altman-backed Volantis is pioneering the integration of photonics into AI accelerators to address the memory wall, with plans to develop a high-performance AI chip that could significantly enhance AI model capabilities.

Technical details

• Photonic interposer
• Optical waveguides
• Micro-VCSELs (micro vertical-cavity surface-emitting lasers)
• HBM (High Bandwidth Memory)
• Interconnects (link speed and number of links)
• Power efficiency (1 picojoule per bit per 24 Gbps lane)
• System power (around 20kW)

FAQ