Avicena partners with ams OSRAM to enable high-volume future production of ultra-low energy chip-to-chip interconnects
AvicenaTech Corp., a privately held company in Sunnyvale, CA, has partnered with ams OSRAM to develop high-volume manufacturing of GaN microLED arrays for its industry-leading LightBundle communication architecture.
The need for next generation computing power is here, driven by strong AI/ML and HPC application demand – for products like ChatGPT, DALL-E, autonomous vehicle training, and many others. Attempts to scale current architectures are running headlong into physical limits leading to slower throughput growth, power-hungry and hard to cool systems.
The Avicena LightBundle architecture breaks new ground by unlocking the performance of xPUs, memory and sensors – removing key constraints of bandwidth and proximity while simultaneously offering an order-of-magnitude reduction in power consumption.
Today’s high-performance ICs use SerDes-based electrical links to achieve adequate IO density. However, the power consumption and bandwidth density of these electrical links degrade quickly with length. Conventional optical communications technologies developed for networking applications have been impractical for inter-processor and processor-memory interconnects due to their low bandwidth density, high power consumption, and high cost. Moreover, co-packaging existing laser sources with hot ASICs does not fit well for reliability reasons unless external laser sources (ELS) are used which increases complexity and cost.
Avicena’s LightBundle links use densely packed arrays of GaN microLEDs to create highly parallel optical interconnects with typical throughputs of > 1Tb/s at energies of < 1 pJ/bit. A LightBundle cable uses a highly multicore multimode fiber to connect a GaN microLED transmitter array to a matching array silicon photodetectors (PDs). Arrays of hundreds or thousands of LightBundle’s microLEDs and PDs are easily integrated with standard CMOS ICs, enabling the closest integration of optical interconnects with electrical circuits.
In addition to high energy efficiency and high bandwidth density, these LightBundle links also exhibit low latency since the modulation format of the individual links is simple NRZ instead of PAM4 which is common in many modern optical links but has the disadvantage of higher power consumption and additional latency.
The need for next generation computing power is here, driven by strong AI/ML and HPC application demand – for products like ChatGPT, DALL-E, autonomous vehicle training, and many others. Attempts to scale current architectures are running headlong into physical limits leading to slower throughput growth, power-hungry and hard to cool systems.
The Avicena LightBundle architecture breaks new ground by unlocking the performance of xPUs, memory and sensors – removing key constraints of bandwidth and proximity while simultaneously offering an order-of-magnitude reduction in power consumption.
Today’s high-performance ICs use SerDes-based electrical links to achieve adequate IO density. However, the power consumption and bandwidth density of these electrical links degrade quickly with length. Conventional optical communications technologies developed for networking applications have been impractical for inter-processor and processor-memory interconnects due to their low bandwidth density, high power consumption, and high cost. Moreover, co-packaging existing laser sources with hot ASICs does not fit well for reliability reasons unless external laser sources (ELS) are used which increases complexity and cost.
Avicena’s LightBundle links use densely packed arrays of GaN microLEDs to create highly parallel optical interconnects with typical throughputs of > 1Tb/s at energies of < 1 pJ/bit. A LightBundle cable uses a highly multicore multimode fiber to connect a GaN microLED transmitter array to a matching array silicon photodetectors (PDs). Arrays of hundreds or thousands of LightBundle’s microLEDs and PDs are easily integrated with standard CMOS ICs, enabling the closest integration of optical interconnects with electrical circuits.
In addition to high energy efficiency and high bandwidth density, these LightBundle links also exhibit low latency since the modulation format of the individual links is simple NRZ instead of PAM4 which is common in many modern optical links but has the disadvantage of higher power consumption and additional latency.
No comments: