💡Volantis Develops AI Accelerator with 240 TB/s Memory Bandwidth
Volantis aims to break GPU memory limits with photonic tech
TL;DR
Volantis is developing an AI accelerator, A-1, with 240 TB/s memory bandwidth and up to 10 TB capacity, targeting 20 trillion parameter models. The photonic interposer technology could revolutionize high-memory workloads.
Volantis is developing an AI accelerator, A-1, with 240 TB/s memory bandwidth and up to 10 TB capacity. This photonic interposer technology aims to break the memory bandwidth bottleneck of modern GPUs. Developers working on large-scale AI models will benefit from faster memory access and higher capacity, enabling more efficient training and inference. The A-1 chip uses optical waveguides to extend memory location, achieving 1 picojoule per bit per 24 Gbps lane, and a total system power of around 20kW. Volantis is aiming to integrate compute and memory from partners onto the interposer, with the goal of scaling from small-scale optical demonstrators to a final product.
Key Points
Volantis' A-1 chip uses photonic interposer technology, achieving up to 240 TB/s memory bandwidth.
A-1 can pack up to 10 TB of memory onto a single package, serving 20 trillion parameter models.
The chip uses optical waveguides to extend memory location, breaking the GPU memory bottleneck.
A-1 aims for 1 picojoule per bit per 24 Gbps lane and a total system power of around 20kW.
Volantis is licensing everything else from IP providers, focusing on the optical interposer.
Why It Matters
If you're working on large-scale AI models, Volantis' A-1 chip could be a game-changer. With 240 TB/s memory bandwidth and up to 10 TB capacity, it can serve 20 trillion parameter models at 10,000 tok/s per user. This technology could significantly improve training and inference efficiency for high-memory workloads.
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.
Join 3,569 builders reading daily.