Volantis develops photonic links to expand AI chip memory capacity
Volantis, a Sam Altman-backed AI chip startup, is developing an accelerator using photonic interconnects to overcome a fundamental constraint in AI chip design. By using light-based connections instead of electrical wires to link memory to processors, the company claims it can dramatically increase the amount of high-bandwidth memory available to large language models. This tackles the "memory wall" – a critical bottleneck that currently limits memory capacity on modern accelerators like AMD's MI455X, which maxes out around 432GB because of physical constraints on the chip's perimeter.
Volantis's approach integrates micro-VCSELs (vertical-cavity surface-emitting lasers) into an interposer to allow memory to sit centimetres away from the compute die rather than millimetres. The company claims its A-1 chip can support up to 10TB of memory at 240TB/s speeds, enough to serve a 20 trillion parameter model at 10,000 tokens per second per user. Rather than developing novel memory or compute technology itself, Volantis is licensing components and focusing solely on differentiating through the optical interposer.
- Optical interconnects let AI chips hold far more memory without spatial constraints
- Volantis claims 10TB memory at 240TB/s on one package
- Company is licensing all other technology, focusing only on the photonic interposer
New here? Start with this
Modern artificial intelligence chips face a critical constraint called the "memory wall". The processors that power large language models need enormous amounts of fast memory positioned directly next to them, but there is only so much physical space along a chip's edge where memory can connect. Current cutting-edge chips like AMD's MI455X can hold around 432GB of memory because of this physical limitation alone.
Volantis is an artificial intelligence startup backed by OpenAI chief Sam Altman. The company is developing tiny lasers that transmit data as light beams, allowing memory to sit several centimetres away from the main processor rather than millimetres away. This approach, if successful, would let chips access vastly more memory whilst maintaining the extreme speeds that artificial intelligence systems require.
If this works, it matters because companies could run much larger artificial intelligence models and serve far more users simultaneously from a single chip. Volantis claims its design could hold ten terabytes of memory, far exceeding current chips, whilst sustaining the high speeds these systems need. This addresses one of the fundamental constraints that has limited how powerful artificial intelligence hardware can become.
Read the full article at the source →
Originally published by The Register as “Altman-backed Volantis reveals plan to vault the memory wall by baking photonics into AI accelerators”.