“As ChatGPT and other generative AI continue reaching a fevered pitch, the semiconductor industry needs to step up with supporting hardware,” explained Khaled Maalej, VSORA founder and CEO. “That’s what we intend to do with our single-chip architecture to trounce the memory wall and accelerate generative AI implementations.”
“The VSORA ChatGPT-on-chip architecture can unleash the unbounded potential of Generative AI such as ChatGPT,” said Nicolas von Bülow, managing partner at Clipperton Finance, a VSORA financial adviser, who went on to suggest that the chip would help to spearhead what he described as a, “paradigm shift in the semiconductor industry.”
Jotunn has been designed to accelerate the proliferation of low-cost, low-power generative AI implementations whose monthly visit numbers have increased geometrically to close to two billion. Current leading processors idle for almost 97% of the time waiting for processing data when implementing GPT-3 or GPT-4.
ChatGPT, based on GPT-3.5, can be handled by Jotunn4 entirely on-chip, dropping power consumption by more than an order of magnitude versus its competitors. It’s claimed that Jotunn4 improves the cost effectiveness of a GPT-4 solution by more than 150 times compared to its closest competitor.
The Jotunn4 solution provides 1.6 petaflops with efficiencies of more than 50% for GPT-3.5 and GPT-4. Typically, efficiencies currently range around 2-4%. Further, Jotunn4 has 192 GB of on-chip memory and consumes less than 100W in a 45 x 45 mm package.
Initially designed as a low-power, low-cost and high-performance chip architecture for L3-L5 autonomous driving vehicles, VSORA extended its patented architecture to accelerate GenAI applications.
Algorithm implementation is quick and easy, utilising the VSORA high-level language and graph compiler development flow. The future-proof VSORA architecture is algorithm agnostic, allowing existing as well as new algorithms to be run on any VSORA chip.
Jotunn4 is expected to ship in 2024.
