Startup Sagence AI Tackles GPU Energy Drain With Analog Chips
Sagence AI develops analog chips to provide a more energy-efficient alternative to power-hungry GPUs for AI workloads. These in-memory processors aim to eliminate data bottlenecks and reduce the massive electricity demands of modern data centers.
GPUs currently power most AI models but consume massive amounts of energy, prompting Goldman Sachs to estimate a 160% increase in electricity demand by 2030. To combat this unsustainable trend, veteran chip designer Vishal Sarin launches Sagence AI to create energy-efficient alternatives. The startup develops both the analog hardware and the software needed to program these specialized chips.
Unlike traditional digital chips that store data as binary ones and zeros, Sagence uses analog chips that represent data through a continuous range of values. This older technological approach requires far fewer components for complex calculations and keeps data directly in the memory modules. By eliminating the need to shuttle information back and forth between memory and processors, these in-memory chips avoid major bottlenecks and achieve higher data density.
Analog technology presents distinct challenges, including difficulties in achieving high precision during manufacturing and increased complexity in programming. Despite these hurdles, Sarin envisions these analog chips complementing rather than completely replacing existing digital infrastructure. The company ultimately strives to break through current performance and economic limitations to make practical AI computing more environmentally responsible.