Amazon Web Services (AWS) has agreed to build Nvidia’s NVLink Fusion interconnect into a future generation of its home-grown Trainium AI accelerator, the cloud unit announced at its annual re:Invent conference in Las Vegas on Tuesday.
NVLink, which provides high-speed, cache-coherent links between processors, will let AWS create larger AI training clusters in which thousands of chips behave as a single system. The companies did not give a launch date for the Trainium4 devices that will carry the technology.
The tie-up also gives AWS customers access to “AI Factories”—on-premise racks that combine Trainium processors with Nvidia GPUs, Elastic Fabric Adapter networking and Amazon’s Bedrock and SageMaker AI services. Enterprises supply floor space, power and cooling; AWS installs and manages the stack.
Nvidia chief executive Jensen Huang said the partnership “creates the compute fabric for the AI industrial revolution,” while AWS compute and AI vice-president Dave Brown told Reuters the goal is to match rivals on raw performance while under-cutting them on price.
Alongside the future NVLink plans, AWS released Trainium3 servers on Tuesday. Each chassis houses 144 Trainium3 dies and, according to Amazon, delivers more than four times the training throughput of Trainium2 while consuming 40 per cent less energy. Brown declined to give absolute wattage or performance figures.
AWS also unveiled updated “Nova” foundation models—Nova 2 for faster text and image outputs, Sonic for speech-to-speech tasks—and a service called Nova Forge that lets companies fine-tune models on private data without erasing base-model knowledge.
Amazon’s share price rose 0.9 per cent to $235.98 in midday trading.