Nvidia Unveils the Ampere A100, Its Biggest Generational Leap Yet
Nvidia's new A100 GPU packs 54 billion transistors and up to 20x the AI performance of its predecessor, launching at a virtual GTC 2020.
Nvidia usually saves its biggest hardware announcements for a stage in San Jose packed with developers. Today it did the reveal from Jensen Huang’s kitchen instead, and the chip itself was worth the awkwardness of a virtual keynote. The A100, built on the new Ampere architecture, is the company’s first real architectural jump since Volta, and by Nvidia’s own numbers it’s not a small one.
The headline spec is 54 billion transistors, which Nvidia is calling the largest 7nm chip ever made. That’s an enormous amount of silicon to fab reliably at that node, and it tells you Nvidia is betting hard on TSMC’s process maturity to make yields work at this scale. The payoff, according to Nvidia, is up to 20x the AI training and inference performance of the previous-generation Volta-based V100. Even accounting for the usual gap between vendor marketing numbers and what customers see in production, that’s a big enough claimed jump to reshape how data centers plan GPU purchases for the next cycle.
Multi-Instance GPU is the interesting part
The transistor count and raw throughput numbers will get the headlines, but the feature I think matters more long-term is Multi-Instance GPU, or MIG. It lets a single A100 be partitioned into as many as seven fully isolated GPU instances, each with its own memory and compute resources. Up to now, if you wanted hard isolation between workloads on a GPU, you bought separate GPUs — full stop. MIG means a cloud provider or an internal ML platform team can carve up one A100 to serve multiple smaller jobs, inference workloads, or even multiple tenants, without the workloads stepping on each other. That’s a utilization story as much as a performance story, and utilization is where a lot of money gets wasted in GPU-heavy data centers.
Nvidia is packaging eight A100s into its DGX A100 system, and the combined box is rated at 5 petaflops. That’s a genuinely startling number for a single appliance, and it puts DGX A100 squarely in “replaces a rack of older servers” territory for AI training clusters. Expect the major cloud providers to start standing up A100 instances relatively quickly, since Nvidia’s data center partners tend to move fast once new silicon ships, and expect the usual suspects — large language model researchers, computer vision teams, recommendation-system builders at big tech companies — to be first in line.
What I’ll be watching for over the next few months is real third-party benchmarking against V100 and against AMD’s data center GPUs, since a 20x claim from the manufacturer is a starting point, not a verdict. I’d also want to see how MIG performs under actual multi-tenant load rather than in a demo. But even with the usual skepticism applied to launch-day numbers, this is clearly the biggest architectural swing Nvidia has taken in years, and it’s arriving at exactly the moment demand for AI training and inference capacity is exploding. Good timing, whether or not it was planned that way.