Skip to content

SYSTEM Cited by 1 source

AWS Graviton3

Overview

AWS Graviton3 is Amazon's third-generation custom Arm server processor, available as EC2 c7g instances (general availability 2022). The design targets compute density in the cloud rather than peak single-thread clock speed: 2.6 GHz base, only 100 MHz above Graviton2, on a modern 5 nm process, packaged three chips per node. The result is a dense, low-power-per-core product priced per-vCPU below comparable Intel/AMD offerings. Source: Graviton 3: First Impressions (Chips and Cheese), via sources/2022-07-11-highscalability-stuff-the-internet-says-on-scalability-for-july-11th-2022.

Design philosophy

"The design seems to be very narrowly targeted at maximizing compute density in the cloud. To that end, AWS has chosen very conservative core clocks. At 2.6 GHz, Graviton 3 clocks a mere 100 MHz over its predecessor, Graviton 2. … Such a core running at 2.6 GHz on 5 nm should absolutely sip power. That in turn lets AWS pack three of these chips into a single node, increasing compute density. The final result is a chip that lets AWS sell each Graviton 3 core at a lower price, while still delivering a significant performance boost over their previous Graviton 2 chip."

James Hamilton on the economic framing (Perspectives blog):

"Graviton3 has the performance needed for modern compute intensive applications such as gaming, video encoding, machine learning, and high-performance computing while also delivering the power-performance and price-performance customers seek for less demanding, general-purpose workloads."

Real-workload datapoints (from Hamilton)

  • Formula 1 CFD (Computational Fluid Dynamics): +40% performance on c7g vs. c6gn.
  • Honeycomb.io telemetry ingestion: +35% throughput and -30% latency vs. c6g.

Both are compute-bound, large-working-set workloads where per-core power efficiency at 5 nm pays off.

Strategic position in AWS

Graviton3 is part of AWS's broader push to make regions the unit of scale:

"something that people often under appreciate about AWS is that new regions are by far the biggest needle mover for our business" — Colm MacCárthaigh.

And the internal transition: "when I joined, regions were pets, but we've been making them cattle." Custom Arm silicon at high compute density is a structural enabler of that pet→cattle shift; you can only make regions disposable when compute is affordable enough to over-provision them.

Trade-off framing: density vs. clock

  • Win: lower $/vCPU, lower power per core, higher rack density.
  • Loss: not the chip to pick for latency-critical single-threaded workloads where every MHz matters (HFT, game servers with tight per-tick deadlines).

For the bulk of general-purpose cloud workloads, the density trade favors Graviton3.

Last updated · 517 distilled / 1,221 read