Graviton versions: from experiment to default
Graviton is AWS’s own Arm processor. In 2024 AWS said more than half of the CPU capacity it had recently added to EC2 was Graviton. In 2018 it was one instance family and a leap of faith.
The version matters when you pick instances. “Graviton” can mean a 2018 Cortex-A72 core or a 2026 chip with triple the throughput, and the discount versus x86 moved with every generation. Here’s the story of all five, with prices and benchmarks from our dataset.
Where it came from
In 2015 AWS bought Annapurna Labs, a small Israeli chip design firm, a deal TechCrunch reported at around $350 million. Annapurna’s first big wins weren’t CPUs. Its chips took over storage and networking work inside EC2 servers, work that had been burning host cycles, and became the Nitro system modern EC2 runs on.
Graviton followed the same playbook: design the chip around what a cloud provider actually needs, then make it boring.
Graviton1 (2018): the proof of concept
The first Graviton appeared at re:Invent 2018 in exactly one family: a1, 16 cores of ARM Cortex-A72 at 2.3 GHz. It was cheap. An a1.xlarge ran $0.102/hour, roughly half of m5.xlarge.
Two problems killed it for most buyers. Every vCPU came with 2 GiB of memory, half the ratio x86 general-purpose instances offered. And per-core performance ran behind: about 11,900 CoreMark per vCPU against m5’s 14,800. You got what you paid for.
Graviton1 proved custom Arm chips could run at cloud scale. It didn’t prove they were better, so most teams watched from the sidelines.
Graviton2 (2019-2020): the turning point
Announced in December 2019, Graviton2 moved to ARM’s Neoverse N1 core, 64 per chip, and fixed the memory ratio: m6g matches m5’s 4 GiB per vCPU exactly. Instances rolled out through 2020, m6g in May, c6g and r6g in June, the burstable t4g in September.
This is where the pitch changed from “cheaper” to “cheaper and faster”. In our data an m6g.xlarge posts about 19,300 CoreMark per vCPU, roughly 30% more than m5.xlarge, while costing 20% less per hour ($0.154 vs $0.192 in us-east-1). The burstable pair tells the same story: t4g.medium runs about 34% more CoreMark per vCPU than t3.medium at 19% less per hour.
Two more things made Graviton2 stick. The software story resolved: Amazon Linux 2, the major distros, Docker, and most popular open source stacks shipped arm64 builds. And the chip escaped EC2: RDS, Aurora, ElastiCache, OpenSearch, EMR, and eventually Lambda and Fargate offered Graviton options, so you could benefit without touching your AMIs.
AWS claimed up to 40% better price performance than comparable x86. That figure was generous at the edges, but the direction was no longer in doubt.
Graviton3 (2021-2023): DDR5 and vector math
Graviton3 was announced at re:Invent 2021 and first shipped in c7g in May 2022, with m7g and r7g following in early 2023. New Neoverse V1 cores, DDR5 memory, and Arm’s SVE vector extensions, which matter for ML inference and media work. AWS claimed up to 25% more performance than Graviton2.
Our measured gap on the m-family is bigger: m7g.xlarge posts about 25,800 CoreMark per vCPU, 34% over m6g. The x86 discount held near 19%: $0.1632 versus m7i.xlarge at $0.2016.
This generation also branched by workload. c7gn pushed network bandwidth up to 200 Gbps, hpc7g targeted tightly coupled HPC jobs, and AWS later sold a higher-clocked Graviton3E for HPC buyers who needed more single-thread speed.
Graviton4 (2023-2024): the big one
Graviton4 jumped to Neoverse V2 cores and a much wider memory system, 12 channels of DDR5-5600. AWS puts the memory bandwidth gain at about 60% over Graviton3, with up to 30% better performance for the launch family.
Two firsts. The memory-optimized family shipped first: r8g went GA in July 2024, months before m8g and c8g in September. And the sizes got serious: r8g tops out at 192 vCPUs and 1,536 GiB, x8g at 3 TiB, territory that used to require x86 high-memory instances.
Measured, m8g.xlarge posts about 31,200 CoreMark per vCPU, another 21% over m7g.
Graviton5 (announced 2024, instances 2026)
AWS announced Graviton5 at re:Invent 2024: 192 cores and a claimed 25% uplift over Graviton4. The first instances took a while to arrive. m9g and c9g showed up in mid-2026 at 3.3 GHz, the highest clock speed any Graviton has shipped.
m9g.xlarge measures about 36,500 CoreMark per vCPU in our data, up 17% from m8g. Region coverage is still filling in, so treat Graviton5 as availability permitting rather than a planning default.
What the benchmarks say
CoreMark per vCPU across five generations, all 4-vCPU xlarge sizes from our dataset:
| Instance | Graviton | CoreMark / vCPU | Introduced |
|---|---|---|---|
| a1.xlarge | Graviton1 | ~11,900 | 2018 |
| m6g.xlarge | Graviton2 | ~19,300 | 2020 |
| m7g.xlarge | Graviton3 | ~25,800 | 2023 |
| m8g.xlarge | Graviton4 | ~31,200 | 2024 |
| m9g.xlarge | Graviton5 | ~36,500 | 2026 |
For reference, m5.xlarge (Skylake-era Intel, 2017) sits at about 14,800. m6g passed it in 2020 while costing 20% less.
CoreMark is a microbenchmark. It tells you roughly what one vCPU can do, not how your database or your Go service will behave, and whole-instance throughput doesn’t scale linearly at the top sizes. Use the table to read the generational direction, then benchmark your own workload before committing.
The discount, generation by generation
The sticker discount versus the same-size Intel instance has narrowed with each generation:
- Graviton2: m6g.xlarge is 20% cheaper than m5.xlarge
- Graviton3: m7g.xlarge is 19% cheaper than m7i.xlarge
- Graviton4: m8g.xlarge is 15% cheaper than m8i.xlarge
(All us-east-1, Linux on-demand, from our data.)
The narrowing doesn’t mean Graviton lost ground. Each generation still raises per-core performance faster than it raises price. What it does mean is that the “Graviton is always 20% off” rule you may have internalized is stale, and newer pairings can land closer to 10-15%. Check the exact types you’ll run in the instance explorer or line them up in the comparison tool before you commit. The migration question itself is covered in Graviton vs. x86.
How to spot a Graviton instance
- A
gafter the generation digit: m7g, c8g, c9g, t4g. - Suffixes stack:
gdadds local NVMe storage (m7gd),gnadds network bandwidth (c7gn),gbadds more bandwidth on top (c8gb). - The x86 neighbors use
ifor Intel (m7i) andafor AMD (m7a). - Family codes like
im4gnandis4genare Graviton2 storage-optimized instances;x2gdpairs memory-heavy sizing with local NVMe.
Every instance page in the compute section lists architecture and processor explicitly, so when a name gets confusing, look it up there.
Bottom line: Graviton1 was a proof of concept, Graviton2 made Arm a mainstream choice, and every generation since has been faster per core than the last while staying cheaper per hour than the Intel instance beside it. For most Linux workloads the question has shifted from whether to run Graviton to which generation to run.