Amazon EC2 offers hundreds of instance types, and the list grows every year. The naming looks cryptic at first, but it follows a simple pattern. Once you can read a name like m7g.2xlarge, picking the right instance becomes a short, logical process instead of guesswork.
How to read an instance name
m 7 g . 2xlarge
│ │ │ └── size
│ │ └── attributes (processor and extras)
│ └── generation
└── family- Family: the workload profile (m general purpose, c compute, r memory, and so on).
- Generation: higher is newer. Newer generations almost always give better price-performance, so prefer the latest available in your Region.
- Attributes:
g= AWS Graviton (Arm) processor,i= Intel,a= AMD,d= local NVMe instance storage,n= enhanced networking,e= extra memory or storage,flex= flexible variant that does not guarantee full CPU all the time. - Size: nano, micro, small, medium, large, xlarge, 2xlarge and up. Within a family, each size step roughly doubles vCPUs, memory and price.
The main families
| Family | Profile | Typical workloads |
|---|---|---|
| T (t3, t4g) | Burstable general purpose | Dev servers, small websites, low-traffic APIs |
| M (m7i, m7g) | Balanced CPU and memory | Application servers, mid-size databases, most default choices |
| C (c7i, c7g) | High CPU per GB of memory | Batch processing, web servers under heavy load, encoding, gaming servers |
| R, X (r7i, x2iedn) | High memory per vCPU | In-memory caches, large relational databases, analytics |
| I, D (i4i, d3) | Fast local storage | NoSQL databases, data warehouses, high-IOPS workloads |
| G, P, Inf, Trn | Accelerated (GPU or AWS chips) | Machine learning training and inference, graphics, video |
Burstable T instances: the credit model
T instances earn CPU credits while running below a baseline (for example 10–20% of a vCPU, depending on size) and spend them when they burst above it. In standard mode, an instance that runs out of credits is throttled to its baseline. In unlimited mode (the default for T3 and T4g), it can keep bursting and you pay extra for the surplus credits.
T instances are excellent for spiky, mostly idle workloads and terrible for steady high CPU. If CPUCreditBalance keeps falling to zero or surplus charges appear on your bill, move to an M or C instance.
Graviton: often the cheapest win
Instances with a g use AWS Graviton Arm processors and typically offer better price-performance than comparable x86 instances. Most Linux workloads on interpreted or JIT runtimes (Python, Node.js, Java, Go, .NET) run unchanged or with a rebuild. Watch for native dependencies and container images that only exist for x86, and test before switching production.
A practical selection method
- Start from the workload, not the catalog. Is it limited by CPU, memory, disk I/O, network or GPU? Measure it on your laptop, an existing server or a load test.
- Pick the family from the table above. When unsure, start with the latest M generation.
- Pick a size with headroom: aim for peak CPU around 60–70% and memory well below the limit.
- Prefer horizontal scaling: two smaller instances across two Availability Zones behind a load balancer beat one large instance for availability.
- Measure and adjust. Use CloudWatch metrics (CPU, network) plus the CloudWatch agent for memory and disk, and AWS Compute Optimizer for right-sizing recommendations after a couple of weeks of data.
Pricing options in one paragraph
The same instance type can cost very different amounts depending on how you buy it. On-Demand is flexible and the most expensive. Savings Plans and Reserved Instances trade a one- or three-year commitment for a large discount. Spot Instances use spare capacity at a deep discount but can be interrupted with a two-minute warning, which suits stateless, fault-tolerant work like batch jobs and Auto Scaling groups that mix Spot and On-Demand.
Common mistakes
- Running production databases on T instances and hitting credit exhaustion at peak.
- Staying on old generations (m4, c4) that cost more for less performance.
- Sizing for an imagined peak and paying for idle capacity all year instead of using Auto Scaling.
- Ignoring memory: CloudWatch does not show memory usage by default, so teams oversize CPU to fix what is really a memory problem.
Practice it
Launch and connect to your first instance in the free lab Getting Started with EC2. Then mix Spot and On-Demand in Mixed Instances ASG: Spot + On-Demand and cut idle cost in Right-Size and Schedule Non-Prod.
FAQ
What is a vCPU?
On most x86 instances, a vCPU is one hyperthread of a physical core. On Graviton instances, each vCPU is a full physical core, which is one reason they perform well per vCPU.
Can I change the instance type later?
Yes. Stop the instance, change its type and start it again (EBS-backed instances keep their data). Some changes, like moving between architectures, need a different AMI.
Which instance is free tier eligible?
Free tier eligibility depends on your account's plan and changes over time. Check the free tier page in the console rather than relying on old blog posts.