EC2 instance families explained: matching workload to hardware

Quick answer

EC2 instance families group instance types by resource profile: general purpose (M, T) balance CPU and memory; compute optimized (C) favor CPU; memory optimized (R, X) favor RAM; storage optimized (I, D) favor local disk throughput; and accelerated (P, G, Inf) add GPUs or accelerators. Matching the family to your workload's bottleneck is the biggest single cost lever, because a memory-heavy workload on a compute family (or vice versa) forces you to oversize and overpay. Within a family, Graviton variants add a further discount, and the T-family adds burstable economics for low-average-CPU work.

Before you tune anything else on EC2, you pick an instance family, and that choice dominates cost. Each family is tuned to a resource profile, and running a workload on the wrong-shaped family forces you to oversize (buying excess CPU to get enough RAM, or vice versa) and overpay. Matching workload to family is the first and biggest cost lever in EC2.

The main families

Family	Optimized for	Typical workloads
General purpose (M, T)	Balanced CPU and memory	Web servers, app servers, microservices
Compute optimized (C)	High CPU per dollar	Batch, encoding, gaming, HPC
Memory optimized (R, X, Z)	High RAM per dollar	Databases, caches, in-memory analytics
Storage optimized (I, D, H)	Local disk throughput/capacity	NoSQL, data warehouses, log processing
Accelerated (P, G, Inf, Trn)	GPUs and accelerators	ML training, inference, graphics

The letter tells you the profile: M and T are balanced, C is CPU-heavy, R and X are memory-heavy, I and D are storage-heavy, and P, G, Inf, and Trn carry GPUs or ML accelerators. The trailing number is the generation (higher is newer and usually better value), and suffixes indicate the processor: g means Graviton (ARM), a means AMD, i means Intel.

Match the family to your bottleneck

Identify what your workload is bound by. CPU-bound (batch processing, encoding, simulation): compute optimized C. Memory-bound (databases, caches, large in-memory datasets): memory optimized R or X. Balanced (typical web and application servers): general purpose M, or T if average CPU is low and bursty (see burstable instances). Local-storage-bound (high-IOPS NoSQL, log crunching): storage optimized I or D. Needs a GPU (ML, graphics): accelerated, and see GPU instance pricing. Getting this right avoids oversizing to compensate for the wrong shape.

Generation and processor: free discounts

Within the right family, two choices lower cost for equivalent work. Newer generations offer better price-performance, so a current-generation instance usually beats an older one at the same or lower price. And Graviton (g) variants deliver similar performance at a lower price than Intel or AMD for compatible workloads, as covered in Graviton cost savings. Choosing the newest generation and Graviton where compatible is close to free money.

Then size and buy right

Family choice sets the shape; sizing sets the amount. Pick the smallest size in the right family that meets your real usage (right-sizing), then reduce the rate with commitments for steady load, as in how much savings plans save, or Spot for interruptible work. The order matters: right family first (avoids structural overspend), then right size, then right purchase model. For how the per-hour rate is constructed across all of these, see EC2 pricing explained.

FAQ

What are EC2 instance families?

Groups of EC2 instance types sharing a resource profile: general purpose (M, T) balance CPU and memory; compute optimized (C) favor CPU; memory optimized (R, X, Z) favor RAM; storage optimized (I, D, H) favor local disk; and accelerated (P, G, Inf, Trn) add GPUs or ML accelerators. Matching the family to your workload's bottleneck is the biggest single EC2 cost lever, because the wrong shape forces you to oversize.

How do I choose the right EC2 instance family?

Identify your workload's bottleneck. CPU-bound work (batch, encoding, simulation) suits compute optimized C; memory-bound work (databases, caches, in-memory analytics) suits memory optimized R or X; balanced web and app servers suit general purpose M or burstable T; local-storage-bound work suits storage optimized I or D; and GPU or ML work needs accelerated families. Matching family to bottleneck avoids oversizing to compensate.

What do the letters and numbers in EC2 instance types mean?

The leading letter indicates the family profile (M and T balanced, C compute, R and X memory, I and D storage, P and G accelerated). The trailing number is the generation, with higher being newer and usually better value. Suffix letters indicate the processor: g is Graviton (ARM), a is AMD, i is Intel. So c7g is a 7th-generation compute-optimized Graviton instance.

What is the difference between compute and memory optimized instances?

Compute optimized (C family) provides high CPU per dollar with a lower memory-to-CPU ratio, suited to CPU-bound work like batch processing, video encoding, and HPC. Memory optimized (R, X, Z families) provides high RAM per dollar with more memory per vCPU, suited to memory-bound work like databases, caches, and in-memory analytics. Running one workload type on the other family forces oversizing and overspend.

Does choosing a newer instance generation save money?

Usually yes. Newer generations offer better price-performance, so a current-generation instance typically delivers equal or better performance at the same or lower price than an older one. Migrating to the latest generation within the same family is a low-risk cost win. Combining it with Graviton variants where compatible compounds the saving for equivalent work.

Should I use Graviton instances to save money?

For compatible workloads, yes. Graviton (ARM, the g suffix) instances deliver similar performance at a lower price than comparable Intel or AMD instances, so choosing the Graviton variant of your chosen family lowers cost for equivalent work. Most common workloads run on ARM with little or no change, making Graviton one of the easiest EC2 discounts to capture once you have picked the right family.

What to do next

Price any EC2 instance family before you launch it. C3X reads your Terraform and prices your resources against a live catalog. Start with the quickstart.