Decode the EC2 instance type name m7g.2xlarge: what does each part of the name tell you, and how many vCPUs and how much memory does that instance have?

## Why the name is a grammar AWS publishes hundreds of EC2 instance types, and no interviewer expects you to recite them. What they expect is that you can look at a type you have never seen and say what it is for. The name is fully systematic, so one parsing rule covers the whole catalogue. ``` m 7 g .2xlarge family generation capability size ``` ## Part 1 — the family letter fixes the ratio The leading letter tells you the vCPU-to-memory ratio and therefore which workload the type is shaped for: - `t` — burstable general purpose. A low sustained baseline with the ability to burst above it. Cheap, and the right default for dev boxes and low-traffic services. - `m` — general purpose, about **4 GiB of memory per vCPU**. The safe default when you have not measured anything. - `c` — compute optimised, about **2 GiB per vCPU**. Batch, encoding, CPU-bound API servers. - `r` — memory optimised, about **8 GiB per vCPU**. Caches, in-memory analytics, large JVM heaps. `x` and `u` families go further still. - `i` — storage optimised with large local NVMe. `d` families are dense HDD storage. - `g`, `p`, `inf`, `trn` in the **leading** position — accelerated computing: GPUs for graphics (`g`), GPUs for training/HPC (`p`), and AWS's own inference and training accelerators. The ratios are the point. If you know a service needs 32 GiB and 8 vCPUs, you already know it is an `m` shape; if it needs 64 GiB with the same 8 vCPUs, it is an `r` shape. ## Part 2 — the generation digit A higher digit is newer silicon. m5 → m6i → m7i are successive general-purpose generations. Newer generations normally deliver more performance for the same or lower on-demand price, so "move to the current generation" is one of the cheapest optimisations available. Generations are not interchangeable in every region — a new type reaches regions in waves, which is why launch failures sometimes trace back to a type simply not existing where you asked for it. ## Part 3 — capability letters after the generation These stack, in a fixed order, and each one is a real change to the hardware: - `g` — AWS Graviton, an **arm64** processor. Requires an arm64 AMI and arm64 binaries. - `a` — AMD EPYC; `i` — Intel. Both x86_64, but different silicon and price points. - `d` — includes local NVMe **instance store** volumes. - `n` — higher network and EBS bandwidth for the same vCPU count. - `e` — extra memory or storage; `z` — higher sustained clock speed. So `c7gn.4xlarge` reads as: compute optimised, 7th generation, Graviton, network enhanced, 16 vCPUs. The collision worth memorising: `g` as the **first** letter is a GPU family (`g5.2xlarge` is an NVIDIA GPU instance), while `g` **after the generation digit** is Graviton (`m7g.2xlarge`). Candidates mix these up constantly. ## Part 4 — the size suffix Sizes double along a fixed ladder: ``` large = 2 vCPU xlarge = 4 vCPU 2xlarge = 8 vCPU 4xlarge = 16 vCPU 8xlarge = 32 vCPU (and on up: 12xlarge, 16xlarge, 24xlarge, 32xlarge, 48xlarge) ``` Memory follows the family ratio, so within a family the whole machine scales linearly: m7g.large is 2 vCPU / 8 GiB, m7g.2xlarge is 8 vCPU / 32 GiB, m7g.4xlarge is 16 vCPU / 64 GiB. Below `large` you find `nano`, `micro`, `small` and `medium`, which exist mainly on burstable `t` families and a few Graviton families — not across the whole catalogue. A `.metal` suffix means no hypervisor at all: the whole physical host, used for nested virtualisation, licence-bound software, or workloads that need direct access to CPU features. ## Checking rather than guessing The grammar tells you the shape; the API tells you the truth. `aws ec2 describe-instance-types` returns exact vCPU counts, memory, supported architectures, and network performance for any type, and it is the right way to settle an argument or to build an allow-list of types for a fleet. ## What this buys you in an interview Being able to parse the name lets you answer the real question that follows — "what would you run this service on?" — out loud, in one sentence: family from the resource ratio, generation from what is current in the region, capability letters from architecture and network needs, size from measured usage plus headroom.