Exam guide §2.1
Choosing a VM starts here: which family , which series , and the exact machine type (vCPU + memory shape). See the Compute Engine overview for how machine type fits alongside region/zone, boot disk, and access.
⛶ Machine family e.g. General-purpose Machine series e.g. E2 Machine type e.g. e2-standard-4 Pick from the general shape outward: a family narrows to a series, a series narrows to one predefined machine type.
Predefined types have fixed vCPU:memory ratios; custom types let you set them independently.
Network egress scales at 2 Gbps per vCPU ; instances with 2 or 4 vCPUs get up to 10 Gbps .
Theoretical maximum is 200 Gbps at 176 vCPUs (C3 machine series).
Each vCPU = one hardware hyper-thread on the underlying CPU platform.
? DECISION Which machine family?⛶ Accelerator-optimized A2 · A3 · G2 GPUs built in - CUDA ML training/inference, HPC. off the memory-per-vCPU axis Compute-optimized C2 · C2D · H3 Highest perf per core: CPU-bound work, gaming, HPC, per-core licensing. General-purpose E2 · N2 · N2D · Tau T2 Balanced, best price-performance. Default for standard workloads. Memory-optimized M1 · M2 · M3 Most RAM, lowest cost per GB. Large in-memory DBs (SAP HANA). less memory per vCPU more memory per vCPU Three families sit on one axis - memory per vCPU. Accelerator-optimized is a separate family with GPUs built in; of the other families, only N1 general-purpose allows attached GPUs. Pick this when: map the workload's bottleneck to a family: balanced = general-purpose; CPU-bound = compute-optimized; RAM-bound = memory-optimized; GPU/ML = accelerator-optimizedGeneral-purpose (E2, N2, N2D, Tau T2)
Best price-performance and the most flexible vCPU:memory ratios. Default choice for most workloads.
⛶ Processor Size Workload Applications E2 standard Intel/AMD (Google picks) 2-32 vCPUs, 0.5-8 GB/vCPU (Cost-optimized) Day-to-day computing at a lower cost • Web serving • App serving • Back office apps • Small-medium databases • Microservices • Virtual desktops • Development environments E2 shared-core Context-switched core 0.25-1 vCPU, 0.5-8 GB (Cost-optimized) Tiny, non-intensive workloads • Small websites & apps • Low-traffic services • Dev/test (e2-micro/small/medium) N2, N2D N2: Intel Cascade Lake/Ice Lake · N2D: AMD EPYC Milan/Rome N2 up to 128 vCPUs; N2D up to 224 (Balanced) Balanced price/performance across a wide range of VM shapes • Web serving • App serving • Back office apps • Medium-large databases • Cache • Media/streaming Tau T2D, T2A T2D: 3rd Gen AMD EPYC (x86) · T2A: Ampere Altra Arm T2D up to 60 vCPUs, 4 GB/vCPU (Scale-out optimized) Best performance/cost for scale-out workloads • Scale-out workloads • Web serving • Containerized microservices • Media transcoding • Large-scale Java applications
General-purpose series: balanced price-performance for standard workloads.
E2 VMs are the lowest price on Compute Engine, especially with committed-use discounts.N2/N2D are the most flexible types and support both committed and sustained use discounts.Tau T2D nodes work with GKE - add the T2D machine type to a node pool to optimize price-performance for containerized workloads. Compute-optimized (C2, C2D, H3)
Highest performance per core; for compute-intensive workloads.
⛶ Processor Size Workload Applications C2 Intel Cascade Lake, up to 3.8 GHz all-core turbo 4-60 vCPUs, up to 240 GB, up to 3 TB local SSD (Compute-intensive) Ultra high performance per core • Compute-bound workloads • High-performance web serving • Gaming (AAA game servers) • Ad serving • High-performance computing (HPC) • Media transcoding • AI/ML C2D 3rd Gen AMD EPYC Milan (largest last-level cache per core) 2-112 vCPUs, 4 GB/vCPU, up to 3 TB local SSD (Compute-intensive) Largest VM sizes • Memory-bound workloads • Gaming (AAA game servers) • High-performance computing (HPC) • High-performance databases • Electronic Design Automation (EDA) • Media transcoding H3 Intel Sapphire Rapids + Google IPU 88 cores, 352 GB DDR5 (Compute-intensive) HPC-focused • High-performance computing (HPC) • Electronic Design Automation (EDA)
Compute-optimized series: highest performance per core. Memory-optimized (M1, M2, M3)
Most memory per instance and the lowest cost per GB of memory. For high memory-to-vCPU ratios beyond what general-purpose high-memory types offer.
⛶ Processor Memory Workload Applications M1 - up to 4 TB (Ultra high-memory) Lowest cost per GB • Medium in-memory databases (SAP HANA) • In-memory analytics & business warehousing (BW) • Genomics analysis, SQL analysis services • Microsoft SQL Server and similar databases M2 - up to 12 TB (Ultra high-memory) Largest in-memory DBs • Large in-memory databases (SAP HANA) • In-memory analytics & business warehousing (BW) • Genomics analysis, SQL analysis services M3 Intel Ice Lake up to 128 vCPUs, up to 30.5 GB/vCPU (Ultra high-memory) Balanced memory + compute • OLAP and OLTP SAP workloads • Memory-intense Electronic Design Automation (EDA)
Memory-optimized series: most memory per instance, lowest cost per GB.
M1/M2 offer the lowest cost per GB of memory on Compute Engine.Up to 30% sustained use discount , plus committed-use discounts adding >60% savings on 3-year commitments.
Accelerator-optimized (A2, G2)
For workloads that require GPUs - massively parallel CUDA compute (ML, HPC).
⛶ GPU Size Workload Applications A2 up to 16 NVIDIA Ampere A100 (40 GB each) 12-96 vCPUs, up to 1,360 GB (GPU / HPC) Optimized for high-performance computing • CUDA-enabled ML training & inference • High-performance computing (HPC) • Massive parallelized computation • LLMs, databases G2 NVIDIA (Cascade Lake CPU) 4-96 vCPUs, up to 432 GB (GPU / HPC) Optimized for high-performance computing • CUDA-enabled ML training & inference • Video transcoding • Remote visualization workstations
Accelerator-optimized series: attached GPUs for CUDA ML and HPC. Custom machine types
When no predefined type fits, set vCPUs and memory independently. Costs slightly more than an equivalent predefined type, so only use custom when your vCPU:memory ratio falls between predefined types, or you need more RAM/CPU without jumping to the next larger (over-provisioned) predefined type. Supported on E2, N2, N2D.
Only 1 vCPU or an even number of vCPUs.
Memory between 1 GB and 8 GB per vCPU .
Total memory must be a multiple of 256 MB .
Need more than 8 GB per vCPU? Pay for extended memory to exceed the ceiling (e.g. N-series up to ~624 GB).