Accelerator-optimized machine family

The accelerator-optimized machine family feature NVIDIA's new Ampere A100 GPUs and is a new machine family available on Compute Engine. Machine types in this family are optimized for massively parallelized Compute Unified Device Architecture (CUDA) workloads, such as machine learning (ML) and high performance computing (HPC).

Machine Workloads
A2 machine types
  • CUDA-enabled ML training and inference
  • HPC
  • Massive parallelized computation

Machine types in the A2 series have a fixed number of A100 GPUs that offers up to 10x improvements in computing speed when compared to previous generation NVIDIA V100 GPUs.

In summary, the A2 machine series provides the following features:

  • Industry-leading NVLink scale that provides peak GPU to GPU NVLink bandwidth of 600 Gbps. For example, systems with 16 GPUs have an aggregate NVLink bandwidth of up to 9.6 Terabytes per second. These 16 GPUs can be used as a single high performance accelerator with unified memory space to deliver up to 10 petaFLOPS of compute power and up to 20 petaFLOPS of inference compute power that can be used for artificial intelligence, deep learning, and machine learning workloads.
  • Next generation NVIDIA A100 GPUs. A100 GPU provides 40 GB of GPU memory—ideal for large language models, databases, and HPC.
  • High performance network bandwidth of up to 100 Gbps.
  • Virtualization optimizations.
  • Optional local SSD support—you can get up to 3 TB of Local SSD with A2 machine types. This can be used as fast scratch disks or for feeding data into the A100 GPUs while preventing I/O bottlenecks.

A2 machine series

The A2 machine series has 12 to 96 vCPUs, and up to 1360 GB of memory. Each A2 machine type has a fixed number of GPUs attached. You can attach up to 257 TB of local storage to these machine types in this series for applications that require higher storage performance.

Machine types GPU count vCPUs* Memory (GB) Max number of persistent disks (PDs) Max total PD size (TB) Local SSD Maximum egress bandwidth (Gbps)
a2-highgpu-1g 1 12 85 128 257 Yes 24
a2-highgpu-2g 2 24 170 128 257 Yes 32
a2-highgpu-4g 4 48 340 128 257 Yes 50
a2-highgpu-8g 8 96 680 128 257 Yes 100
a2-megagpu-16g 16 96 1360 128 257 Yes 100

*A vCPU is implemented as a single hardware hyper-thread on one of the available CPU platforms.
Persistent disk usage is charged separately from machine type pricing.
Maximum egress bandwidth cannot exceed the number given. Actual egress bandwidth depends on the destination IP address and other factors. See Network bandwidth.

Limitations

  • You cannot use regional persistent disks with the A2 machine series.
  • The A2 machine series is only available in certain regions and zones.
  • a2-megagpu-16g machine types are not supported on Windows operating systems. When using Windows operating systems, choose a different machine type.
  • The A2 machine series is only available on the Cascade Lake platform.

What's next