The ND A100 v4 series virtual machine(VM) is a new flagship addition to the Azure GPU family. These sizes are designed for high-end Deep Learning training and tightly coupled scale-up and scale-out HPC workloads.
The ND A100 v4 series starts with a single VM and eight NVIDIA Ampere A100 40GB Tensor Core GPUs. ND A100 v4-based deployments can scale up to thousands of GPUs with an 1.6 TB/s of interconnect bandwidth per VM. Each GPU within the VM is provided with its own dedicated, topology-agnostic 200 GB/s NVIDIA Mellanox HDR InfiniBand connection. These connections are automatically configured between VMs occupying the same Azure Virtual Machine Scale Set, and support GPU Direct RDMA.
Each GPU features NVLINK 3.0 connectivity for communication within the VM with 96 physical 2nd-generation AMD Epyc™ 7V12 (Rome) CPU cores behind them.
These instances provide excellent performance for many AI, ML, and analytics tools that support GPU acceleration 'out-of-the-box,' such as TensorFlow, Pytorch, Caffe, RAPIDS, and other frameworks. Additionally, the scale-out InfiniBand interconnect supports a large set of existing AI and HPC tools that are built on NVIDIA's NCCL2 communication libraries for seamless clustering of GPUs.
Host specifications
| Part |
Quantity
Count Units |
Specs
SKU ID, Performance Units, etc. |
| Processor |
96 vCPUs |
AMD EPYC 7V12 (Rome) [x86-64] |
| Memory |
900 GiB |
|
| Local Storage |
1 Disk |
6000 GiB |
| Remote Storage |
32 Disks |
80000 IOPS
800 MBps |
| Network |
8 NICs |
24000 Mbps |
| Accelerators |
8 GPUs |
Nvidia A100 GPU (40GB) |
Feature support
Premium Storage: Supported
Premium Storage caching: Supported
Live Migration: Not Supported
Memory Preserving Updates: Not Supported
Generation 2 VMs: Supported
Generation 1 VMs: Not Supported
Accelerated Networking: Supported
Ephemeral OS Disk: Supported
Nested Virtualization: Not Supported
Sizes in series
vCPUs (Qty.) and Memory for each size
| Size Name |
vCPUs (Qty.) |
Memory (GB) |
| Standard_ND96asr_v4 |
96 |
900 |
VM Basics resources
Local (temp) storage info for each size
| Size Name |
Max Temp Storage Disks (Qty.) |
Temp Disk Size (GiB) |
| Standard_ND96asr_v4 |
1 |
6000 |
Storage resources
Table definitions
- 1Temp disk speed often differs between RR (Random Read) and RW (Random Write) operations. RR operations are typically faster than RW operations. The RW speed is usually slower than the RR speed on series where only the RR speed value is listed.
- Storage capacity is shown in units of GiB or 1024^3 bytes. When you compare disks measured in GB (1000^3 bytes) to disks measured in GiB (1024^3) remember that capacity numbers given in GiB may appear smaller. For example, 1023 GiB = 1098.4 GB.
- Disk throughput is measured in input/output operations per second (IOPS) and MBps where MBps = 10^6 bytes/sec.
- To learn how to get the best storage performance for your VMs, see Virtual machine and disk performance.
Remote (uncached) storage info for each size
| Size Name |
Max Remote Storage Disks (Qty.) |
Uncached Disk IOPS |
Uncached Disk Speed (MBps) |
| Standard_ND96asr_v4 |
32 |
80000 |
800 |
Storage resources
Table definitions
- 1Some sizes support bursting to temporarily increase disk performance. Burst speeds can be maintained for up to 30 minutes at a time.
- 2Special Storage refers to either Ultra Disk or Premium SSD v2 storage.
- Storage capacity is shown in units of GiB or 1024^3 bytes. When you compare disks measured in GB (1000^3 bytes) to disks measured in GiB (1024^3) remember that capacity numbers given in GiB may appear smaller. For example, 1023 GiB = 1098.4 GB.
- Disk throughput is measured in input/output operations per second (IOPS) and MBps where MBps = 10^6 bytes/sec.
- Data disks can operate in cached or uncached modes. For cached data disk operation, the host cache mode is set to ReadOnly or ReadWrite. For uncached data disk operation, the host cache mode is set to None.
- To learn how to get the best storage performance for your VMs, see Virtual machine and disk performance.
Network interface info for each size
| Size Name |
Max NICs (Qty.) |
Max Bandwidth (Mbps) |
| Standard_ND96asr_v4 |
8 |
24000 |
Networking resources
Table definitions
- Expected network bandwidth is the maximum aggregated bandwidth allocated per VM type across all NICs, for all destinations. For more information, see Virtual machine network bandwidth
- Upper limits aren't guaranteed. Limits offer guidance for selecting the right VM type for the intended application. Actual network performance will depend on several factors including network congestion, application loads, and network settings. For information on optimizing network throughput, see Optimize network throughput for Azure virtual machines.
- To achieve the expected network performance on Linux or Windows, you may need to select a specific version or optimize your VM. For more information, see Bandwidth/Throughput testing (NTTTCP).
Accelerator (GPUs, FPGAs, etc.) info for each size
| Size Name |
Accelerators (Qty.) |
Accelerator-Memory (GB) |
| Standard_ND96asr_v4 |
8 |
320 |
List of all available sizes: Sizes
Pricing Calculator: Pricing Calculator
Information on Disk Types: Disk Types
Next steps
Take advantage of the latest performance and features available for your workloads by changing the size of a virtual machine.
Utilize Microsoft's in-house designed ARM processors with Azure Cobalt VMs.
Learn how to Monitor Azure virtual machines.