The ND MI300X v5 series virtual machine (VM) is a new flagship addition to the Azure GPU family. It was designed for high-end Deep Learning training and tightly coupled scale-up and scale-out Generative AI and HPC workloads.
The ND MI300X v5 series VM starts with eight AMD Instinct MI300 GPUs and two fourth Gen Intel Xeon Scalable processors for a total 96 physical cores. Each GPU within the VM is then connected to one another via 4th-Gen AMD Infinity Fabric links with 128 GB/s bandwidth per GPU and 896 GB/s aggregate bandwidth.
ND MI300X v5-based deployments can scale up to thousands of GPUs with 3.2 Tb/s of interconnect bandwidth per VM. Each GPU within the VM is provided with its own dedicated, topology-agnostic 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand connection. These connections are automatically configured between VMs occupying the same virtual machine scale set, and support GPUDirect RDMA.
These instances provide excellent performance for many AI, ML, and analytics tools that support GPU acceleration "out-of-the-box," such as TensorFlow, Pytorch, and other frameworks. Additionally, the scale-out InfiniBand interconnect supports a large set of existing AI and HPC tools that are built on AMD’s ROCm Communication Collectives Library (RCCL) for seamless clustering of GPUs.
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.
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.
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