The NP-series virtual machines are powered by Xilinx U250 FPGAs for accelerating workloads including machine learning inference, video transcoding, and database search & analytics. NP-series VMs are also powered by Intel Xeon 8171M (Skylake) CPUs with all core turbo clock speed of 3.2 GHz.
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.
1Some sizes support bursting to temporarily increase disk performance. Burst speeds can be maintained for up to 30 minutes at a time.
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
Size Name
Accelerators (Qty.)
Accelerator-Memory (GB)
Standard_NP10s
1
64
Standard_NP20s
2
128
Standard_NP40s
4
256
Frequently asked questions
Q: What's the difference between Xilinx U250 and the AMD Alveo U250?
A: AMD Acquired Xilinx and renamed their FPGA line to Alveo. They are identical and use the same drivers, but the original Xilinx page redirects to AMD's new site.
Q: How to request quota for NP VMs?
A: Follow this page Increase VM-family vCPU quotas. NP VMs are available in East US, West US2, SouthCentral US, West Europe, SouthEast Asia, Japan East, and Canada Central.
Q: What version of Vitis should I use?
A: Xilinx recommends Vitis 2022.1, you can also use the Development VM marketplace options (Vitis 2022.1 Development VM for Ubuntu 18.04, Ubuntu 20.04, and CentOS 7.8)
Q: Do I need to use NP VMs to develop my solution?
A: No, you can develop on-premises and deploy to the cloud. Make sure to follow the attestation documentation to deploy on NP VMs.
Q: What shell version is supported and how can I get the development files?
A: The FPGAs in Azure NP VMs support Xilinx Shell 2.1 (gen3x16-xdma-shell_2.1). See Xilinx Page Xilinx/Azure with Alveo U250 to get the development shell files.
Q: Which file returned from attestation should I use when programming my FPGA in an NP VM?
A: Attestation returns two xclbins, design.bit.xclbin and design.azure.xclbin. Use design.azure.xclbin.
Q: Where should I get all the XRT / Platform files?
Q: What are the differences between on-premises FPGAs and NP VMs?
A: - Regarding XOCL/XCLMGMT:
On Azure NP VMs, only the role endpoint (Device ID 5005), which uses the XOCL driver, is present.
In on-premises FPGAs, both the management endpoint (Device ID 5004) and role endpoint (Device ID 5005), which use the XCLMGMT and XOCL drivers respectively, are present.
- Regarding XRT:
On Azure NP VMs, the XDMA 2.1 platform only supports Host_Mem(SB).
To enable Host_Mem(SB) (up to 1-Gb RAM): sudo xbutil host_mem --enable --size 1g
To disable Host_Mem(SB): sudo xbutil host_mem --disable
Starting on XRT2021.1:
On-premises FPGA in Linux exposes
M2M data transfer.
This feature isn't supported in Azure NP VMs.
Q: Can I run xbmgmt commands?
A: No, on Azure VMs there's no management support directly from the Azure VM.
Q: Do I need to load a PLP?
A: No, the PLP is loaded automatically for you, so there's no need to load via xbmgmt commands.
Q: Does Azure support different PLPs?
A: Not at this time. We only support the PLP provided in the deployment platform packages.
Q: How can I query the PLP information?
A: Need to run xbutil query and look at the lower portion.
Q: Do Azure NP VMs support FPGA bitstreams with Networking GT Kernel connections?
A: No. The FPGA Attestation service performs a series of validations on a design checkpoint file and generates an error if the user's application contains connections to the FPGA card's QSFP networking ports.
Azure HPC is a purpose-built cloud capability for HPC & AI workload, using leading-edge processors and HPC-class InfiniBand interconnect, to deliver the best application performance, scalability, and value. Azure HPC enables users to unlock innovation, productivity, and business agility, through a highly available range of HPC & AI technologies that can be dynamically allocated as your business and technical needs change. This learning path is a series of modules that help you get started on Azure HPC - you