Some GPUs support displayless and display-enabled modes but must be used in NVIDIA AI Enterprise deployments in displayless mode.

The GPUs listed in the following table support multiple display modes. As shown in the table, some GPUs are supplied from the factory in displayless mode, but other GPUs are supplied in a display-enabled mode.

A GPU that is supplied from the factory in displayless mode, such as the NVIDIA A40 GPU, might be in a display-enabled mode if its mode has previously been changed.

To change the mode of a GPU that supports multiple display modes, use the displaymodeselector tool, which you can request from the NVIDIA Display Mode Selector Tool page on the NVIDIA Developer website.

Because C-Series vCS vGPUs have large BAR memory settings, using these vGPUs has some restrictions on VMware ESXi.

Some GPUs require 64 GB or more of MMIO space. When a vGPU on a GPU that requires 64 GB or more of MMIO space is assigned to a VM with 32 GB or more of memory on ESXi , the VM’s MMIO space must be increased to the amount of MMIO space that the GPU requires.

For more information, refer to VMware Knowledge Base Article: VMware vSphere VMDirectPath I/O: Requirements for Platforms and Devices (2142307).

With ESXi 6.7 or later, no extra configuration is needed.

No extra configuration is needed.

The following table lists the GPUs that require 64 GB or more of MMIO space and the amount of MMIO space that each GPU requires.

In a Linux VM, if the requirements for using C-Series vCS vGPUs or GPUs requiring large MMIO space in pass-through mode are not met, the following error messages are written to the VM's dmesg log during installation of the NVIDIA AI Enterprise graphics driver:

NVRM: BAR1 is 0M @ 0x0 (PCI:0000:02:02.0)
[  90.823015] NVRM: The system BIOS may have misconfigured your GPU.
[  90.823019] nvidia: probe of 0000:02:02.0 failed with error -1
[  90.823031] NVRM: The NVIDIA probe routine failed for 1 device(s).

The releases in this release family of NVIDIA AI Enterprise support NVIDIA CUDA Toolkit 11.4.

To build a CUDA application, the system must have the NVIDIA CUDA Toolkit and the libraries required for linking. For details of the components of NVIDIA CUDA Toolkit, refer to NVIDIA CUDA Toolkit Release Notes for CUDA 11.4.

To run a CUDA application, the system must have a CUDA-enabled GPU and an NVIDIA display driver that is compatible with the NVIDIA CUDA Toolkit release that was used to build the application. If the application relies on dynamic linking for libraries, the system must also have the correct version of these libraries.

For more information about NVIDIA CUDA Toolkit, refer to CUDA Toolkit 11.4 Documentation.

is supported only on a subset of supported GPUs, VMware vSphere Hypervisor (ESXi) releases, and guest operating systems.

is supported on all supported GPUs, hypervisor software releases, and guest operating systems.

To support applications and workloads that are compute or graphics intensive, multiple vGPUs can be added to a single VM. The assignment of more than one vGPU to a VM is supported only on a subset of vGPUs and VMware vSphere Hypervisor (ESXi) releases.

Peer-to-peer CUDA transfers enable device memory between vGPUs on different GPUs that are assigned to the same VM to be accessed from within the CUDA kernels. NVLink is a high-bandwidth interconnect that enables fast communication between such vGPUs. Peer-to-Peer CUDA transfers over NVLink are supported only on a subset of vGPUs, VMware vSphere Hypervisor (ESXi) releases, and guest OS releases.

Unified memory is a single memory address space that is accessible from any CPU or GPU in a system. It creates a pool of managed memory that is shared between the CPU and GPU to provide a simple way to allocate and access data that can be used by code running on any CPU or GPU in the system. Unified memory is supported only on a subset of vGPUs and guest OS releases.

NVIDIA GPU Operator simplifies the deployment of NVIDIA AI Enterprise with software container platforms on immutable operating systems. An immutable operating system does not allow the installation of the NVIDIA AI Enterprise graphics driver directly on the operating system. NVIDIA GPU Operator is supported only on specific combinations of VMware vSphere Hypervisor (ESXi) release, container platform, and guest OS release.

If error-correcting code (ECC) memory is enabled on a physical GPU that does not have HBM2 memory, the amount of frame buffer that is usable by vGPUs is further reduced. All types of vGPU are affected, not just vGPUs that support ECC memory.

On all GPUs that support ECC memory and, therefore, dynamic page retirement, additional frame buffer is allocated for dynamic page retirement. The amount that is allocated is inversely proportional to the maximum number of vGPUs per physical GPU. All GPUs that support ECC memory are affected, even GPUs that have HBM2 memory or for which ECC memory is disabled.

The approximate amount of frame buffer that NVIDIA AI Enterprise reserves can be calculated from the following formula:

max-reserved-fb = vgpu-profile-size-in-mb÷16 + 16 + ecc-adjustments + page-retirement-allocation + compression-adjustment

max-reserved-fb

The maximum total amount of reserved frame buffer in Mbytes that is not available for vGPUs.

vgpu-profile-size-in-mb

The amount of frame buffer in Mbytes allocated to a single vGPU. This amount depends on the vGPU type. For example, for the T4-16Q vGPU type, vgpu-profile-size-in-mb is 16384.

ecc-adjustments

The amount of frame buffer in Mbytes that is not usable by vGPUs when ECC is enabled on a physical GPU that does not have HBM2 memory.

• If ECC is enabled on a physical GPU that does not have HBM2 memory ecc-adjustments is fb-without-ecc/16, which is equivalent to 64 Mbytes for every Gbyte of frame buffer assigned to the vGPU. fb-without-ecc is total amount of frame buffer with ECC disabled.
• If ECC is disabled or the GPU has HBM2 memory, ecc-adjustments is 0.

page-retirement-allocation

The amount of frame buffer in Mbytes that is reserved for dynamic page retirement.

• On GPUs based on the NVIDIA Maxwell GPU architecture, page-retirement-allocation = 4÷max-vgpus-per-gpu.
• On GPUs based on NVIDIA GPU architectures after the Maxwell architecture, page-retirement-allocation = 128÷max-vgpus-per-gpu

max-vgpus-per-gpu

The maximum number of vGPUs that can be created simultaneously on a physical GPU. This number varies according to the vGPU type. For example, for the T4-16Q vGPU type, max-vgpus-per-gpu is 1.

compression-adjustment

The amount of frame buffer in Mbytes that is reserved for the higher compression overhead in vGPU types with 12 Gbytes or more of frame buffer on GPUs based on the Turing architecture.

compression-adjustment depends on the vGPU type as shown in the following table.

vGPU Type	Compression Adjustment (MB)

For all other vGPU types, compression-adjustment is 0.