Replacing NVIDIA A40 and A16 vGPU hosts: support dates, successors and profile mapping

A40 and A16 vGPU hosts: what to replace them with

NVIDIA's vGPU support table lists the A40 in "Extended Support" until December 2027 and the A16 still in full support, with no end date yet. For office desktops on vPC, the replacement is the RTX PRO 4500 Blackwell Server Edition, which NVIDIA calls "the primary recommended GPU for NVIDIA vPC", or the L4 where 1 GB desktops or low-profile slots matter; for CAD and 3D users on RTX vWS, the NVIDIA RTX PRO 6000 Blackwell Server Edition or the L40S.

This article starts from an installed estate. It takes the A40 and A16 profiles your desktops run today, maps each to the profile on the successor card, counts desktops per card and per host before and after, and lists what changes in the hypervisor, the server and the licences. How to size a new VDI environment from zero is in our guide to VDI GPU sizing. All NVIDIA figures were read on 24 October 2026.

What NVIDIA's support table says about the A40 and A16

NVIDIA publishes one table with the support status of every GPU that runs its vGPU software. The rows for the cards in this article read as follows.

NVIDIA, GPUs supported by vGPU software, updated 2 October 2026, read 24 October 2026. "Last supported branch" and "maintenance officially ends" read "TBD" for all six cards. SE stands for Server Edition.

NVIDIA defines the phases on its vGPU lifecycle pages, also updated 2 October 2026. During Extended Full Support a GPU "is supported in the same way as during Phase I: Full Support", and that phase lasts three years; Maintenance Support follows, with bug fixes and security updates at NVIDIA's choice but no new features or technical support. NVIDIA's end-of-support table there gives December 2024 as the end of the A40's full support, the month the table above shows as last shipped, and December 2027 as the end of its extended full support; the last vGPU branch and the end of maintenance are not yet set. Older cards on the same page show the pattern that follows. The Quadro RTX 6000 and 8000, for example, are now in "Maintenance", with 19 as their last branch and maintenance ending in July 2028. Expect the A40 to get a last branch and a maintenance end date in a later update of the table, and plan new hosts before the end of 2027, while the A40 still gets new features and technical support, so that the change happens on your schedule.

The A16 is in a different position. It is still in full support, and NVIDIA's vPC sizing guide lists it as a "Multi-GPU board for high-density vPC deployments". An A16 estate is replaced when capacity, power or a hypervisor refresh calls for it; no support deadline forces the step yet. Neither the A40 nor the A16 is in our range, so this article compares the cards we supply as successors.

The successor cards side by side

The A40 is a 48 GB, 300 W passive card with one video encoder, and its natural successors are the L40S at the same memory size and the RTX PRO 6000 Server Edition at twice the memory.

NVIDIA A40 product page and datasheet (March 2022); NVIDIA L40S product page; RTX PRO 6000 Blackwell Server Edition product page and product brief SP-12355-001_v02. All read 24 October 2026.

The A16 is one board with four GPUs of 16 GB each, built for many small desktops. Its successors are single GPUs: the RTX PRO 4500 Server Edition, a 32 GB single-slot card, and the 24 GB low-profile L4.

NVIDIA A16 datasheet (February 2022); RTX PRO 4500 Blackwell Server Edition product page (July 2026); NVIDIA L4 product page as cited in our L4 vGPU profiles guide; profile sizes from NVIDIA's vGPU types reference (29 September 2026) and vPC sizing guide. FHFL: full height, full length.

The encoder row matters more on vGPU hosts than on inference servers, because every desktop's screen is encoded for the remote display protocol. At NVIDIA's maximum of 24 desktops of 2 GB per A40, all 24 share one encoder. On the successors the same arithmetic gives 8 per encoder on an L40S, 8 to 12 on an RTX PRO 6000 Server Edition and about 5 on an RTX PRO 4500 Server Edition. NVIDIA's guides set no session limit per encoder, so encoder load belongs in the pilot.

Profile mapping from A40 and A16 types

A vGPU type fixes the frame buffer a desktop gets. The mapping keeps each desktop at the same frame buffer where the new card offers it and shows how many fit per card. A40 and A16 figures are NVIDIA's maximums per GPU in equal-size mode.

NVIDIA Virtual GPU Types Reference, vGPU user guide (updated 29 September 2026) for A16, A40, L40S and L4 types; NVIDIA vPC sizing guide (updated 19 August 2026) for the Blackwell B profiles and the 48 MIG-backed desktops on VCF 9.1; NVIDIA RTX vWS sizing guide (19 August 2026) for users per GPU on DC-4Q and DC-8Q and the four DC-24Q per card. 4500 SE: RTX PRO 4500 Server Edition.

Two rows need a decision rather than a lookup. NVIDIA states that 1 GB profiles are not available on Blackwell GPUs, so a desktop on A16-1B either moves to an L4, which still has 1 GB types, or doubles its frame buffer to 2 GB on an RTX PRO 4500 Server Edition. NVIDIA's vPC guide adds that modern operating systems "require larger frame buffer sizes" and that 2B or 3B profiles are often needed, so the move to 2 GB is often the profile your users needed anyway. The workstation rows are ranges because NVIDIA's vWS guide counts users per GPU there, not profiles per GPU: an RTX PRO 6000 Server Edition holds twelve DC-8Q profiles by memory, but the guide puts 3 to 6 medium CAD users on it, since they share its compute.

The names change with the generation. A40 and A16 types carry the card name (A40-8Q), the Blackwell types a DC prefix and, for MIG-backed types, the slice count (DC-8Q time-sliced, DC-1-8Q on one 24 GB slice). The full lists are in our guides to RTX PRO 6000 Server Edition vGPU profiles, RTX PRO 4500 Server Edition vGPU profiles, L40S vGPU profiles and L4 vGPU profiles.

We supply the RTX PRO 4500 Server Edition, RTX PRO 6000 Server Edition, L40S and L4 in vGPU hosts built to order; send us your current profile list and peak sessions per pool, and our team replies with the mapping and a host count.

Desktops per host before and after

Two example hosts show the effect: four A16 boards in a 2U server, as NVIDIA's vPC sizing guide counts them, and a host with four A40 cards. The new layouts below use NVIDIA's per-card maximums; host totals are our arithmetic.

Desktops per card from NVIDIA's vGPU types reference and vPC sizing guide; four A16 boards per 2U and 192 users per 2U with MIG-backed, time-sliced vGPU from the vPC guide's maximum density table; card power from NVIDIA's product pages. Maximum densities, not recommended deployment points.

An A16 host on 2 GB desktops maps one to one onto an eight-card RTX PRO 4500 Server Edition host. Supermicro lists eight of these cards in its 3U, 4U and 5U models and up to six in 2U, so the new host is taller, and by our estimate it draws about 2.2 kW at the wall, as in our article on a server with eight RTX PRO 4500 Server Edition. An A16 host on 1 GB desktops holds twice as many users as any RTX PRO host can at 2 GB, so that pool needs either twice the RTX PRO hosts or L4 cards at 1 GB.

On the A40 side, four RTX PRO 6000 Server Edition cards carry a third more desktops than four A40 on time-sliced vGPU and twice as many with MIG-backed, time-sliced vGPU, which on vSphere needs VCF 9.1. The card power rises to as much as 2.4 kW, and four cards at 600 W come to about 3.4 kW at the wall by our estimate, or about 2.7 kW at a 450 W cap. For a CAD pool that runs A40-8Q today, the L40S is the like-for-like card: same profile sizes, same six desktops of 8 GB per card, three encoders instead of one.

Hypervisor, vGPU branch and guest images

A new card often means a new hypervisor build. NVIDIA's vGPU product support matrix for vSphere, updated 29 September 2026, lists VCF 9.1, VCF 9.0 and ESXi 8.0, with ESXi 8.0 only from Update 3 P06; it lists no vSphere 7 release. The two Blackwell Server Editions need VCF 9.0.1 or later on VCF 9.0, or ESXi 8 Update 3g or later on 8.0. vGPU on vSphere also requires the vSphere Foundation edition or an Enterprise Plus licence, per the same matrix. The A40 is listed on all three releases in displayless mode only.

The branch decides how long a host runs before its next driver change. As of October 2026, NVIDIA's vGPU documentation lists vGPU 19 on the R580 driver as the long-term support branch, supported until July 2028, and vGPU 20 on R595 as a production branch, supported until March 2027. The A40, A16, L40S, L4 and RTX PRO 6000 Server Edition all run on vGPU 19. The RTX PRO 4500 Server Edition is supported only from 20.0, so a pool with that card moves to the next branch within its first year in service. A mixed estate that should stay on one long-term branch during the transition can therefore keep A40 and A16 hosts and add RTX PRO 6000 Server Edition, L40S or L4 hosts on vGPU 19.

The guest side moves with the host. A desktop moves live only to a host with a free vGPU of the same type, so an A40-2B desktop cannot be migrated onto an RTX PRO host; it is shut down, given the new profile and started on the new host. Plan the guest driver update into the golden image at the same time and test it on one pilot host. Our guide to GPUs in VMware vSphere covers host preparation and vMotion of vGPU desktops.

Power, slots and cables in the existing chassis

Fitting the new cards into the old hosts is rarely the better path. The A40 and the A16 take an 8-pin CPU power connector; the L40S and both RTX PRO Server Editions take a 16-pin connector, and the RTX PRO 6000 Server Edition's product brief states that the card reads the power level from the 16-pin cable's sense pins. The RTX PRO 6000 Server Edition can draw twice the A40's 300 W per slot. Whether a given server takes the new card is settled by its maker's GPU support list for that exact machine type, together with the cable kit, airflow and power supplies. The checks are set out in our guide to upgrading an L40S server to the RTX PRO 6000 Server Edition, which applies to A40 hosts in the same way.

The L4 is the exception: a 72 W low-profile card powered from the slot, it fits servers that have no GPU power cables at all. NVIDIA's vPC table counts 16 L4 in a 2U server as a maximum; Lenovo lists 8 for its SR650 V3 and 10 for the SR650 V4, as our L4 profiles guide sets out.

The A40's NVLink bridge has no successor among these cards. Neither the L40S nor the RTX PRO Server Editions have NVLink, so a vWS user who relied on a bridged pair of A40 for 96 GB gets that memory in one RTX PRO 6000 Server Edition instead.

Licences: what carries over

NVIDIA's vGPU packaging and licensing guide, updated 24 March 2026, states that vApps, vPC and RTX vWS "are available on a per Concurrent User (CCU) model". The count follows peak concurrent sessions, not cards, so 500 concurrent desktops on A16 boards need the same 500 vPC licences on RTX PRO 4500 Server Edition cards. The edition follows the profile series: B types need vPC or vWS, Q types vWS, A types vApps. Compute (C-series) profiles are licensed separately per GPU through NVIDIA AI Enterprise. Our guide to NVIDIA vGPU licensing for RTX vWS, vPC and vApps covers terms and the move from one edition to another, and the NVIDIA licence server guide covers DLS and CLS.

Which successor for which pool

• A16 pool, office desktops on 2 or 3 GB: RTX PRO 4500 Server Edition, 16 or 10 desktops per card, eight cards per host in a 3U to 5U chassis.
• A16 pool that must stay at 1 GB per desktop, or servers with low-profile slots only: L4, 24 desktops per card at 1 GB.
• A40 pool for office desktops: RTX PRO 6000 Server Edition, 32 desktops per card, or 48 with MIG-backed vGPU on VCF 9.1.
• A40 pool for CAD and 3D on vWS: L40S for the same profiles at the same density, or RTX PRO 6000 Server Edition for larger profiles up to 96 GB and MIG isolation.
• A pool that should stay on the long-term vGPU 19 branch: RTX PRO 6000 Server Edition, L40S or L4, not the RTX PRO 4500 Server Edition.

We supply these four cards in GPU servers built to order, assembled and tested before delivery, with the vPC and vWS licences on the same quote. Describe your A40 or A16 pools and hypervisor version, and we size the new hosts and the spare capacity.

FAQ