Replacing NVIDIA A10 and A30 server GPUs: RTX PRO 4500 Server Edition, L4 or larger cards

Which card replaces the A10 and the A30

For an NVIDIA A10 in a server, the replacement is the RTX PRO 4500 Blackwell Server Edition: it fits the same single-slot, full-length slot, adds 8 GB of memory and a third more bandwidth, and brings FP8 and FP4 Tensor Cores and MIG. Where the slot is low-profile or the power budget is under 100 W, the L4 is the A10's successor. The NVIDIA A30 has no single successor: inference with MIG moves to the RTX PRO 6000 Blackwell Server Edition, double-precision work to the H200 NVL.

Neither card is being pushed out by NVIDIA. The A10 has full vGPU support, and the A30 is still in the current NVIDIA AI Enterprise support matrix. The case for replacing them is what the newer cards compute, how much memory they carry and how many tenants they isolate, which this comparison sets out card by card. Specifications are from NVIDIA's product pages and datasheets, read on 24 October 2026.

NVIDIA product pages and datasheets for the A10, A30, RTX PRO 4500 and RTX PRO 6000 Blackwell Server Edition, L4 and H200 NVL; NVIDIA vGPU types reference (29 September 2026) and vPC sizing guide. All read 24 October 2026. SE stands for Server Edition.

A10, A30, RTX PRO 4500 Server Edition and L4 compared

The four cards share a power class between 72 and 165 W, which is why they meet in the same servers.

NVIDIA A10 product page and datasheet (March 2022), A30 product page and datasheet (March 2022), RTX PRO 4500 Blackwell Server Edition product page (July 2026), L4 figures as in our L4 and RTX PRO 4500 Server Edition comparison; compute capability (CC) from NVIDIA's CUDA GPUs page. NVIDIA does not state whether the RTX PRO 4500 Server Edition's FP8 and FP4 figures assume sparsity. FHFL: full height, full length.

Read the bandwidth row first if your A30 serves one large model to few users at a time. Token generation for a single stream follows memory bandwidth, and here the A30's HBM2 is ahead of the RTX PRO 4500 Server Edition. The Blackwell card compensates with smaller weights: an FP8 or FP4 model reads fewer bytes per token than the 16-bit model the A30 runs. For bandwidth itself, the step up from an A30 is the RTX PRO 6000 Server Edition at 1,597 GB/s.

The media engines also differ more than the power class suggests. The A30 has no video encoder at all, so it was never a VDI card, and the A10 has one. The RTX PRO 4500 Server Edition has three encoders and three decoders.

FP8 and FP4: what the Ampere cards cannot compute

The A10 and A30 datasheets list Tensor Core rates for TF32, BF16, FP16, INT8 and INT4, and none for FP8. vLLM's FP8 documentation states that FP8 computation is supported on NVIDIA GPUs with compute capability 8.9 or higher, and that GPUs without native FP8 support, such as Turing and Ampere, "land on weight-only (W8A16) Marlin". An FP8 checkpoint on an A10 therefore halves the weights in memory, but the arithmetic stays 16-bit. The L4, with compute capability 8.9, falls on the other side of that line: of the two low-power options, it computes FP8 natively, while its 300 GB/s of bandwidth is half the A10's. For a model that is limited by bandwidth when serving one user, an L4 running FP8 weights reaches about the speed of an A10 running the same model in 16 bits, by the bandwidth ceiling; for many parallel requests, the native FP8 arithmetic is what it adds.

On the RTX PRO 4500 Server Edition, FP8 and NVFP4 run on the Tensor Cores. This matters most in the prompt phase, which is compute-bound, and for the number of parallel requests a card serves. StorageReview's test of 14 October 2026 (vLLM container 26.08, 32 concurrent streams) measured Ministral 3 14B in FP8 at 1,107 output tokens per second on one RTX PRO 4500 Server Edition, against 995 on an L40S and 364 on an L4. The same test ran Qwen3.8-27B in NVFP4 at 320 tokens per second on the 4500 Server Edition, a model that did not fit on the L4. StorageReview did not test the A10 or A30, and it did not meter power in that round.

Memory sets the model size. By our users-per-card rule, 90 per cent of the memory the driver reports minus 3 GiB for the runtime, an RTX PRO 4500 Server Edition leaves 25.7 GiB for weights and cache, the figure we use in our article on a server with eight RTX PRO 4500 Server Edition. An 8B model in BF16 needs about 16 GB at 2 bytes per parameter by our estimate, so it fits both generations; a 14B model in FP8, at about 14 GB, computes in FP8 on the Blackwell card and on the L4 but runs as W8A16 on an A10; and the 27B model in NVFP4 from StorageReview's test ran on the 32 GB card and did not fit on the 24 GB L4. Which model families run on 24 and 32 GB cards is covered in our guide to 30B MoE models on 24 GB GPUs.

We supply the RTX PRO 4500 Server Edition and the L4 as cards and in servers built to order; send us the models and request volumes your A10 or A30 servers carry today, and our team replies with a card count and configuration.

A10 in VDI hosts: profiles and desktops per card

The A10 is on NVIDIA's vGPU list since release 12.2 with full support, and NVIDIA's packaging and licensing guide recommends it for mid-range RTX vWS. Its vGPU types range from A10-1B to A10-24Q. The RTX PRO 4500 Server Edition, which NVIDIA's vPC sizing guide calls "the primary recommended GPU for NVIDIA vPC", takes more desktops per card at 2 and 3 GB and none at 1 GB.

NVIDIA Virtual GPU Types Reference, vGPU user guide (updated 29 September 2026), equal-size maximum per GPU for A10 and L4; NVIDIA vPC sizing guide (19 August 2026) for the RTX PRO 4500 Server Edition B profiles; NVIDIA RTX vWS sizing guide (19 August 2026) for light users per card on DC-4Q.

A host with eight A10 on 2 GB desktops holds 96; eight RTX PRO 4500 Server Edition hold 128 at 1,320 W of card power, against 1,200 W for the A10s. A pool that runs 1 GB desktops today either moves to 2 GB on the Blackwell card or to the L4. The other side of the change is the branch: the A10 runs on the long-term vGPU 19 branch, supported until July 2028, while the RTX PRO 4500 Server Edition is supported from vGPU 20.0 only, a production branch NVIDIA supports until March 2027. The full profile tables and hypervisor list are in our guide to RTX PRO 4500 Server Edition vGPU profiles, and replacing A40 and A16 vGPU hosts covers the move for larger desktop estates.

A30: MIG tenants and FP64 go to different cards

The A30 splits into up to four MIG instances of 6 GB, two of 12 GB or one of 24 GB, and two A30 can be joined by an NVLink bridge at 200 GB/s. Its vGPU support is compute only: NVIDIA's datasheet names NVIDIA AI Enterprise and Virtual Compute Server, and NVIDIA's vGPU list notes that support for GPUs with only C-series vGPUs was withdrawn in vGPU software 16.0 and continues with NVIDIA AI Enterprise. That is why the A30 is absent from the vGPU list and present in the AI Enterprise 8.2 support matrix of 2 September 2026; NVIDIA's vGPU lifecycle page, updated 2 October 2026, also names it among the GPUs that remain fully supported.

For MIG users, the RTX PRO 6000 Blackwell Server Edition keeps the four-tenant layout and gives each tenant 24 GB instead of 6 GB; our guide to RTX PRO 6000 MIG profiles lists the instance sizes. Two RTX PRO 4500 Server Edition cards give four isolated tenants of 16 GB at the same 330 W as two A30s. On Ampere, enabling MIG makes the driver attempt a GPU reset and the mode survives a reboot; from Hopper on it persists only while the driver is resident, a change for boot automation that our article on replacing A100 cards sets out.

For double precision, none of the Blackwell RTX PRO cards replaces an A30. NVIDIA's RTX PRO Blackwell architecture whitepaper runs FP64 at 1/64 of the FP32 rate, about 1.9 TFLOPS on the RTX PRO 6000 Server Edition by our arithmetic, and NVIDIA publishes no FP64 figure for the RTX PRO 4500 Server Edition. The H200 NVL delivers 30 TFLOPS of FP64 and 60 on its Tensor Cores, almost six times the A30's 5.2 and 10.3, with 141 GB of HBM3e at 4.8 TB/s. It draws up to 600 W and includes a five-year NVIDIA AI Enterprise subscription. Our FP64 GPU server guide compares the options for simulation codes.

Slots, power and cables in the existing server

The RTX PRO 4500 Server Edition looks like an A10 from the outside: single slot, 4.4 by 10.5 inches, passive cooling. The difference is the cable. The A10 takes a PCIe 8-pin auxiliary connector, the RTX PRO 4500 Server Edition one 16-pin (CEM5) connector, so an A10 server needs the maker's 16-pin GPU cables and a support listing for the new card. Supermicro lists up to eight RTX PRO 4500 Server Edition cards in its 3U, 4U and 5U models and up to six in 2U. Whether your machine type takes the card is settled by its maker's GPU support list, as our guide to single-slot server GPUs explains.

The card power changes little: eight A10 draw 1,200 W, eight RTX PRO 4500 Server Edition 1,320 W, which by our estimate puts an eight-card server at about 2.2 kW at the wall. The L4 needs no cable at all, since it is powered from the slot, and eight of them draw 576 W. An A30 server that moves to RTX PRO 6000 Server Edition cards goes from 165 W to as much as 600 W per card, which few A30 chassis were built for; plan a new server for that step.

Software: drivers, branches and containers

The Blackwell card needs a newer software stack than the Ampere cards. NVIDIA's RTX PRO 4500 Server Edition is on the vGPU list from 20.0, and compute capability 12.0 needs CUDA builds that target Blackwell, which CUDA 12.8 introduced, as our driver branches and CUDA versions guide sets out. A container built for an A10 on an older CUDA toolkit carries no code for the new chip; rebuild the images or move to current vLLM or NGC containers before the cards arrive. For bare-metal inference the MIG halves of the RTX PRO 4500 Server Edition need no licence; as vGPU in VMs, compute profiles need NVIDIA AI Enterprise per GPU, as our NVIDIA AI Enterprise licensing guide explains.

We supply the RTX PRO 4500 Server Edition, RTX PRO 6000 Server Edition, L4 and H200 NVL, assembled and tested before delivery in servers built to order. Tell us which A10 or A30 jobs you want to move, and we propose the card for each.

FAQ