RTX 4500 Ada and RTX 5000 Ada vs RTX A4500, A5000 and A6000: the upgrade from Ampere

Which Ada card takes the slot of an RTX A4500, A5000 or A6000

The NVIDIA RTX 4500 Ada Generation and the NVIDIA RTX 5000 Ada Generation fit the slot space of the Ampere cards NVIDIA RTX A4500, RTX A5000 and RTX A6000: all five are 4.4 in high, 10.5 in long and two slots wide. For an RTX A4500 or A5000 the card with the nearest memory size is the RTX 4500 Ada with 24 GB. The RTX 5000 Ada is the step to 32 GB and the only one of the two with vGPU. For an RTX A6000 with its 48 GB, either Ada card means less memory. We supply both Ada cards to companies across the EU.

NVIDIA datasheets 2895216 (October 2023) and 2898211 (August 2023); NVIDIA's RTX A4500 datasheet (May 2022) and its product pages for the RTX A5000 and RTX A6000.

The pairs are our reading by memory size and format. NVIDIA's documents for the two Ada cards name no previous-generation card to compare them with.

The five cards in NVIDIA's figures

NVIDIA datasheets 2895216 (October 2023) and 2898211 (August 2023); RTX A4500 datasheet (May 2022); NVIDIA's product pages for the RTX A5000 and A6000 and its professional graphics line card of March 2023. All five cards have ECC memory and measure 4.4 × 10.5 in, dual slot.

the same documents; connectors of the Ada cards from NVIDIA's quick start guide for RTX Ada cards; encoders from NVIDIA's Video Encode and Decode GPU Support Matrix of 16 October 2026; vGPU supported-GPU list of 2 October 2026.

The sheets of the two Ada cards are in the RTX 4500 Ada specs and the RTX 5000 Ada specs.

More compute in the same slots, less memory bandwidth

Core counts say little across two generations. The RTX 4500 Ada has fewer CUDA cores than the RTX A5000, 7,680 against 8,192, and a higher single-precision figure. NVIDIA's TFLOPS rows are the better basis, with the reservation that they are peak rates. By our arithmetic from those rows:

• RTX A4500 to RTX 4500 Ada: 67 per cent more;
• RTX A5000 to RTX 4500 Ada: 42 per cent more, at 20 W less;
• RTX A5000 to RTX 5000 Ada: 2.3 times;
• RTX A6000 to RTX 5000 Ada: 69 per cent more, at 50 W less;
• RTX A6000 to RTX 4500 Ada: about the same figure, 38.7 against 39.6 TFLOPS, at 210 W instead of 300 W.

Memory bandwidth moves the other way. The RTX 4500 Ada reads its 24 GB at 432 GB/s over a 192-bit interface, where the RTX A4500 has 640 GB/s over 320 bits and the RTX A5000 768 GB/s. That is about a third less than the RTX A4500 and 44 per cent less than the RTX A5000. The RTX 5000 Ada, at 576 GB/s, has a quarter less than the RTX A5000 and A6000.

One published measurement fits this picture. In Puget Systems' test of LLM inference on NVIDIA RTX GPUs of 22 August 2024, updated 4 October 2024 and run with llama.cpp and Phi-3-mini in 4-bit, the RTX 5000 Ada reached about 90 per cent of the token generation of the RTX A6000, with about 75 per cent of its memory bandwidth. Our reading: where work leans on memory throughput, the older card with the higher bandwidth can stay ahead, and where it mainly computes, NVIDIA's figures put the Ada cards ahead by the margins above. A test with your own workload settles which applies.

Power cables, from one 8-pin plug to the 16-pin connector

Each of the three Ampere cards has a single 8-pin connector. NVIDIA's datasheet gives the RTX A4500 one 8-pin PCIe connector, NVIDIA's quick start guide for the Ampere cards names an 8-pin PCIe cable for the RTX A5000, and the RTX A6000 has an 8-pin connector of the CPU type, for which the guide lists an adapter from two 8-pin PCIe cables in the box.

The Ada cards changed the connector. NVIDIA's quick start guide names a 16-pin PCIe CEM5 connector for the RTX 4500 Ada and the RTX 5000 Ada, lists an adapter from two 8-pin PCIe cables in the box and tells the installer to run two separate 8-pin cables from the power supply to it. A workstation that fed an RTX A4500 or A5000 through one 8-pin PCIe lead therefore needs a second lead for either Ada card. With an RTX A6000 the question is how it was fed: its own adapter took two 8-pin PCIe cables, the same pair the Ada adapter asks for.

In watts the steps are small: 10 W more from the RTX A4500 to the RTX 4500 Ada, 20 W less from the RTX A5000 to the RTX 4500 Ada, 20 W more from the RTX A5000 to the RTX 5000 Ada. NVIDIA publishes no power-supply wattage for the Ada cards. Our overview of RTX power connectors lists the connector of each card of this series.

If the Ada cards go into workstations that hold Ampere cards today, send us the workstation models, the cards in them and the free power leads, and we reply with a written quote within one business day.

NVLink existed on the A-series and not on Ada

NVIDIA's line card of March 2023 marks NVLink for the RTX A5000 and A6000, and the RTX A4500 datasheet names NVLink bridges; in each case the link joins two cards. The datasheets of the RTX 4500 Ada and the RTX 5000 Ada give NVLink as "No".

The practical meaning lies in memory pooling for GPU renderers. Chaos writes that V-Ray GPU can pool memory across pairs of GPUs that support NVLink and have a bridge installed, and that NVLink is found on the larger GPUs "previous to the Ada generation". Blender's manual says of Cycles that each GPU can only access its own memory, with NVIDIA GPUs connected by NVLink as the exception.

For an upgrade this has one consequence. A bridged pair of RTX A5000 cards that rendered scenes larger than 24 GB with pooled memory has no counterpart in a pair of Ada cards: with two Ada cards, the scene has to fit the memory of each card. A claim that two RTX 4500 Ada add up to 48 GB for a renderer has no basis in these documents. The Ada route to a larger scene is a larger card, 32 GB on the RTX 5000 Ada.

A language model spread over several cards is a separate subject; our article on one model on several GPUs covers it.

Video engines, vGPU and the sync board

Video. The Ampere cards have one encoder without AV1 encoding; the RTX A4500 datasheet lists one encoder and one decoder with AV1 decoding. The Ada cards double that: two encoders and two decoders each, AV1 encoding as well as decoding, and the entry "Unrestricted" for NVENC sessions in NVIDIA's support matrix of 16 October 2026.

Displays. All five cards have four full-size DisplayPort outputs. NVIDIA's documents print DisplayPort 1.4 for the RTX A4500 and A5000 and 1.4a for the RTX A6000 and the two Ada cards.

vGPU. NVIDIA's vGPU list of 2 October 2026 includes the RTX A5000 from release 12.2, the RTX A5500 from 14.0, the RTX A6000 from 12.0 and the RTX 5000 Ada from 16.1 (read 12 November 2026). The RTX A4500 and the RTX 4500 Ada are not on it. A host that shares RTX A5000 cards among virtual workstations therefore continues with the RTX 5000 Ada, and a move to the RTX 4500 Ada would end vGPU on that host. Which of the newer cards are on that list is set out in MIG and vGPU on RTX PRO and RTX Ada cards.

Sync board. NVIDIA's RTX PRO Sync page and the sync table in the 621.14 driver release notes both list the RTX A4500, A5000 and A6000. For each of the two Ada cards the documents differ: NVIDIA's datasheet and its RTX PRO Sync page list the card as compatible; the sync table in the 621.14 driver release notes does not. A synchronised display wall on Ampere cards should have this point confirmed for its driver branch before the cards are swapped.

FP8 and compute capability

NVIDIA's "CUDA GPU Compute Capability" page of 16 October 2026 lists 8.6 for the RTX A5000 and A6000 and 8.9 for both Ada cards; the RTX A4500 is not on that page. NVIDIA's NIM support matrix for visual generative AI of 20 October 2026 requires compute capability 8.0 for BF16 and 8.9 for FP8. In that matrix, FP8 profiles need the Ada generation or newer.

File sizes show what that is worth. Ministral 3 14B is a 25.97 GiB file in BF16, and its Instruct version in FP8 is a 14.65 GiB file, as read on Hugging Face on 10 November 2026. Our sizing rule gives a 24 GB card a budget of 18.5 GiB for weights plus KV cache and a 32 GB card 25.7 GiB: 0.9 times the memory the driver reports, less 3 GiB, with the reported memory scaled from the 95.6 GiB the driver shows on the 96 GB card. The scaling is an assumption, and the driver on the delivered card decides. The FP8 file fits the RTX 4500 Ada with 3.9 GiB left for context, while the BF16 file does not fit a 24 GB card.

An RTX A6000 with 48 GB needs another card

Against the RTX A6000 the RTX 5000 Ada has the higher compute figure and the lower board power, and 16 GB less memory. Where scenes or models fill more than 32 GB today, neither Ada card of this page is large enough.

The card we supply for that case is the NVIDIA RTX PRO 5000 Blackwell with 48 GB: GDDR7 with ECC at 1,344 GB/s, 14,080 CUDA cores and 65 TFLOPS in single precision, at the same 300 W and in the same 4.4 × 10.5 in dual-slot format as the RTX A6000. It uses the 16-pin connector as well. One difference matters for virtualised hosts: the RTX A6000 is on NVIDIA's vGPU list and the 48 GB RTX PRO 5000 is not. See the page to buy the RTX PRO 5000 Blackwell 48 GB and its 48 GB and 72 GB spec comparison. The step from the RTX A6000 to the 96 GB cards is covered in our RTX A6000 vs RTX PRO 6000 Blackwell comparison.

Recommendation by starting point

• RTX A4500 in a CAD or rendering seat. The RTX 4500 Ada: 4 GB more memory and 67 per cent more in NVIDIA's single-precision figure at 10 W more. Plan the second 8-pin lead. Ordering is on the page to buy the RTX 4500 Ada.
• RTX A5000, and 24 GB is enough. The RTX 4500 Ada, with the same memory size at 20 W less, provided the seat uses neither vGPU nor an NVLink bridge.
• RTX A5000 on a vGPU host, or work that outgrows 24 GB. The RTX 5000 Ada: buy the RTX 5000 Ada. The two Ada cards are compared in RTX 4500 Ada vs RTX 5000 Ada.
• A bridged pair used for pooled memory. One card with more memory in place of the pair.
• RTX A6000. The RTX 5000 Ada where 32 GB is enough, the RTX PRO 5000 Blackwell 48 GB where it is not.
• No tie to Ada. The Blackwell cards with 24 and 32 GB are set against these two Ada cards in RTX 4500 Ada vs RTX PRO 4500 and 4000 Blackwell and RTX 5000 Ada vs RTX PRO 4500 Blackwell; the Blackwell vs Ada overview shows every tier.

Owners of the single-slot RTX A4000 find their path in our RTX A4000 upgrade article. For a fleet with several of these starting points, describe which cards run in which workstations and what they are used for; the first consultation is free, and the quote names a card for each seat.

FAQ