RTX A6000 vs NVIDIA RTX PRO 6000 Blackwell: which edition replaces the A6000

Which RTX PRO 6000 edition replaces the A6000

The NVIDIA RTX A6000 is replaced card for card by the RTX PRO 6000 Blackwell Max-Q Workstation Edition: both draw 300 W, cool actively, measure 4.4 by 10.5 inches and take two slots, and the Max-Q doubles memory to 96 GB and raises bandwidth from 768 to 1,792 GB/s. The 600 W Workstation Edition is the faster card for one or two per tower, and the passive Server Edition replaces A6000 cards in rack servers and vGPU hosts. All three are supplied by us; the RTX A6000 itself is not.

The A6000, an Ampere card, is still current at NVIDIA: the vGPU list of 2 October 2026 shows it in "Full Support" with its date last shipped as "Active". Owners therefore upgrade for memory, speed and formats, without a support deadline. Owners of the older Turing Quadro RTX 6000 and 8000 do have a date, covered further down.

Our recommendation from NVIDIA's RTX A6000 datasheet (May 2022), RTX PRO 6000 product pages and datasheets and NVIDIA's vGPU list, read on 24 October 2026.

Our comparison of the RTX PRO 6000 Blackwell and the RTX 6000 Ada covers the step from Ada; this page starts one generation earlier, where the differences in memory type, formats and NVLink are larger.

RTX A6000 and the three RTX PRO 6000 editions

NVIDIA RTX A6000 datasheet (May 2022); NVIDIA professional desktop GPU overview; RTX PRO 6000 Max-Q datasheet (June 2026); RTX PRO 6000 Server Edition product brief SP-12355-001_v02; NVIDIA Video Codec SDK support matrix; NVIDIA vGPU list of 2 October 2026; all read on 24 October 2026.

On the A6000 datasheet the 309.7 TFLOPS of tensor performance are "effective" figures "using the new sparsity feature", with no precision named, and NVIDIA gives the RTX PRO 6000 its tensor rate in FP4 (3,511 AI TOPS for the Max-Q, 4,000 for the Workstation Edition), so the two do not compare directly. FP32 and bandwidth do: the Max-Q has 2.8 times the A6000's FP32 rate and 2.3 times its bandwidth, at the same 300 W, by our arithmetic.

96 GB on one card instead of a bridged pair

The A6000 is the last card in this line with NVLink. Its datasheet lists a 2-way bridge at 112.5 GB/s bidirectional that "connects two NVIDIA RTX A6000 GPUs", for up to 96 GB of combined memory. The RTX PRO 6000 has no NVLink in any edition; NVIDIA's Server Edition brief lists it as "Not supported", and traffic between cards runs over PCIe 5.0 x16.

For most A6000 pairs this loss matters less than it sounds, because one RTX PRO 6000 holds 96 GB on a single card. Work that needed the bridge to fit a scene or a model now fits one GPU with no split at all, at 300 W on the Max-Q against 600 W for two A6000. Where a model needs more than 96 GB, two RTX PRO 6000 split it over PCIe with tensor or pipeline parallelism, as our article on the RTX PRO 6000 without NVLink explains.

Language models: what 48 GB and 96 GB hold

Our users-per-card rule takes 90 per cent of the memory the driver reports and subtracts 3 GiB for CUDA graphs and activations; the rest after the weights holds the KV cache. GDDR6 with ECC switched on gives up 6.25 per cent of its capacity, so an A6000 has 40.0 GiB with ECC off and 37.3 GiB with ECC on. On GDDR7 NVIDIA states ECC costs no capacity, and an RTX PRO 6000 has 83.0 GiB.

Our estimate by the users-per-card rule: concurrent conversations of 8,192 tokens, 16-bit KV cache unless stated (2.0 GiB per conversation for Qwen3-32B, 2.5 GiB for Llama 3.3 70B); weights 32.0 GiB and 67.7 GiB. With ECC on, one A6000 holds 2 Qwen3-32B conversations, and an A6000 pair 21 of Qwen3-32B and 2 of Llama 3.3 70B.

The cache column for the A6000 stays at 16 bits because of the format. vLLM's FP8 documentation of 24 September 2026 states that "FP8 computation is supported on NVIDIA GPUs with compute capability >= 8.9", while "Turing/Ampere GPUs are supported for W8A16 (weight-only FP8) utilizing Marlin kernels". The A6000 is compute capability 8.6: it can load an FP8 checkpoint, store the weights in 8 bits and compute in 16. The RTX PRO 6000 computes in FP8 and in NVFP4, the 4-bit format of Blackwell Tensor Cores, which lets Llama 3.3 70B run in 39.8 GiB on one card. Our guide to how many users one RTX PRO 6000 can serve has the figures for more models.

Bandwidth sets the speed of each conversation. For Llama 3.3 70B in FP8, 70.6 GB read per token, a bridged A6000 pair reaches at most about 22 tokens per second (2 × 768 GB/s ÷ 70.6 GB) and one Max-Q about 25, by our arithmetic. These are ceilings, and the pair also loses time exchanging partial results after every layer. One card with the whole model avoids that exchange. We supply all three editions as cards and in workstations built to order; send us the models, context and number of users you plan for, and our team replies with the edition and the count.

Fitting a Max-Q where an A6000 was

The Max-Q has the A6000's footprint: NVIDIA's overview gives 4.4 by 10.5 inches, dual slot, for the Max-Q, the size the A6000 datasheet lists. Two things differ. The A6000 takes "1x 8-pin CPU" power, while the Max-Q datasheet lists one PCIe CEM5 16-pin connector, so the power supply needs a 16-pin cable approved by the workstation maker. And the Max-Q is a PCIe 5.0 card; in a workstation with PCIe 4.0 slots it runs at Gen4, as StorageReview reported for the second Max-Q slot of a Dell Precision 7875 in its HP Z8 Fury G6i review of 23 June 2026.

Several workstation makers list the Max-Q in four-card configurations: Lenovo's ThinkStation PX specification of 9 October 2026 "supports up to four NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition (300W)", HP lists up to four for the Z8 Fury G6i, and Dell's Precision 7960 configurator offers one to four. NVIDIA's Max-Q page states that it "enables up to four GPUs in a single system". A four-card A6000 tower therefore maps to a four-card Max-Q tower, with 384 GB instead of 192 GB, as our page on the four-card Max-Q workstation and our comparison of four RTX 6000 Ada or four Max-Q describe.

The Workstation Edition is a different size: 5.4 by 12.0 inches, 600 W and double flow-through cooling that exhausts into the case. Our product page advises one card comfortably and two at most per tower, because the first card's exhaust feeds the second. It is the choice for a single-GPU workstation where speed matters more than density; our edition comparison gives the measured gap between the two desktop editions, 5 to 14 per cent in Puget Systems' content-creation tests.

A6000 in rack servers and vGPU hosts

The A6000 is an active card. Where it sits in a rack server, its successor is the RTX PRO 6000 Server Edition: passive, configurable from 300 to 600 W, and built for the airflow of a server. The Max-Q belongs in a rack server only where the server maker lists it for that model, and many makers do not, as our article on the RTX PRO 6000 Max-Q in a rack server shows.

For virtual workstations the choice is narrower. NVIDIA's vGPU list carries the RTX A6000 from release 12.0 and the RTX PRO 6000 Server Edition from 19.0; the Max-Q and the Workstation Edition are not on it. An A6000 host serving RTX vWS desktops therefore moves to the Server Edition, whose profiles our guide to RTX PRO 6000 Server Edition vGPU profiles lists. Both cards switch their display outputs off under vGPU; the A6000 datasheet notes that its "display ports are not active when using vGPU software".

MIG is new for A6000 owners. The A6000 datasheet names no MIG, while every RTX PRO 6000 edition splits into up to four instances of 24 GB, two of 48 GB or one of 96 GB, on Linux with a recent driver and the display outputs switched off, as our guide to RTX PRO 6000 MIG profiles describes. A team that shared one A6000 by taking turns can give four people a 24 GB GPU each on one Max-Q.

Video, rendering and display

The A6000 has one 7th-generation encoder without AV1 encoding and two decoders with AV1 decode, by its datasheet and NVIDIA's codec matrix. Each RTX PRO 6000 edition has four 9th-generation encoders with AV1, and the Max-Q datasheet lists four decoders. For editing and streaming workstations that is four times the encode engines. The Max-Q has four DisplayPort 2.1b outputs against the A6000's four DisplayPort 1.4a.

For rendering, the 96 GB matter as much as the FP32 rate: scenes that had to be split or simplified for 48 GB fit whole. Our article on the GPU render server with RTX PRO 6000 covers renderers and card counts.

Quadro RTX 6000 and 8000: the Turing cards

The Quadro RTX 8000 datasheet lists 48 GB of GDDR6 at 672 GB/s, 16.3 TFLOPS of FP32, 295 W, PCIe 3.0 x16 and an NVLink bridge at 100 GB/s. NVIDIA's vGPU list of 2 October 2026 puts the Quadro RTX 6000 and 8000 in "Maintenance", with branch 19 as the last vGPU release and maintenance ending in July 2028; the RTX 8000 is listed as last shipped in May 2022. CUDA 13.0 still supports Turing, compute capability 7.5, so current toolkits still target these cards, and vLLM runs on them with FP8 as weight-only.

For an RTX 8000 owner the Max-Q is the same move at larger scale: about 6.7 times the FP32 rate and 2.7 times the bandwidth at almost the same power, by our arithmetic from the two datasheets. Virtualised RTX 6000 and 8000 hosts face July 2028 and move to the Server Edition.

Drivers and software after the swap

The RTX PRO 6000 is compute capability 12.0 and needs CUDA 12.8 or later and NVIDIA's open kernel modules on Linux, which our article on Blackwell RTX software support covers with framework versions. An old workstation image with a driver from the A6000's early years will not run it. Applications that ship only compiled code for Ampere need a build that includes Blackwell; NVIDIA's Blackwell compatibility guide states that binaries which "only include cubins" "need to be rebuilt". Where an A6000 and an RTX PRO 6000 run side by side in one machine during the change, both need a driver that lists both cards, as our guide on mixing Ada and Blackwell GPUs explains for the generation in between.

What we supply

We supply the RTX PRO 6000 Blackwell as Max-Q, Workstation Edition and Server Edition, as single cards or in workstations and servers built to order, on one EU contract and invoice with manufacturer warranty on every component. Quantities and lead time are confirmed in writing with the quote. When the cards go into a workstation or server you already own, we check platform, power and cooling first and say in advance if something will not work. Tell us the workstation model and how many A6000 cards it holds, and we reply with the edition and the parts that fit.

FAQ