Replacing NVIDIA V100 after CUDA 13: H200 NVL or RTX PRO Server Edition, job by job

What CUDA 13 changed for the V100

CUDA 13.0 ended new development for the NVIDIA V100: its release notes state that "offline compilation and library support for these architectures have been removed", meaning Maxwell, Pascal and Volta, and cuFFT and cuSPARSE dropped every GPU older than Turing. A V100 server keeps running on CUDA 12.x and the R580 driver branch until June 2028. The replacement depends on the job: the H200 NVL for double-precision codes, the RTX PRO 6000 Blackwell Server Edition or RTX PRO 4500 Server Edition for AI work.

The release notes also say what still works: "The use of CUDA Toolkits through the 12.x series to build applications for these architectures will continue to be supported, but newer toolkits will be unable to target these architectures." A research team can therefore keep its V100 nodes on CUDA 12 containers. What it loses is the rest of the stack as it moves on. TensorRT 11.4.0, in its support matrix of 5 October 2026, "supports NVIDIA hardware with compute capability SM 7.5 or higher". vLLM's installation guide lists "compute capability 7.5 or higher" as its GPU requirement. NVIDIA's CUDA GPUs page, read on 24 October 2026, no longer lists the V100 at all.

Our recommendation from NVIDIA's datasheets, product pages and vGPU list, read on 24 October 2026.

How long a V100 server stays supported

Three NVIDIA documents set the end. NVIDIA's release notes for data-centre driver 580.178.04 still list "NVIDIA V100" among the supported Volta products, for bare metal and virtualisation; those for driver 595.91.07 list no V100 among their supported products. NVIDIA's driver table, last updated on 9 September 2026, designates R580 as a long term support branch with an end of life in June 2028, and R595 as a production branch to March 2027. A V100 server therefore stays on R580, with its security updates, until June 2028.

For virtualised V100s, NVIDIA's list of GPUs supported by vGPU, updated on 2 October 2026, puts all six V100 variants (SXM2 16 and 32 GB, PCIe 16 and 32 GB, FHHL 16 GB and V100S PCIe) in "Maintenance", with branch 19 as the last supported vGPU release and maintenance ending in July 2028. The V100 PCIe 32 GB is listed as last shipped in June 2021.

That gives a V100 cluster about twenty months from October 2026 on supported drivers, on a software stack frozen at CUDA 12. Our guide to GPU server lifecycle over five years explains the branch types, and NVIDIA driver branches and CUDA versions covers containers that need a newer CUDA than the driver.

V100 against the cards that replace it

NVIDIA V100 datasheet (January 2020); NVIDIA H200 product page, RTX PRO 6000 Server Edition product brief SP-12355-001_v02 and RTX PRO 4500 Server Edition page, as quoted in our FP64 and specification articles; RTX Blackwell PRO architecture whitepaper v1.0 for the FP64 ratio, about 1.9 TFLOPS being our arithmetic from 120 TFLOPS FP32. Read on 24 October 2026.

The V100 datasheet lists the PCIe card at 250 W with 7 TFLOPS of FP64, 14 of FP32 and 112 of tensor performance, the SXM2 module at 300 W with 7.8, 15.7 and 125, and the later V100S PCIe at 8.2, 16.4 and 130 with 1,134 GB/s. All carry HBM2 with ECC; the PCIe versions connect at PCIe Gen3 with 32 GB/s, the SXM2 module through NVLink with 300 GB/s.

FP64 codes: only the H200 NVL is faster

The V100 was a full double-precision card, and that is where the choice of successor narrows. The H200 NVL computes 30 TFLOPS of FP64, about 3.8 times a V100 SXM2, and 60 TFLOPS on its FP64 Tensor Cores for matrix products, by NVIDIA's H200 page. The RTX PRO Blackwell cards compute FP64 at 1/64 of their FP32 rate; NVIDIA's whitepaper says the few FP64 cores are there "to ensure any programs with FP64 code operate correctly". An RTX PRO 6000 Server Edition therefore reaches about 1.9 TFLOPS of FP64, a quarter of a V100.

A team that runs VASP, Quantum ESPRESSO, a CFD solver in double precision or its own FP64 code moves to H200 NVL. Teams whose codes run in mixed or single precision, GROMACS and AMBER by default, can use the RTX PRO 6000 Server Edition with its 120 TFLOPS of FP32. Our FP64 GPU server comparison lists the codes and their precision and covers FP64 emulation in cuBLAS on Blackwell, and our article on GPU servers for CFD applies it to Fluent.

Bandwidth matters for solvers as much as FP64 rate. The H200 NVL reads its memory at 4.8 TB/s, more than five times the V100's 900 GB/s, so a memory-bound kernel gains even where the FP64 units are not the limit.

AI training and inference: from 16 GB to 96 or 141 GB

A V100 trains in mixed precision with FP16 on its Tensor Cores, and its 16 or 32 GB set the batch size and the largest model a job can hold. The successors add BF16 and FP8; the Blackwell cards add FP4. vLLM's FP8 documentation of 24 September 2026 states that FP8 computation needs compute capability 8.9 or higher, which the H200 NVL (9.0) and the RTX PRO Blackwell cards (12.0) have.

On paper, NVIDIA's sparse rating of 1,671 TFLOPS of BF16 for the H200 NVL halves to 835.5 dense, about 6.7 times the 125 TFLOPS of a V100 SXM2 in FP16, by our arithmetic. Specifications are a ceiling; the gain a training job sees depends on how much of it runs in matrix products.

For serving language models, a V100 is now outside the main engines: vLLM requires compute capability 7.5. On the new cards our users-per-card rule (90 per cent of the memory the driver reports, less 3 GiB) gives 83.0 GiB per RTX PRO 6000 Server Edition and 123.4 GiB per H200 NVL for weights and KV cache. Llama 3.3 70B in FP8, 67.7 GiB of weights, fits one card of either; in FP16, about 141 GB of weights, it needs at least five 32 GB V100 before any cache, since FP8 does not run on Volta. Our guide to how many users one RTX PRO 6000 can serve has the figures per model.

For a department that ran notebooks and small training jobs on single V100s, the RTX PRO 4500 Server Edition gives 32 GB at 165 W in a single slot, and MIG splits it into two instances of 16 GB. Its 800 GB/s is slightly below the V100's 900 GB/s, so a bandwidth-bound job that already fitted a V100 gains little in speed; the gains are FP8, FP4 and lower power. We supply all three cards in servers built to order; send us the codes, their precision and typical job sizes, and our team replies with the card for each.

V100 SXM2 servers and NVLink

Eight-way V100 SXM2 systems linked their GPUs through NVLink at 300 GB/s per GPU, and training jobs written for them expect fast all-reduce. PCIe cards differ. The H200 NVL pairs two cards or joins four through NVLink bridges at 900 GB/s per GPU, so an eight-card H200 NVL server forms two NVLink domains of four, as our article on H200 NVL NVLink bridges explains. The RTX PRO 6000 Server Edition has no NVLink; NVIDIA's product brief lists it as "Not supported", and traffic between cards runs over PCIe 5.0. Data-parallel jobs on PCIe cards are possible, as our article on one model on several GPUs over PCIe or NVLink shows, but a job tuned for SXM2 NVLink belongs on H200 NVL bridges.

The GPU positions in an SXM2 server are sockets for SXM2 modules, so a V100 SXM2 system is replaced as a whole server.

Our arithmetic from the datasheet figures above, cards at their maximum board power of 300 W (V100 SXM2) and 600 W (H200 NVL, RTX PRO 6000 SE).

Two H200 NVL in one bridged pair match a full eight-way V100 SXM2 server in FP64 and hold more memory, at half the card power. Our estimate at the wall for the new servers is about 2.0 kW for two H200 NVL and 3.4 kW for four, each with two AMD EPYC 9355, and 6.3 kW for eight with two EPYC 9555, as in our guides to the two-card and eight-card H200 NVL server.

Fitting new cards into a V100 PCIe server

A V100 PCIe server was designed around 250 W passive cards on PCIe Gen3. The H200 NVL and the RTX PRO 6000 Server Edition draw up to 600 W each over a 16-pin cable and expect PCIe 5.0. Whether a given chassis accepts them is a question for its maker's support list, together with the power supplies, the cable and the airflow per slot; NVIDIA's product brief asks for 43 to 125 CFM through each RTX PRO 6000 Server Edition card, depending on inlet temperature. The RTX PRO 6000 Server Edition can be capped at 300 W, its minimum, which brings it close to a V100's power budget.

Our checklist for an H200 NVL in an existing server walks through the points. A server from the V100 years may also be out of warranty, and a new server avoids mixing new cards with old power supplies and firmware. We build on a customer's chassis only after a compatibility check of platform, power and cooling, and we say in advance if something will not work.

Pascal cards: the P40, P100 and P4

Pascal cards are further along. NVIDIA's vGPU end-of-support page, updated on 2 October 2026, gives the Tesla P40, P100, P4 and P6 branch 16 as their last vGPU release and July 2026 as the end of maintenance support, which has passed. CUDA 13.0 removed Pascal together with Volta, and vLLM and TensorRT 11.4 do not run on it. The same mapping by job applies: FP64 work to the H200 NVL, inference and training to the RTX PRO Server Editions, and low-profile slots to the NVIDIA L4, as our page on the NVIDIA T4 successor describes.

Moving the software

Three things change for code and containers. CUDA 12.8 or later is needed for Blackwell, which vLLM's installation guide states for the B200 and which applies to the RTX PRO Blackwell cards as well; the H200 NVL, a Hopper card, is covered by CUDA 13, which supports every architecture from Turing on. Containers pinned to old CUDA 11 or early CUDA 12 images are rebuilt on a current base, and NVIDIA's Blackwell compatibility guide states that binaries which "only include cubins" "need to be rebuilt to run on the Blackwell GPUs". Mixed precision code written for FP16 runs on the new cards as it is, and BF16 or FP8 are optional changes after validation. The V100 nodes can stay in the scheduler as a separate partition on CUDA 12 until June 2028, which our guide to a GPU server for a university research lab describes for mixed clusters.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build servers with two, four or eight H200 NVL and their NVLink bridges, or with RTX PRO 6000 Server Edition or RTX PRO 4500 Server Edition cards, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. Our training node on AI servers carries eight H200 NVL with NVLink bridges, two AMD EPYC processors, 2.3 TB of DDR5 ECC, eight NVMe drives and 400G networking. Every H200 NVL includes a five-year NVIDIA AI Enterprise subscription. Describe your V100 cluster and the codes it runs, and we reply with a replacement plan and a written quote.

FAQ