FP64 GPU server in 2026: H200 NVL, B300 or RTX PRO for double-precision HPC

Which GPUs still compute fast in FP64

For a server that runs double-precision codes in 2026, the card to choose among those we supply is the NVIDIA H200 NVL, which NVIDIA rates at 30 TFLOPS of FP64 and 60 TFLOPS on its FP64 Tensor Cores. The newer B300 gives most of that up: NVIDIA's HGX table lists 10 TFLOPS of FP64 for all eight B300 GPUs in a system. The NVIDIA RTX PRO 6000 Blackwell Server Edition runs FP64 at 1/64 of its FP32 rate, about 1.9 TFLOPS by our arithmetic, enough for code to run correctly and far too little for an FP64 solver.

The figures below come only from NVIDIA's product pages, its HGX table and its RTX Blackwell PRO architecture whitepaper, read on 24 October 2026. Where NVIDIA prints no FP64 figure, the value is our arithmetic and marked so.

NVIDIA H200 product page; NVIDIA HGX page (HGX B300: "FP64/FP64 Tensor Core 10 TFLOPS", HGX B200: 296 TFLOPS, eight GPUs each), divided by eight by us; RTX Blackwell PRO architecture whitepaper v1.0 ("The FP64 TFLOP rate is 1/64th the TFLOP rate of FP32 operations") applied to NVIDIA's FP32 ratings of 120 and 110 TFLOPS; B300 and B200 memory per GPU as in our DGX B300 articles. All read on 24 October 2026.

Read the FP64 column for general double-precision code and the Tensor column for dense matrix products that libraries such as cuBLAS route to the Tensor Cores. On the H200 NVL the two differ by a factor of two; NVIDIA prints one figure for both on the B300 and the B200. The RTX PRO 6000 Workstation Edition, rated at 125 TFLOPS of FP32, comes to about 1.95 TFLOPS of FP64 by the same rule. One H200 NVL therefore delivers in FP64 what sixteen RTX PRO 6000 Server Edition cards deliver, and three times what a whole eight-GPU B300 system lists.

Why the RTX PRO and B300 have so little FP64

NVIDIA's RTX Blackwell PRO architecture whitepaper (v1.0) gives the GB202 chip of the RTX PRO 6000 two FP64 cores per SM and states: "The small number of FP64 Cores are included to ensure any programs with FP64 code operate correctly." The card is built for FP32 graphics and simulation, AI in FP8 and FP4, and video. A code that declares its variables as double compiles and runs on it, at 1/64 of the FP32 speed. NVIDIA's product pages for the RTX PRO 6000 editions list no FP64 figure at all.

The B300 is NVIDIA's Blackwell Ultra GPU. Its HGX table keeps FP8, FP16 and TF32 Tensor rates equal to the B200 and raises dense FP4 from 72 to 108 PFLOPS for eight GPUs, while FP64 falls from 296 to 10 TFLOPS and INT8, given with sparsity, from 72 to 3 POPS. The gains are all in low-precision AI. For a team that runs both AI training and double-precision solvers, the B300 is the AI half of the answer only. Our comparison of DGX B300, DGX B200 and HGX B300 covers the other differences.

The H200 NVL is a Hopper card with full FP64 units, on the same architecture as the H100 that VASP's developers list among the data-centre GPUs they tested.

Codes that need FP64 and codes that do not

The precision a code computes in decides the card, and the developers' documentation says which it is. These statements were read on 24 October 2026; our guides to chemistry and materials research and CFD simulation quote them in full.

VASP wiki (OpenACC GPU port), GROMACS 2026.4 installation guide, AMBER GPU pages, LAMMPS GPU package page (30 September 2026), Ansys Fluent GPU hardware buying guide (9 April 2026), MACE README, NVIDIA's cuBLAS emulation blog (24 October 2025) for Quantum ESPRESSO. Card fit is our reading of these documents.

VASP's wiki advises against NVIDIA GPUs outside the data-centre cards it has tested, from the P100 to the H100: running VASP on other NVIDIA GPUs, "gaming" hardware for example, "is technically possible but not advisable", because "these GPUs are not well suited since they do not offer fast double-precision floating-point arithmetic (FP64) performance". The RTX PRO 6000 Server Edition is not on the tested list, and by our reading its FP64 rate of 1/64 of FP32 is the property that sentence describes. GROMACS describes a double-precision build as "slower, and not normally useful", and AMBER describes the SPFP model of pmemd.cuda as a hybrid single, double and fixed-precision model. Those molecular dynamics codes suit the RTX PRO 6000 Server Edition, with its 120 TFLOPS of FP32. For an in-house or research code, ask the developers whether the kernels compute in double, or count the double-precision instructions with a profiler; the answer separates the two card families by a factor of about 16.

We build servers for both kinds of code; tell us the codes, versions and the precision each runs in, and our team replies with a configuration for each.

Memory bandwidth for HPC codes

Many solvers move more data than they compute, and then bandwidth sets the speed. The H200 NVL reads its HBM3e at 4.8 TB/s, the RTX PRO 6000 Server Edition its GDDR7 at 1,597 GB/s, and NVIDIA's Blackwell Ultra technical blog gives the B300 8 TB/s per GPU. Dividing the FP64 rate by the bandwidth gives the operations per byte at which a kernel stops waiting for memory: about 6.3 on the H200 NVL, 1.2 on the RTX PRO 6000 Server Edition and 0.16 on the B300, by our arithmetic.

A vector update y = a·x + y does two operations for 24 bytes moved, about 0.08 per byte, so it waits for memory on all three, and on the B300 it would run fastest. As soon as a kernel does more than about 1.25 operations per double it loads, the B300 is bound by its FP64 units; above about two per double the H200 NVL is faster, and for kernels bound by compute on both the gap reaches 24 times, by our arithmetic. Real solvers mix both kinds of kernel; the roofline view of NVIDIA Nsight Compute shows where yours sit. Our comparison of GPU memory bandwidth covers the other cards.

FP64 emulation in cuBLAS on Blackwell

NVIDIA's developer blog of 24 October 2025 describes "the first release of FP64 matmul emulation" in cuBLAS with CUDA Toolkit 13.0 Update 2. It uses the Ozaki scheme on Blackwell INT8 Tensor Cores and names the RTX PRO 6000 Blackwell Server Edition and GB200 NVL72 as targets. By default, the blog says, "cuBLAS maintains accuracy equivalent to or better than native hardware": automatic dynamic precision analyses the inputs, picks the emulation parameters and falls back to native execution where emulation brings nothing.

On the RTX PRO 6000 Server Edition, NVIDIA reports speed-ups of up to 13 times on individual matrix shapes and, end to end, 1.5 times for Quantum ESPRESSO's Ausurf benchmark with automatic precision, and nearly three times with a fixed 39 mantissa bits, which trades accuracy for speed. The speed-up applies to matrix products called through cuBLAS. Every other FP64 operation in the code still runs at 1/64 of the FP32 rate, and the blog lists further BLAS level-3 and LAPACK routines only as future work. The blog names no B300 result, and NVIDIA's HGX table gives the B300 3 POPS of INT8 for eight GPUs.

Emulation is worth a test on an RTX PRO 6000 Server Edition when a code spends most of its time in DGEMM or ZGEMM and its results have been validated with emulation switched on. For a code that runs general FP64 kernels, it does not change the card choice.

Eight H200 NVL as an FP64 node

Our training node carries eight H200 NVL with NVLink bridges, two AMD EPYC processors, 2.3 TB of DDR5 ECC, eight NVMe drives and 400G networking, and it doubles as an FP64 compute node. The table sets it beside the smaller H200 NVL servers and the systems it is compared with.

Sums are our arithmetic from the first table. Wall power is our estimate (cards at 600 W; two AMD EPYC 9355 at 280 W for two and four cards, two EPYC 9555 at 360 W for eight cards; plus 15 per cent) as in our 2-, 4- and 8-card H200 NVL and RTX PRO 6000 Server Edition guides; DGX B300 memory from NVIDIA's HGX table and "~14 kW" from NVIDIA's DGX B300 product page, which our DGX B300 specs article quotes with the datasheet's 14.5 to 15.1 kW.

Per kilowatt at the wall, the eight-card H200 NVL server gives about 38 TFLOPS of FP64 by our arithmetic and the DGX B300 under 1. The eight cards form two NVLink domains of four: Dell's technical guide for the PowerEdge XE7740 requires two 4-way bridges for eight H200 NVL, at 900 GB/s per GPU within a domain, with PCIe 5.0 between the domains. VASP runs "a single MPI-rank per available GPU" and gains from NCCL, so an eight-card job places its ranks so that the heaviest traffic stays inside a domain. For many small jobs, MIG splits each H200 NVL into up to seven instances, 56 in the server, each with its own memory; our guide to H200 NVL MIG profiles lists them.

At about 6.3 kW the eight-card server needs a 400 V three-phase position per feed, and Dell limits the inlet temperature to 30 °C with eight H200 NVL in the XE7740. Two- and four-card servers suit a smaller team: four H200 NVL on one 4-way bridge give 120 TFLOPS of FP64 at about 3.4 kW, as our guide to the 4× H200 NVL server shows. Every H200 NVL includes a five-year NVIDIA AI Enterprise subscription, which starts 90 days after the card ships to the server maker.

B300 per project, for the AI side

We supply the DGX B300 and HGX B300 servers per project. Before we can quote, the distribution asks for the company, the project and the intended use, and we then quote for that project. The B300 belongs where large models are trained or served in FP8 and FP4 with 288 GB per GPU and NVLink across all eight GPUs. For a site that runs both, a B300 system for AI and an H200 NVL server for FP64 codes is a layout that gives each workload the silicon built for it. Our article on buying a DGX B300 in the EU describes how such an order runs, and the DGX B300 specs cover power and rack.

Recommendation by workload

Our recommendation from the FP64 rates, code documentation and emulation results above.

A university or research centre that serves both kinds of user often runs two partitions in one scheduler, H200 NVL for FP64 and RTX PRO 6000 Server Edition for FP32 and inference, as our guides to a university research lab server and to H200 NVL and RTX PRO 6000 in one platform describe.

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 AI and HPC servers to order with two, four or eight H200 NVL and their NVLink bridges, or with RTX PRO 6000 Server Edition cards, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component and the assembly covered by us. Operating system, driver, CUDA and a container runtime are installed on request. Quantities and lead time are confirmed in writing with the quote. Send us the codes, typical job sizes and the power at the rack, and we reply with a configuration and a written quote.

FAQ