4 min readAceCAST

WRF GPU vs CPU cost: the arithmetic, shown line by line

One published 2023 benchmark and two Azure list prices, worked line by line: what a GPU VM costs against one CPU VM and against sixteen, and what the sums leave out.

Stacked bar chart of elapsed time in seconds, split into initialization, input and output, and compute, for WRF on two, four and eight Azure HBv3 CPU nodes and for AceCAST on one NDm A100 v4 GPU node. The GPU bar is the shortest.

First published November 11, 2026 · Updated November 16, 2026

If you run WRF, the Weather Research and Forecasting model maintained by NCAR, someone will ask whether GPUs would make it cheaper. This article works the WRF GPU vs CPU cost question in the open. It uses one published benchmark and two public cloud list prices, and it shows every sum so that you can redo it with your own figures.

A note on scope: the benchmark measured AceCAST 2.1 and 3.0.1, both earlier releases.

The inputs: one benchmark and two list prices

The benchmark. On April 28, 2023, the NVIDIA Technical Blog published Microsoft and TempoQuest Accelerate Wind Energy Forecasts with AceCAST. The runs were made on Microsoft Azure. WRF ran on CPU virtual machines (VMs) of type Standard HB120rs_v3, each with 120 AMD EPYC Milan-X cores. AceCAST, which runs WRF on NVIDIA GPUs, ran on a Standard_ND96amsr_A100_v4 VM with 8 NVIDIA A100 GPUs. The article reports:

  • "AceCAST on one node achieved ~9x acceleration compared to CPU-based WRF on one node."
  • "18 CPU nodes are necessary to achieve a similar performance to one GPU node."
  • For the configuration its authors found optimal, WRF on 16 CPU VMs against AceCAST on 1 GPU VM: "AceCAST runs 7% faster at 75% lower cost than CPU-based WRF."

The prices. Azure list prices from the Azure Retail Prices API for the East US region, Linux, pay-as-you-go, on November 11, 2026:

  • One CPU VM (Standard_HB120rs_v3): $3.60 per hour.
  • One GPU VM (Standard_ND96amsr_A100_v4): $32.77 per hour.

The 2023 Azure study reported about 9× speed for AceCAST 2.1 against one CPU node, and similar performance to 18 CPU nodes; our cost example uses its separate AceCAST 3.0.1 comparison against 16 CPU nodes. One GPU node contained eight GPUs.

The arithmetic, line by line

The cost of a run is the hourly price multiplied by the hours the run takes. Speed matters to cost because a faster run is billed for fewer hours.

ComparisonCPU sideGPU sideResult
Hourly price, one VM each$3.60$32.7732.77 ÷ 3.60 = 9.10. The GPU VM costs about 9.1 times as much per hour.
Same forecast, one VM each, GPU at about 9 times the speed. Example: a job of 9 hours on the CPU VM.9 hours × $3.60 = $32.401 hour × $32.77 = $32.7732.77 ÷ 32.40 = 1.011. The GPU run costs about 1% more, which is effectively the same, and the run takes about one ninth as long.
Similar speed, per hour: 18 CPU VMs against 1 GPU VM18 × $3.60 = $64.80$32.77(64.80 − 32.77) ÷ 64.80 = 0.494. The GPU side costs about 49% less.
Per hour: 16 CPU VMs against 1 GPU VM16 × $3.60 = $57.60$32.77(57.60 − 32.77) ÷ 57.60 = 0.431. The GPU side costs about 43% less per hour.
Same forecast, GPU at 1.07 times the speed: 16 CPU VMs against 1 GPU VM. Example: a job of 1 hour on the CPU VMs.1 hour × $57.60 = $57.60(1 ÷ 1.07) hours × $32.77 = $30.63(57.60 − 30.63) ÷ 57.60 = 0.468. 47% lower compute cost, before the AceCAST license.

Two things stand out. With one VM on each side, the price ratio (about 9.1) and the speed ratio (about 9) cancel: you pay about the same and wait about one ninth as long. And by the benchmark's own numbers it took 18 CPU VMs, not 9, to match one GPU VM, so adding CPU VMs did not speed the run up in proportion. That is where the saving in the larger comparisons comes from.

The published speed is 1.07 times the CPU speed, so the GPU run takes 1 ÷ 1.07 = 0.934579 hours, displayed as 0.935 GPU node-hours. Costs use the unrounded value.

Why this is not the 75% in the benchmark

The authors reported 75% lower cost. Using the list rates shown here gives about 47% lower compute cost, before the AceCAST license. The original cost basis has not been reproduced here; a change in cloud prices alone does not explain the difference.

The 47% is a worked example, not a measurement: the 2023 speed ratios for AceCAST 2.1 and 3.0.1 multiplied by 2026 prices. Your own workload will have its own number.

The same article also reports that "AceCAST runs 16.8x faster than WRF for inner domain compute and communication times." That describes one part of a run, not the whole elapsed time you are billed for, so it is not used in the sums above.

What the sums assume, and what they leave out

Assumptions:

  • Azure list prices for one region, Linux, pay-as-you-go, in November 2026. Other regions and other clouds differ.
  • Whole VMs are billed only while the run is going, with no idle or queue time on either side.
  • The speed ratios are used only for these fixed historical examples. They were measured in 2023 on the authors’ cases and do not predict your own workload.
  • "7% faster" means 1.07 times the speed.

Left out:

  • Software license. AceCAST is licensed software. Its license price is not in these sums.
  • Storage for input data and model output.
  • Data transfer into and out of the cloud.
  • Staff time to set up, run and maintain either system.
  • Reserved or spot pricing, which can move either side of the comparison.
  • Your own domain and physics. Grid size, nesting, physics options and output frequency all change how long a run takes on each kind of hardware.

The only number that settles the question is your own case, timed on both kinds of hardware.

What to do next

Edit hourly rates and runs per month in the fixed benchmark example in the cost estimator, or request a trial and time your own case.

Your forecast

Have a forecast question of your own?

Tell us the decision it has to serve. One short form reaches the team, and we review every request.