Quantum · Official
NVIDIA's Daviator Speeds Up Materials Simulations on GPUs With Batched Processing
NVIDIA Technical blog ·

At a glance
NVIDIA described a new library called Daviator that speeds up plane-wave DFT, a common method for simulating materials, on GPUs. It processes many k-points together instead of one at a time, which cuts wasted GPU time on small systems. It also adds a mixed precision layer that uses faster lower-precision math where possible.
Key points
- Daviator is an NVIDIA-built GPU library that handles many k-points (sample points used in materials simulations) at once.
- Before this, small systems with many k-points could run slower on expensive GPUs than on CPUs because of launch overhead.
- A new mixed precision layer adds more speed, especially on GPUs that are much faster at single-precision math.
- On a projected reference dataset workload, the results show up to 22x faster time-to-solution and up to 75% lower cost than CPU runs, depending on hardware and setup.
Why it matters
Faster, cheaper simulation data helps train the AI models used to study materials and chemistry.
Original source
Tackling GPU Bottlenecks in Plane-Wave DFT with K-Point Batching and Mixed Precision
NVIDIA Technical blog · Nov 19, 00:00 KST
Read on NVIDIA Technical blog ↗AI summarized the original article. Check the original for accuracy. Copyright belongs to NVIDIA Technical blog.
More from this field
- Yale Team Confirms Decade-Old Prediction on Fast-Moving Excitons in Thin Materials
Yale researchers have, for the first time, watched a type of energy-carrying particle called an exciton move in an unexpected way. The result backs a prediction made about ten years ago by Professor Diana Qiu. The team studied an atom-thin material.
Phys.org · Nov 21, 01:40 KST - Aalborg University Orders ConScience Quantum Computer for Engineering Training
Aalborg University in Denmark has chosen Swedish company ConScience to supply an on-site superconducting quantum computer for hands-on engineering training. It is set to be installed in 2027, starting with two qubits and growing to seven. Students will work directly with the hardware instead of only using cloud simulations or remote access.
Quantum Computing Report · Nov 21, 01:29 KST - France Backs €14M Qubit Control 2030 Project to Simplify Quantum Computer Wiring
France's National Quantum Strategy is putting €14 million (about $16 million) into a 48-month project called Qubit Control 2030. Led by the startup VIQTHOR, five French companies will build a domestic system to control quantum computers and cut down the bulky copper wiring that runs into the cold chamber. The project runs from October 2026 to September 2030.
Quantum Computing Report · Nov 21, 01:15 KST - Harvard Team Gets Diamond Quantum Spins to Move in Coordinated Spirals
Researchers led by Piotr Put at Harvard University made thousands of nitrogen-vacancy spins inside diamond act together at room temperature. The spins were randomly placed, yet they moved in a coordinated way. The result was published in Physical Review X.
Phys.org · Nov 20, 22:00 KST