RTX 5070 Ti (Blackwell) performs worse than RTX 4060 Ti (Ada) on same simulation

[Ma mike]

## Hardware Comparison | | RTX 4060 Ti (old) | RTX 5070 Ti (new) | |---|---|---| | Architecture | Ada Lovelace (SM 8.9) | Blackwell (SM 12.0) | | VRAM | 8 GB GDDR6 / 128-bit | 16 GB GDDR7 / 256-bit | | Bandwidth | 288 GB/s | 896 GB/s (measured 866) | | SM count | 34 | 70 | | L2 Cache | 32 MB | 48 MB | | TDP | 160W | 300W | ||| Shared hardware: - CPU: Intel i5-12600KF - RAM: 32 GB DDR4 - Motherboard: Gigabyte B760M GAMING AC DDR4 - OS: Windows 10 22H2 (19045), WDDM - CUDA Driver: 13.3 (610.62) - Mumax3: v3.12 (CUDA 12.9, cc=120 PTX, cuFFT 11) - ReBAR: enabled (BAR1 = 16 GB) - ASPM: off - Power plan: High Performance ## Problem On a 512×224×1 grid simulation (fixed timestep 5e-14s, same .mx3 script): - 4060 Ti: ~26 output files per minute - 5070 Ti: ~15 output files per minute The 5070 Ti is 1.7x **slower** despite having 3x the memory bandwidth and 2x the SM count. ## What we checked (all normal) - Memory bandwidth test: 866 GB/s (97% of theoretical) ✓ - ReBAR: enabled, 16 GB BAR1 ✓ - No thermal throttling (53°C, 200W/300W) ✓ - No power throttling ✓ - PCIe Gen 5 x16 capable, link width x16 ✓ - Clean CUDA toolkit removal, only driver remains ✓ - Killed background GPU processes (GCC, NVIDIA Overlay) ✓ - Kernel cache warm after first run ✓

[Антон Луценко]

Hi, there can be a number of issues (seems like the CPU is bottlenecking), but here is some observation:

The simulation area is tiny. The overhead of launching the CUDA kernels might be larger than actual computing time. Try a larger grid, maybe the standard benchmark.

[Ma mike]

Hi, you are right. very thanks! The CPU will indeed be a bottleneck. But the main cause is the source code of mumax3 adopts an old architecture and emits kernels one by one.

A larger grid is indeed faster.

Related issues have also been reported in the literature: Optimization of MuMax3 by Using Claude Code (J. Magnetics, 2026).