Skip to content

Multi-GPU setups: SLI, TCC mode, and utilization

  • Status: unverified
  • Applies to: Metashape Pro 2.x (Python) ; Metashape Standard 2.x (GUI)
  • Edition: Pro / Standard
  • Diátaxis: explanation
  • Confidence: high
  • Last reviewed: 2026-05-27

Confidence: high. GPU configuration is documented in the official manual (§ System Requirements → GPU recommendations), and the SLI / TCC claims are directly from Agisoft support with permalinks.

Metashape uses CUDA (NVIDIA) or OpenCL (AMD / Apple) for GPU- accelerated stages. It addresses each GPU independently — SLI and NVLink bridges do not accelerate processing.

"SLI doesn't help for CUDA computations. SLI is used to improve the visualization only." — Alexey Pasumansky, 2017-05-08, PhotoScan 1.3 (permalink)

If you have two GPUs connected via SLI/NVLink, Metashape will use both independently for compute. The bridge adds no benefit and can be removed without affecting processing speed.

TCC mode on Windows (Quadro / Tesla / Titan only)

On Windows, NVIDIA GPUs use the WDDM (Windows Display Driver Model) by default. WDDM adds overhead for GPU command submission because it routes through the Windows graphics stack.

For GPUs not connected to a display (dedicated compute cards), switching to TCC (Tesla Compute Cluster) mode removes this overhead:

nvidia-smi.exe -g <card_id> -dm 1   # switch to TCC (requires reboot)

Caveat: TCC is only supported on:

  • Quadro cards
  • Tesla / A100 / H100 data-centre cards
  • Titan cards (varies by generation)

GeForce GTX / RTX consumer cards do NOT support TCC. Attempting to switch them will fail silently or error.

After switching, verify with nvidia-smi.exe — the card should show "TCC" instead of "WDDM".

Point cloud fusion phase is CPU-only

GPU utilization drops to near-zero during the second phase of point cloud generation (fusion / filtering). This is expected behaviour, not a driver issue or configuration problem.

The two phases are:

  1. Depth map computation — GPU-accelerated (high utilization).
  2. Point cloud fusion — CPU-only (GPU idle).

If you observe high GPU utilization during depth maps followed by a long CPU-only phase, this is normal. See GPU usage by stage for the full breakdown.

Linux has better multi-GPU utilization

On multi-GPU Windows systems, users frequently report low GPU utilization (20-30%) during depth map generation. The same hardware under Linux typically shows higher utilization because Linux's GPU driver model has lower per-command overhead than WDDM.

See Linux vs Windows performance for benchmarks and root-cause analysis.

See also: GPU usage by stage.

Configuring multi-GPU in Metashape

In the GUI: Tools → Preferences → GPU tab. Check all GPUs you want to use. Optionally uncheck "Use CPU when processing" if you want GPU-only processing (useful for isolating GPU performance).

In Python:

Demo verified: ✓ — Metashape 2.2.3, Apple M1 Pro (attributes confirmed via introspection).

import Metashape

# List available GPUs
print(Metashape.app.gpu_mask)  # bitmask of enabled GPUs
print(Metashape.app.enumGPUDevices())  # list of GPU device strings

# Enable all GPUs (set all bits)
Metashape.app.gpu_mask = (1 << len(Metashape.app.enumGPUDevices())) - 1

# Disable CPU during GPU processing
Metashape.app.cpu_enable = False

Key points

  • SLI/NVLink: irrelevant for processing. Remove if desired.
  • TCC mode: improves utilization on Quadro/Tesla/Titan (not GeForce).
  • Point cloud fusion: CPU-only by design. GPU idle is expected.
  • Linux: better multi-GPU utilization due to lower driver overhead.
  • Multiple GPUs: each is used independently; more GPUs = faster depth maps (near-linear scaling for that stage).

References

  • Metashape Pro User Manual (2.3), § System Requirements → GPU recommendations — lists supported GPU families and CUDA/OpenCL requirements.
  • Metashape Python API Reference (2.3.1): Application.gpu_mask, Application.enumGPUDevices(), Application.cpu_enable.