Skip to content

Performance tuning: CPU, RAM, GPU, and OS

  • Status: unverified
  • Applies to: Metashape Standard 2.x ; Metashape Pro 2.x
  • Edition: Pro / Standard
  • Diátaxis: explanation
  • Confidence: medium
  • Last reviewed: 2026-06-05

Confidence: medium. The CPU / GPU / OS findings are consistent across multiple forum threads (2013–2025) but the specific BIOS workarounds and CPU-stability claims are hardware- and firmware-dependent. Linux-vs-Windows speedups are community-benchmarked, not officially confirmed.

The official manual's System Requirements page lists hardware recommendations but does not cover what actually matters for sustained Metashape throughput, why the differences arise, or the BIOS / OS-level interactions that most users underestimate. This article consolidates the operational findings from the forum.

The starting point: the existing GPU usage by stage article enumerates which processing steps use the GPU, the CPU, or both. Read that first if you have not. This article addresses everything else: CPU selection, RAM sizing, multi-GPU setups, the OS gap, and known hardware instability.

CPU: frequency matters more than core count

For most Metashape stages, 6–8 high-frequency cores beat 40+ low-frequency cores. The classic finding from a 2013 PhotoScan benchmark thread:

"For dual-Xeon systems I can suggest to use 6-8 core Xeons with highest possible frequency. Using 40 cores with 2.2 GHz could be even slower than desktop six-core i7." — Alexey Pasumansky, 2015-02-09, PhotoScan 1.1 (permalink, archived 2026-09-23)

Reasons:

  • Thread synchronisation overhead scales super-linearly past some saturation threshold. Beyond ~32 cores on most workloads, threading overhead consumes more cycles than the additional cores contribute. The exact threshold varies by stage and OS.
  • Turbo Boost only sustains briefly. Sustained AVX2 / AVX-512 loads thermally throttle to base clock within seconds. Marketing labels (e.g., "5.8 GHz Turbo") are misleading; benchmark against the base clock.
  • Per-stage scaling differs. On Agisoft's 2015 dual-Xeon tests:
Stage Dual-Xeon vs Single speedup
Align Photos 20–40%
Build Point Cloud (CPU phase) ~30%
Build Mesh 10–15%

Modern CPUs scale further than 2015's, but the diminishing-returns pattern remains.

RAM: peak is during mesh / depth-map storage

See the existing RAM and quality settings article. Briefly: peak RAM is usually during Build Mesh (Ultra quality) or during depth-map fusion if the project has many high- resolution depth maps in memory simultaneously.

A practical guideline (Agisoft's): 4 GB → 30–50 photos at 10 MPx; 16 GB → 300–500 photos. Modern builders typically target 64–128 GB for medium-large projects.

GPU: depth maps + texture only

Per GPU usage by stage, the GPU is heavily used during:

  • Build Depth Maps (the most GPU-intensive stage)
  • Build Point Cloud (depth-maps phase only — phase 2 is CPU)
  • Build Texture (since 2.x)

Other stages (alignment, mesh, orthomosaic, DEM) are CPU-only. This means GPU upgrades only help if depth maps or texture generation dominates your wall-clock time. For projects dominated by alignment or orthomosaic export, GPU spending is wasted.

Multi-GPU: SLI is irrelevant; TCC mode helps Quadro/Tesla

See Multi-GPU setups. Three operational facts:

  • SLI / NVLink does not accelerate compute. Bridges are for visualisation only; remove them without consequence.
  • TCC mode improves utilisation on Quadro / Tesla / Titan by removing WDDM overhead. Not supported on GeForce GTX/RTX consumer cards.
  • Linux has lower multi-GPU overhead than Windows (no WDDM in the path).

OS: Linux 10–40% faster than Windows on the same hardware

See Linux vs Windows performance. Three ways the gap manifests:

  1. Stock Windows vs stock Linux: 30–40% faster on Linux (community benchmarks, December 2024, multiple users).
  2. Debloated Windows 10 LTSC vs Linux: narrows to ~5–10%.
  3. Multi-GPU Windows: further compounded by WDDM driver overhead per GPU.

Root causes:

  • Windows background services (telemetry, Defender, Superfetch) consume cycles.
  • Linux's GPU driver model has lower per-command overhead than WDDM.
  • File-system differences (NTFS vs ext4/XFS) may contribute for I/O-heavy stages but are not conclusively isolated.

For sustained Metashape work, Linux pays off — either as a dedicated processing machine or via dual-boot.

Hardware instability: Intel i9-13900K / i9-14900K

A documented stability issue affects Intel Core i9-13900K and i9-14900K CPUs under sustained AVX2 load (which Metashape exercises during alignment and point-cloud computation). The crashes are not Metashape bugs — they affect other CPU-heavy applications too. Multiple forum users have confirmed BIOS-level workarounds:

  • Lower the Turbo Power Limits in BIOS (Gigabyte and others).
  • Cap the Performance Active-Core Tuning Ratio in Intel Extreme Tuning Utility (e.g., to 53× on a 14900K).
  • Apply the Intel microcode update released in 2024 (newer BIOS revisions include this automatically).

Diagnostic tools the Agisoft team recommends:

  • Memtest86 — multithreaded RAM check (run from bootable USB)
  • Intel Processor Diagnostic Tool — official CPU stress test
  • Prime95 with AVX2 — exercises the same instructions Metashape uses
  • Intel Extreme Tuning Utility — AVX2 stress test + tuning

If your workstation crashes on alignment but other apps run cleanly, run the AVX2 stress test first; the 13/14900K instability often passes Prime95 SSE but fails AVX2.

"Quite a lot of cases with i9-13900K and i9-14900K (and similar) CPU system crashes were related to the mentioned hardware issues, as confirmed by users." — Alexey Pasumansky, 2023-10-31, Metashape 2.1 (permalink, archived 2026-09-23)

For the full step-by-step BIOS-workaround procedure with vendor- specific menu paths (ASUS, Gigabyte, MSI, ASRock), the diagnostic ladder, and the crash-report locations Agisoft support needs, see Intel i9-13900K / 14900K instability: BIOS workarounds.

Practical recommendations

For a new workstation primarily running Metashape:

Component Choose Avoid
CPU 8–16 cores at 4.0+ GHz base, AMD Ryzen 7/9 or Intel i7/i9 (post-microcode-fix) 32+ slow cores; pre-2024-microcode i9-13900K/14900K
RAM 64–128 GB ECC (more for >1000-image projects) 32 GB for serious work
GPU Quadro / RTX 4090 with 24 GB VRAM (or multi-GPU on Linux for depth maps) SLI bridges; GTX cards under 8 GB VRAM
OS Linux for primary work; Windows for GUI-heavy editing Stock Windows 11 with full bloat
Storage NVMe for project working directory; large HDD/SAN for archive Spinning HDD as primary

For an existing workstation:

  • Run Memtest86 before blaming Metashape for crashes.
  • Disable Windows Defender real-time scanning during long-running batches.
  • Consider a Linux dual-boot if alignment / depth maps dominate your throughput.
  • Upgrade BIOS if on i9-13/14900K to get the 2024 microcode fix.

References

See also