MetaTrader 5 Build 5572's CUDA backend requires NVIDIA GPUs with compute capability 7.5 or higher (Turing architecture and newer). Anything below that throws CUBLAS_STATUS_ARCH_MISMATCH on session load. This page is the lookup table.
Quick lookup — works
| Architecture | Compute | Year | Example cards |
|---|---|---|---|
| Turing | 7.5 | 2018–2019 | GTX 1660 / 1660 Super / 1660 Ti, RTX 2060 / 2070 / 2080 / 2080 Ti, Quadro RTX 4000–8000, Tesla T4 |
| Ampere | 8.0 / 8.6 | 2020–2021 | RTX 3050 / 3060 / 3070 / 3080 / 3090, A10 / A30 / A40 / A100 |
| Ada Lovelace | 8.9 | 2022–2023 | RTX 4050 / 4060 / 4070 / 4080 / 4090, L4 / L40 |
| Hopper | 9.0 | 2022–2024 | H100, H200 |
| Blackwell | 10.0 / 12.0 | 2024–2025 | B100 / B200 (datacenter), RTX 5060 / 5070 / 5080 / 5090 (consumer) |
Quick lookup — doesn't work
| Architecture | Compute | Year | Example cards |
|---|---|---|---|
| Pascal | 6.0–6.2 | 2016–2017 | GTX 1050 / 1060 / 1070 / 1080 / 1080 Ti, Titan X-Pascal, P4 / P40 / P100, Quadro P-series |
| Maxwell | 5.0–5.2 | 2014–2016 | GTX 750 / 950 / 960 / 970 / 980 / 980 Ti, M40 / M60, Quadro M-series |
| Kepler | 3.0–3.7 | 2012–2014 | GTX 600 / 700-series, K20 / K40 / K80, Quadro K-series |
| Older | < 3.0 | pre-2012 | Fermi, Tesla 1.x — all unsupported |
How to check your card right now
command prompt
> nvidia-smi --query-gpu=name,compute_cap --format=csv
If compute_cap is 7.5 or higher, you're good. Below 7.5, you have two options: force CPU with ONNX_USE_CPU_ONLY, or replace the card.
Buying for ONNX in MT5: minimum, recommended, comfortable
- Minimum compatible (lowest cost): GTX 1660 Super used. 6 GB VRAM. ~$150 in 2026.
- Recommended (new card, good headroom): RTX 4060 or RTX 4070. 8–12 GB VRAM. $300–$600.
- Comfortable (training + inference + room to grow): RTX 4090 or RTX 5080. 16–24 GB VRAM. $1200–$2000.
if you'd rather rent
Hourly access to compatible GPUs.
Skip the hardware purchase entirely — rent a Turing-or-newer card by the hour:
Affiliate links — see full comparison.