CUDA GPU support in MetaTrader 5: the compatible cards, the floor, the gotchas.

MetaTrader 5 Build 5572 added native CUDA acceleration for ONNX inference. The list of GPUs that actually work with it is shorter than NVIDIA's full catalog — Turing architecture (compute capability 7.5) is the floor, and anything below that fails on session creation. This article documents the supported cards, the compute capability of each architecture, what to look for when buying, and the renting alternative for people who don't want to commit to hardware.

The architecture floor: Turing (compute 7.5)

NVIDIA labels each GPU generation with both a marketing name (Pascal, Turing, Ampere, Ada Lovelace, Blackwell) and a numeric compute capability. MT5's CUDA backend is compiled with compute 7.5 as the lowest target. Any card below that — Pascal at 6.x, Maxwell at 5.x, Kepler at 3.x — throws CUBLAS_STATUS_ARCH_MISMATCH on session creation.

Generations that meet or exceed the floor: Turing, Ampere, Ada Lovelace, Hopper, Blackwell. In practical terms, every NVIDIA gaming card from 2018 onward and every datacenter card from the T4 onward.

Compatible card list (by tier)

Entry: GTX 16-series & RTX 20-series (Turing, compute 7.5)

• GTX 1660, 1660 Super, 1660 Ti — the cheapest Turing cards. 6 GB VRAM. Used market $150–$200 in 2026. Enough for any retail-scale inference workload, marginal for training large models.
• RTX 2060, 2070, 2080, 2080 Ti — first-generation RTX. 6–11 GB VRAM. Tensor cores present (relevant for FP16 inference).
• Quadro RTX 4000, 5000, 6000, 8000 — workstation cards. Higher VRAM, ECC memory.
• Tesla T4 — the cheapest datacenter Turing. 16 GB VRAM, low-power. Ubiquitous on cloud (RunPod, AWS, GCP, Vast.ai).

Mid: RTX 30-series & A-class (Ampere, compute 8.0–8.6)

• RTX 3060, 3070, 3080, 3090 — 8–24 GB VRAM. Strong 2nd-hand market.
• A40, A100 — datacenter Ampere. A100 (40 or 80 GB) is the workhorse of mid-2020s ML.

Current: RTX 40 / 50-series & Hopper (compute 8.9 / 9.0 / 10.0)

• RTX 4060, 4070, 4080, 4090 — Ada Lovelace. 8–24 GB. RTX 4090 (24 GB) is the new gold standard for desktop ML.
• RTX 50-series — Blackwell consumer. Compute 12.0. Latest at writing.
• H100, H200 — Hopper datacenter. Overkill for retail trading; commonly rented for training.

Cards that don't work

For completeness, these will throw CUBLAS_STATUS_ARCH_MISMATCH:

• Pascal (compute 6.x): GTX 1050, 1060, 1070, 1080, 1080 Ti, Titan X-Pascal, Quadro P-series, Tesla P4/P40/P100.
• Maxwell (compute 5.x): GTX 750, 950, 960, 970, 980, 980 Ti, Quadro M-series, Tesla M40/M60.
• Kepler (compute 3.x): GTX 600/700-series, Quadro K-series, Tesla K20/K40/K80.
• All older NVIDIA architectures (Fermi 2.x, Tesla 1.x, etc.).

Choosing a card for ONNX in MT5

The decision depends almost entirely on what you do with the GPU besides MT5:

• Just running inference (24/7 EA), retail-scale model: a used GTX 1660 Super at $150 is enough. The inference workload barely touches the card.
• Training models locally: at minimum RTX 3060 (12 GB) or RTX 4070 (12 GB). The training step needs the memory more than the inference step.
• Training transformer-scale models locally: RTX 4090 (24 GB) or rent A100/H100 by the hour. 24 GB is the realistic upper bound for desktop without paying datacenter prices.
• Just want it to work, no commitment: skip the hardware purchase, rent in the cloud.

Renting one instead

If you're not sure your trading workload justifies a card purchase — or your dev machine is a laptop — GPU clouds let you rent Turing-or-newer cards by the hour. Importantly, since MT5 requires Windows, make sure the cloud provider supports Windows images on GPU instances (not all do).