Which Mining GPUs Still Work with CUDA in 2026? P102, P104, P106 and CMP HX Compared
Old NVIDIA mining cards did not suddenly lose their CUDA cores when cryptocurrency mining became less profitable. A P102-100 can still run a CUDA kernel. A CMP 30HX can still process data. Even a card without display connectors may be useful as a secondary compute device. The difficult part in 2026 is everything around the chip: driver installation, CUDA Toolkit compatibility, limited VRAM, missing video features and software that has quietly dropped older architectures. This comparison covers the dedicated NVIDIA mining GPUs most often found on the used market: - P106-090 and P106-100; - P104-090, P104-100 and P104-101; - P102-100; - CMP 30HX, 40HX and 50HX; - CMP 70HX, 90HX and 170HX. The short answer is simple: all of these cards can run at least some CUDA workloads with a suitable driver, but they are not equally useful in 2026. Pascal-based P102, P104 and P106 cards should stay on CUDA 12.x. Turing and Ampere CMP cards can use CUDA 13, although NVIDIA development tools may still reject a card because it is identified as a Crypto Mining Processor. Quick Compatibility Table P106-090 / P106-100 Architecture: Pascal Compute capability: 6.1 Typical VRAM: 3–6 GB Recommended toolkit: CUDA 12.9 or an older 12.x release 2026 status: Works, but VRAM and software support are restrictive. P104-090 / P104-100 / P104-101 Architecture: Pascal Compute capability: 6.1 Typical VRAM: 4–8 GB Recommended toolkit: CUDA 12.9 or older 12.x 2026 status: Usable for legacy CUDA and light compute. P102-100 Architecture: Pascal Compute capability: 6.1 Typical VRAM: 5–10 GB Recommended toolkit: CUDA 12.9 or older 12.x 2026 status: The best common P10x option, especially with 10 GB. CMP 30HX Architecture: Turing Compute capability: 7.5 Typical VRAM: 6 GB Recommended toolkit: CUDA 13.x or 12.x 2026 status: Modern enough, but limited by 6 GB and its mining-only design. CMP 40HX Architecture: Turing Compute capability: 7.5 Typical VRAM: 8 GB Recommended toolkit: CUDA 13.x or 12.x 2026 status: More balanced than 30HX when the price is low. CMP 50HX Architecture: Turing Compute capability: 7.5 Typical VRAM: 10 GB Recommended toolkit: CUDA 13.x or 12.x 2026 status: The most interesting Turing CMP option. CMP 70HX Architecture: Ampere Compute capability: 8.6 Typical VRAM: usually 8 GB Recommended toolkit: CUDA 13.x or 12.x 2026 status: Fast, although VRAM can limit current AI workloads. CMP 90HX Architecture: Ampere Compute capability: 8.6 Typical VRAM: 10 GB Recommended toolkit: CUDA 13.x or 12.x 2026 status: Strong compute potential; power use and price decide its value. CMP 170HX Architecture: Ampere GA100 Compute capability: 8.0 Typical VRAM: 8 GB HBM2e Recommended toolkit: CUDA 13.x or 12.x 2026 status: A specialist card with unusual platform and cooling requirements. Memory size and device identifiers can vary between manufacturers and board revisions. Always verify the exact card with its PCI hardware ID and `nvidia-smi` instead of trusting a seller's title. The Important Difference: CUDA-Capable Is Not the Same as Fully Supported Three separate questions are often mixed together: 1. Can the GPU execute CUDA code? 2. Can the operating system load a working driver for it? 3. Does the application support that GPU architecture and memory size? A card may pass the first test and fail the other two. For example, the P102-100 is built from NVIDIA's Pascal generation and supports compute capability 6.1. It can execute CUDA code compiled for `sm_61`. However, a normal modern GeForce installer may not recognize the mining-specific PCI device. A patched driver may therefore be required before Windows exposes the card to CUDA. Even after `nvidia-smi` works, a current application can refuse to start because its binaries contain kernels only for newer architectures such as `sm_75`, `sm_86` or `sm_89`. The safest rule is this: A visible GPU in Device Manager is not proof of CUDA compatibility. A working `nvidia-smi` is not proof that a specific application supports the card. Test all three layers. Why CUDA 13 Changes the Answer for Pascal Mining GPUs CUDA Toolkit 13 removed offline compilation and CUDA library support for Maxwell, Pascal and Volta GPUs. NVIDIA describes these architectures as feature-complete and directs developers who still target them to the CUDA 12.x series. That directly affects: - P106-090 and P106-100; - P104-090, P104-100 and P104-101; - P102-100; - other Pascal-based mining boards. These GPUs normally use compute capability 6.1. CUDA 12.9 still recognizes the Pascal targets: `sm_60`, `sm_61` and `sm_62` CUDA 13 cannot be treated as a drop-in development toolkit for them. A new driver may remain capable of running older binaries through backward compatibility, but the CUDA 13 compiler and libraries no longer provide normal Pascal target support. For a P102, P104 or P106 system in 2026, the practical choice is: - install a driver that correctly recognizes the mining GPU; - use CUDA Toolkit 12.9 Update 1, or the 12.x version required by the application; - compile custom kernels for compute capability 6.1; - avoid replacing the patched display driver when installing the toolkit. An explicit `nvcc` target looks like this: `nvcc kernel.cu -gencode arch=compute_61,code=sm_61 -o kernel` For CMake: `cmake -S . -B build -DCMAKE_CUDA_ARCHITECTURES=61` If a prebuilt program reports `no kernel image is available for execution on the device`, the package was probably built without Pascal code. Changing the driver alone will not add a missing `sm_61` binary to that application. P106-090 and P106-100: CUDA Works, but Memory Is the Wall P106 cards are based on the same Pascal generation associated with the GTX 1060 family, but mining editions commonly lack display outputs and use mining-specific device IDs. Their main weakness in 2026 is not basic CUDA execution. It is capacity. A 3 GB P106-090 is difficult to recommend for modern GPU computing. After driver overhead, the usable memory is too small for many current AI models, large image workflows and serious datasets. It remains suitable for: - learning CUDA C++; - small custom kernels; - password-auditing workloads where legally authorized; - lightweight OpenCL or CUDA utilities; - distributed jobs with very small per-GPU memory requirements. A 6 GB P106-100 is less restrictive, but it still lacks Tensor Cores and remains tied to Pascal-compatible software. It can be fun as a cheap laboratory card. It is rarely a good foundation for a new workstation. Verdict: buy only if it is extremely inexpensive and the objective is experimentation rather than efficient production work. P104-100: A Better Middle Ground, with the Same Pascal Deadline P104 mining cards generally offer more memory bandwidth and compute resources than P106 models. An 8 GB P104-100 can handle workloads that simply do not fit on a 3 GB or 6 GB card. The limitations remain familiar: - compute capability 6.1; - CUDA 12.x for development; - no Tensor Cores; - possible patched-driver requirement; - board-dependent display, fan and power behavior; - uncertain history after years of mining operation. For custom CUDA experiments, Blender versions that still support Pascal, older machine-learning environments and general-purpose compute, an inexpensive 8 GB P104 can still be useful. Before buying, compare it with a normal used GTX 1070 or GTX 1080. A consumer card may cost slightly more but usually has display outputs, broader driver support and better resale value. Verdict: usable as a low-cost 8 GB compute card, but only when the price compensates for the driver and hardware compromises. P102-100: The Most Useful Pascal Mining Card Among the common P10x mining GPUs, the 10 GB P102-100 is the strongest candidate for CUDA work. Its GP102-class silicon and wider memory subsystem give it considerably more room than P106 hardware. The card can be useful for: - CUDA programming and architecture experiments; - rendering software that retains Pascal support; - selected AI inference workloads with compatible builds; - compute tasks that benefit from FP32 throughput and memory bandwidth; - secondary-GPU jobs where no display output is required. But the P102-100 should not be confused with a modern AI accelerator. It has no Tensor Cores, lacks current CUDA 13 library support and may draw a substantial amount of power. Its 10 GB of VRAM is helpful, but current AI software may require newer compute capabilities even when the model itself fits into memory. On Windows, the card may appear as an unknown `3D Video Controller` until a suitable driver is installed. Our separate installation guide covers the P102-100 hardware ID, matched NVIDIA-patcher release, certificate installation, `pnputil`, CUDA 12.9 and `sm_61` compilation: Link available to registered users Verdict: the best P10x option for a hobbyist who accepts patched drivers. For a new production machine, a normal Turing or Ampere card is safer. CMP 30HX, 40HX and 50HX: Turing Survives CUDA 13 CMP 30HX, 40HX and 50HX belong to the Turing generation. Their compute capability is 7.5, so they are on the supported side of the CUDA 13 cutoff. This gives them an important advantage over P102, P104 and P106 cards: - CUDA 13 can target Turing; - current CUDA libraries are not excluded merely because the architecture is too old; - Turing supports Tensor Cores, although exact card configuration and application support still matter; - software compatibility is generally better than on Pascal. That does not make every CMP HX card a bargain. The 30HX has only 6 GB of memory. The 40HX is more flexible at 8 GB, while the 10 GB 50HX is the most practical of the three for memory-heavy workloads. There is also a less obvious limitation. NVIDIA documents that some development and profiling tools do not support devices identified as Crypto Mining Processors and may display an `ERR_NVCMPGPU` error. CUDA runtime workloads can work while Nsight or another development tool refuses to profile the same card. Verdict: CMP 50HX is the most attractive Turing CMP card; 30HX makes sense only at a very low price. CMP 70HX and 90HX: Fast Ampere Compute with Mining-Card Compromises CMP 70HX and 90HX move to the Ampere generation. These cards support CUDA 13 and are much closer to current software expectations than Pascal mining GPUs. The 90HX is the more desirable model for raw performance, but three details decide whether it is a smart purchase: 1. VRAM: 8–10 GB is useful, but current AI workloads can exceed it quickly. 2. Power: an inexpensive card can become expensive if it consumes significantly more electricity than a newer alternative. 3. Driver behavior: mining-specific IDs and vendor-specific boards can still require extra work. For stable diffusion, local inference or rendering, compare the complete system cost against a used RTX 3060 12 GB. The RTX 3060 may be slower in some raw comparisons, but its 12 GB memory, display outputs, conventional drivers and normal application support often make it the more practical device. Verdict: technically capable and CUDA 13 compatible, but compare price, power and VRAM against ordinary RTX cards before buying. CMP 170HX: Interesting Hardware, Awkward Everyday Card CMP 170HX is fundamentally different from the smaller HX products. It is based on GA100-class Ampere silicon and uses HBM2e memory. That makes it technically fascinating, but not automatically convenient. Potential obstacles include: - unusual board design; - no conventional display role; - demanding cooling and airflow requirements; - server-oriented power considerations; - only 8 GB of accessible memory on common CMP 170HX boards; - application or tool restrictions associated with CMP identification. The card can deliver strong compute performance in a properly prepared system, but it is not a beginner-friendly way to run CUDA. Treat it as specialist hardware rather than a cheap substitute for a normal RTX card. Verdict: worthwhile for an experienced builder with a specific workload and suitable chassis; poor choice for a first CUDA system. CUDA 12.9 vs CUDA 13 for Mining GPUs P102, P104 and P106 CUDA 12.9: supported through the Pascal `sm_61` target. CUDA 13.x: not supported as a normal compilation and library target. CMP 30HX, 40HX and 50HX CUDA 12.9: supported. CUDA 13.x: supported through the Turing `sm_75` target. CMP 70HX and 90HX CUDA 12.9: supported. CUDA 13.x: supported through the Ampere `sm_86` target. CMP 170HX CUDA 12.9: supported. CUDA 13.x: supported through the Ampere `sm_80` target. Practical choice For legacy Pascal hardware, use CUDA 12.9. For a new Turing or Ampere project, CUDA 13 is normally preferable when the required application and CMP driver work correctly. Do not select a toolkit only because `nvidia-smi` prints a high CUDA version. The `CUDA Version` field in `nvidia-smi` shows the maximum CUDA driver API level supported by the installed driver. It does not prove that the corresponding CUDA Toolkit is installed, and it does not change the physical compute capability of the GPU. Check the toolkit separately: `nvcc --version` Check the driver and detected GPU: `nvidia-smi` On Linux, list the PCI device: `lspci -nn | grep -i nvidia` On Windows PowerShell: `Get-PnpDevice -Class Display | Format-Table Status, FriendlyName, InstanceId` Can These Mining GPUs Run AI Models? Some can, but "supports CUDA" is a very low bar for modern AI. For AI inference, consider: - available VRAM after driver allocation; - compute capability required by the framework build; - whether the package includes kernels for the card; - Tensor Core availability; - FP16, BF16 and INT8 support; - PCIe bandwidth; - cooling and sustained power draw. P102, P104 and P106 cards can run compatible CUDA code, but they have no Tensor Cores. Many current prebuilt packages are optimized for newer architectures, and some no longer ship Pascal kernels. Building an older framework from source may be possible, but that turns a cheap GPU into a maintenance project. Turing CMP cards are a better minimum for experimental AI because compute capability 7.5 remains supported by CUDA 13. Ampere CMP cards are stronger again. Nevertheless, memory capacity often matters more than theoretical throughput. A conventional 12 GB or 16 GB GPU can be more useful than a faster 8 GB mining card. What to Check Before Buying a Used Mining GPU Ask the seller for: - a clear photo of the exact board and memory chips; - PCI hardware ID; - `nvidia-smi` output; - reported memory capacity; - BIOS version; - temperature and power readings under load; - information about display outputs, if any; - confirmation that fans respond to control; - a short CUDA or rendering test. Watch for cards advertised under a more familiar GeForce name. Patched drivers may deliberately map a mining GPU to a consumer profile, so Device Manager alone is not reliable proof of the physical model. Compare the PCI ID, memory size, board layout and `nvidia-smi` result. Mining history itself is not an automatic death sentence. A card operated at a stable temperature and reduced voltage may be healthier than a gaming GPU repeatedly exposed to thermal cycling. The real risks are worn fans, corroded boards, damaged memory, modified firmware and repairs that the seller does not disclose. Which Mining GPU Should You Choose in 2026? For the lowest-cost CUDA learning system: - choose a P106-100 only if it is nearly free; - prefer a normal GTX or RTX card when prices are close. For legacy Pascal experiments: - choose a 10 GB P102-100; - use CUDA 12.9; - expect patched-driver work; - compile for `sm_61`. For broader modern CUDA compatibility: - start with CMP 50HX or newer; - prefer Ampere CMP 70HX or 90HX when the price and cooling make sense; - verify that required developer tools do not reject CMP devices. For a dependable daily workstation: - buy a conventional RTX GPU instead; - prioritize VRAM, standard drivers and application support over theoretical bargain performance. Final Answer NVIDIA P102-100, P104-100 and P106-100 mining GPUs still run CUDA workloads in 2026, but they should be paired with CUDA Toolkit 12.9 or another compatible CUDA 12.x release. They are Pascal devices with compute capability 6.1, and CUDA 13 no longer provides normal compiler and library support for that generation. CMP 30HX, 40HX and 50HX use Turing compute capability 7.5 and remain compatible with CUDA 13. CMP 70HX, 90HX and 170HX are Ampere-based and also remain viable for current CUDA software. However, mining-specific drivers, absent display outputs, tool restrictions, VRAM capacity and power consumption can matter more than raw CUDA support. The 10 GB P102-100 is the most useful common Pascal mining card for a low-cost experiment. CMP 50HX is a more future-proof Turing option. CMP 90HX offers stronger modern compute, while CMP 170HX is specialist hardware. For a machine that must work every day without driver experiments, a standard RTX card remains the safer purchase. Official References NVIDIA CUDA Toolkit 13.0 release notes — removal of Maxwell, Pascal and Volta support: Link available to registered users NVIDIA CUDA Toolkit 12.9 release notes and driver compatibility: Link available to registered users NVIDIA legacy compute capability table: Link available to registered users NVIDIA CMP development-tool limitation: Link available to registered users NVIDIA CMP HX specifications: Link available to registered users NVIDIA-patcher project and supported mining cards: Link available to registered users
Computer Hardware