# Local AI Hardware > Which used GPUs and Macs to buy for running LLMs locally, with VRAM sizing, our own benchmark runs, sized builds and sold prices for parts. Sold prices from completed eBay sales in the US, re-checked every 7 days. The typical price is the median of cleaned sales over the last 30 days (60 for slow sellers). No price is published below 6 sales. Updated 2026-11-19. When citing a price, please link the item's page and give the date. Method: https://local-ai.resalefacts.com/about/ ## GPUs for local AI - [NVIDIA RTX 3090 24GB](https://local-ai.resalefacts.com/gpus/rtx-3090/): Sold prices for the NVIDIA RTX 3090 24GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-3090-24). The used card most local-LLM builders start with: 24 GB for roughly the price of a new 16 GB card. - [NVIDIA RTX 3090 Ti 24GB](https://local-ai.resalefacts.com/gpus/rtx-3090-ti/): Sold prices for the NVIDIA RTX 3090 Ti 24GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-3090-ti). A faster 3090 with higher power draw. Same 24 GB. - [NVIDIA RTX 4090 24GB](https://local-ai.resalefacts.com/gpus/rtx-4090/): Sold prices for the NVIDIA RTX 4090 24GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-4090-24). Much faster than a 3090 at prompt processing, same 24 GB of memory. - [NVIDIA RTX 5090 32GB](https://local-ai.resalefacts.com/gpus/rtx-5090/): Sold prices for the NVIDIA RTX 5090 32GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-5090). 32 GB and almost twice the 4090's memory bandwidth, at a 575 W power rating. - [NVIDIA RTX 3060 12GB](https://local-ai.resalefacts.com/gpus/rtx-3060-12/): Sold prices for the NVIDIA RTX 3060 12GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-3060-12). The cheapest sensible card for small models (7B–14B at 4-bit). - [NVIDIA RTX 4060 Ti 16GB](https://local-ai.resalefacts.com/gpus/rtx-4060-ti-16/): Sold prices for the NVIDIA RTX 4060 Ti 16GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-4060-ti-16). 16 GB at low power, held back by a narrow memory bus. - [NVIDIA RTX 5060 Ti 16GB](https://local-ai.resalefacts.com/gpus/rtx-5060-ti-16/): Sold prices for the NVIDIA RTX 5060 Ti 16GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-5060-ti-16). 16 GB with much more bandwidth than the 4060 Ti 16GB; a popular pair for 32 GB builds. - [NVIDIA RTX A6000 48GB](https://local-ai.resalefacts.com/gpus/rtx-a6000/): Sold prices for the NVIDIA RTX A6000 48GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-a6000). 48 GB in one two-slot, blower-cooled card. Easier to stack than consumer cards. - [NVIDIA RTX A5000 24GB](https://local-ai.resalefacts.com/gpus/rtx-a5000/): Sold prices for the NVIDIA RTX A5000 24GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-a5000). 24 GB at 230 W in a two-slot blower card. - [NVIDIA RTX Pro 6000 Blackwell 96GB](https://local-ai.resalefacts.com/gpus/rtx-pro-6000/): Sold prices for the NVIDIA RTX Pro 6000 Blackwell 96GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-pro-6000). 96 GB on one card. Expensive, but it replaces four 24 GB cards and their power and cooling. - [NVIDIA Tesla P40 24GB](https://local-ai.resalefacts.com/gpus/tesla-p40/): Sold prices for the NVIDIA Tesla P40 24GB are tracked every day on GPU Sold (https://gpusold.com/gpu/tesla-p40). Cheap 24 GB, but old, slow, and passively cooled. - [NVIDIA Tesla V100 32GB (PCIe)](https://local-ai.resalefacts.com/gpus/tesla-v100-32/): Sold prices for the NVIDIA Tesla V100 32GB (PCIe) are tracked every day on GPU Sold (https://gpusold.com/gpu/tesla-v100-32). 32 GB of HBM2 with high bandwidth, passively cooled. - [AMD Instinct MI50 32GB](https://local-ai.resalefacts.com/gpus/instinct-mi50-32/): Sold prices for the AMD Instinct MI50 32GB are tracked every day on GPU Sold (https://gpusold.com/gpu/instinct-mi50-32). 32 GB of HBM2 for very little money, with real software caveats. - [AMD Radeon RX 7900 XTX 24GB](https://local-ai.resalefacts.com/gpus/rx-7900-xtx/): Sold prices for the AMD Radeon RX 7900 XTX 24GB are tracked every day on GPU Sold (https://gpusold.com/gpu/rx-7900-xtx). 24 GB with ROCm support; good bandwidth for the price. - [NVIDIA RTX 2080 Ti 22GB (modded)](https://local-ai.resalefacts.com/gpus/rtx-2080-ti-22/): Sold prices for the NVIDIA RTX 2080 Ti 22GB (modded) are tracked every day on GPU Sold (https://gpusold.com/gpu/rtx-2080-ti-22). A 2080 Ti with its memory upgraded to 22 GB, sold mainly by modders. No factory warranty applies. ## Multi-GPU build parts - [NVLink bridge for RTX 3090](https://local-ai.resalefacts.com/build-parts/nvlink-bridge-3090/): There are not yet enough recent completed sales of a used NVLink bridge for RTX 3090 to publish a price (we need at least 6 in the last 60 days). The 3090 and 3090 Ti were the last GeForce cards with NVLink. Spacing matters: buy the bridge that matches your slot gap. - [Supermicro H12SSL-i motherboard](https://local-ai.resalefacts.com/build-parts/supermicro-h12ssl-i/): There are not yet enough recent completed sales of a used Supermicro H12SSL-i motherboard to publish a price (we need at least 6 in the last 60 days). A single-socket EPYC board with five PCIe 4.0 x16 slots and two x8: the usual base for multi-GPU servers. - [ASUS Pro WS WRX80E-SAGE SE WiFi](https://local-ai.resalefacts.com/build-parts/asus-wrx80e-sage/): There are not yet enough recent completed sales of a used ASUS Pro WS WRX80E-SAGE SE WiFi to publish a price (we need at least 6 in the last 60 days). Workstation board for Threadripper Pro, seven PCIe 4.0 x16 slots. - [AMD EPYC 7532 (32 cores)](https://local-ai.resalefacts.com/build-parts/amd-epyc-7532/): There are not yet enough recent completed sales of a used AMD EPYC 7532 (32 cores) to publish a price (we need at least 6 in the last 60 days). A 32-core Zen 2 EPYC with 128 PCIe 4.0 lanes, popular for GPU servers. - [AMD EPYC 7302 (16 cores)](https://local-ai.resalefacts.com/build-parts/amd-epyc-7302/): There are not yet enough recent completed sales of a used AMD EPYC 7302 (16 cores) to publish a price (we need at least 6 in the last 60 days). A cheaper 16-core EPYC with the same 128 PCIe lanes. - [64GB DDR4 ECC RDIMM](https://local-ai.resalefacts.com/build-parts/ddr4-64gb-rdimm/): There are not yet enough recent completed sales of a used 64GB DDR4 ECC RDIMM to publish a price (we need at least 6 in the last 60 days). Single 64 GB registered ECC modules, the building block for 256–512 GB CPU-inference rigs. - [1600 W power supply (EVGA SuperNOVA 1600)](https://local-ai.resalefacts.com/build-parts/evga-1600-psu/): There are not yet enough recent completed sales of a used 1600 W power supply (EVGA SuperNOVA 1600) to publish a price (we need at least 6 in the last 60 days). Used EVGA SuperNOVA 1600 W units (P2, T2, G2), common in multi-GPU rigs. - [Dell Precision 7920 tower](https://local-ai.resalefacts.com/build-parts/dell-precision-7920/): There are not yet enough recent completed sales of a used Dell Precision 7920 tower to publish a price (we need at least 6 in the last 60 days). A used dual-Xeon tower that takes big GPUs and a lot of memory. - [HP Z8 G4 workstation](https://local-ai.resalefacts.com/build-parts/hp-z8-g4/): There are not yet enough recent completed sales of a used HP Z8 G4 workstation to publish a price (we need at least 6 in the last 60 days). HP's big dual-socket tower, popular for used GPU builds. ## Macs for local AI - [Mac Studio M2 Ultra 192GB](https://local-ai.resalefacts.com/macs/mac-studio-m2-ultra-192/): There are not yet enough recent completed sales of a used Mac Studio M2 Ultra 192GB to publish a price (we need at least 6 in the last 60 days). 192 GB of unified memory that the GPU can use: large models at low power, slower prompt processing than NVIDIA. - [Mac Studio M1 Ultra 128GB](https://local-ai.resalefacts.com/macs/mac-studio-m1-ultra-128/): There are not yet enough recent completed sales of a used Mac Studio M1 Ultra 128GB to publish a price (we need at least 6 in the last 60 days). The older Ultra, now the cheapest route to 128 GB of GPU-usable memory. - [Mac Studio M3 Ultra](https://local-ai.resalefacts.com/macs/mac-studio-m3-ultra/): There are not yet enough recent completed sales of a used Mac Studio M3 Ultra to publish a price (we need at least 6 in the last 60 days). The 2025 Mac Studio, configurable to 512 GB. Used prices depend heavily on the memory configuration. ## Guides - [How we benchmark GPUs for local LLMs](https://local-ai.resalefacts.com/guides/how-we-benchmark/): The exact method behind every speed on this site: llama.cpp's llama-bench, pinned versions, fixed model files and quantizations, prompt and generation tests, and what we record with each run. (updated 2026-11-18) - [How much VRAM a model needs: weights, context and overhead](https://local-ai.resalefacts.com/guides/how-much-vram/): Work out the memory an LLM needs from its size, quantization and context length, with worked examples for 8B, 32B and 70B models and the formula behind the VRAM calculator. (updated 2026-11-18) - [Local AI builds at 24, 48 and 96 GB: what to buy used](https://local-ai.resalefacts.com/guides/builds-24-48-96/): Three sized builds for running LLMs at home, from a single used RTX 3090 to a four-GPU EPYC server, with the parts, the power, the trade-offs and the Mac alternative at each size. (updated 2026-11-18) ## Tool - [VRAM calculator for local LLMs](https://local-ai.resalefacts.com/tool/): Estimate the GPU memory a model needs from its size, quantization and context length, and see which cards and multi-card setups it fits on. ## Machine-readable data - https://local-ai.resalefacts.com/v1/items (all items with current prices, JSON) - https://local-ai.resalefacts.com/v1/items/{id} (one item with price history and recent sales, JSON) - https://local-ai.resalefacts.com/sitemap.xml Links to eBay on this site are affiliate links.