LLM / AI Workstation — Used GPU, RAM, NVMe Build
Local LLMs (Llama, Mistral, Qwen) are realistic for many in 2025/2026. A used RTX 3090 with 24 GB VRAM for 600-800€ runs most 70B models quantized. Here are the components for a serious local AI workstation or inference server — from GPU to the right motherboard.
What you need
- →GPU with lots of VRAM: RTX 3090 (24 GB) or 4090 (24 GB) for single-card; 70B+ needs 2x
- →RAM: 64 GB DDR4 minimum, 128 GB more comfortable
- →NVMe SSD: at least 2 TB for models and datasets
- →Motherboard: enough PCIe lanes for multi-GPU (X570 or TRX40)
- →CPU: not critical — modern 12-core is fine, AVX-512 nice to have
- →PSU: 850W to 1200W depending on GPU config
Top current deals
INTEL Xeon E5-1620 V3 / 4x 3,5 - 3,6 GHz LGA 2011-v3 10MB Cache / Quad Core CPU
INTEL Xeon E5-2620 / 6x 2,00 - 2,50 GHz / Sockel 2011 SERVER Six Core CPU
INTEL Xeon E5-2620 V3 / 6x 2,4 - 3,2 GHz / Sockel 2011-v3 SERVER Six Core CPU
808409-001 HPE Nvidia Tesla M6 Mezzanine GPU Adapter
Gaming Grafikkarte | MSI NVIDIA RTX 3050 Gaming X 8gb | PCIE 4.0 | Cores 2560
Dell MZ-WLL1T6C Enterprise SSD 1.6TB NVME PCIe U.2 2.5'' 4WDXY
Intel 1.6TB NVMe PCIe SSD 2.5" U.2 DC P4600 Series SSDPE2KE016T701 Enterprise
Western Digital Enterprise NVMe SSD: DC SN620, 1.6TB, PCIe3x4, U.2
Intel Xeon E5-2680 V3 CPU/12x 2,5GHz-3,3GHz/12 Core Prozessor/LGA 2011-3/SR1XP
Intel Xeon E5-2680 V3 CPU/12x 2,5GHz-3,3GHz/12 Core Prozessor/LGA 2011-3/SR1XP
Samsung 4GB DDR4 SO DIMM 3200 MHz CL22 (M471A5244) RAM Speicher
Intel Xeon E5-1620 v2 CPU Prozessor Sockel 2011
Intel Xeon E5-2680 v3 CPU 12x 2,5GHz-3,3GHz Prozessor Sockel 2011-3 SR1XP
GIGABYTE H610M S2H V3 DDR4 LGA 1700 mit HDMI und Display Port
INTEL Xeon E5-2660 V3 / 10x 2,6 - 3,3 GHz / LGA 2011-v3 25MB Cache / 10 Core CPU
Samsung 8GB DDR4 3200MHz RAM M471A1G44BB0-CWE
Intel DC P4510 1TB U.2 NVMe SSD PCIe 3.1 x4 2.5" Enterprise SSDPE2KX010T8
MSI GeForce RTX 2060 6GB GDDR6 Grafikkarte NVIDIA Gaming GPU HDMI DisplayPort
Intel Xeon E5-2680 8-Core SR0KH 2,70GHz Server CPU
960GB 2,5" Micron 7300 Pro Datacenter Enterprise 24/7 U.2 NVMe 220K IOPS +NEW+
GIGABYTE GeForce RTX 3070 GAMING OC 8GB GDDR6 Grafikkarte
Intel P4510 2TB PCIe Gen3 x4 NVMe U.2 2.5" Enterprise SSD SSDPE2KX020T8
Asus GeForce RTX2080 Strix 8GB GDDR6 Gaming Grafikkarte
1TB 2,5" Intel DC P4500 Enterprise 24/7 Raid NVMe SSD 3200Mb/s +NEW+
MSI NVIDIA GeForce RTX 4060 Ventus 2x Black 8g OC Gaming Grafikkarte Top Zustand
WNEU! ASUS ROG Strix Gaming GeForce RTX 3070 OC 8GB GDDR6 Grafikkarte
RTX 2080 Ti 11GB GDDR6 Grafikkarte Dell Blower Gaming GPU Top
ASUS Dual GeForce RTX 3050 Edition 6GB GDDR6 Gaming Grafikkarte
SK hynix DDR4 RAM SODIMM 16gb 2x 8gb 3200mhz
Arbeitsspeicher G-SKILL AEGIS F4-3200C 16D-16GIS 2x8GB DDR4 16GB KIT (lesen)
GIGABYTE B760 DS3H AX ATX Mainboard Intel LGA1700 DDR5 4 Slots HDMI
ASRock B760 Pro RS WiFi LGA 1700 ATX Intel Mainboard DDR5 ATX Motherboard
MSI PRO B760M-P DDR4 LGA 1700 MicroATX Intel Mainboard
ASUS ROG Strix Gaming Grafikkarte 8GB Nvidia Geforce RTX 2060 super
ASUS TUF Gaming GeForce RTX 3070 Ti OC 8GB GDDR6X Grafikkarte
GIGABYTE GeForce RTX 2080 GAMING OC Grafikkarte 8 GB GDDR6
NVIDIA Geforce RTX 2060 Super 8GB GDDR6 Gainward GHOST Gaming PC Grafikkarte
Grafikkarte Rtx 3060 ti gaming trio z
Intel Xeon E5-2620 v4 2.10GHz LGA2011-3 8-Core Server CPU SR2R6
PROZESSOR CPU INTEL XEON E5-1620 v3 3.5GHz SR20P LGA2011-3
Crucial 16GB DDR4-3200 RAM UDIMM (CT16G4DFRA32A) Verschiedene Samsung 16GB 8GB 4
ASUS Dual GeForce RTX 3050 6GB GDDR6 Gaming Grafikkarte Nvidia Grafikkarte HDMI
ASUS Prime Z690-P LGA 1700 ATX Intel Mainboard
Intel HP Optane 375GB P4800X U.2 SFF Enterprise NVMe PCIe P02559-001 878014-B21
ASUS PRIME B760-PLUS Mainboard Sockel LGA 1700 DDR5 PCIe 5.0 ATX B-Ware
Palit GeForce RTX2070 Dual Fan 8GB GDDR6 NVIDIA GPU Gaming Grafikkarte -sehr gut
ASUS PRIME B760M-K Intel B760 LGA 1700 micro ATX
EVGA GeForce RTX 3050 XC Gaming 8GB GDDR6 Grafikkarte G-Sync VR-Ready HDMI 3x DP
Frequently asked questions
RTX 3090 or 4090 for LLMs?
3090 is the price-performance king for LLM: 24 GB VRAM, used ~600-800€. 4090 is 50% more tokens/sec but costs double and has the same VRAM. For most local setups the 3090 is the right pick unless you do throughput-critical inference.
Do I need datacenter GPUs (Tesla, A100)?
Tesla P40 (24 GB, ~250€ used) is a cheap entry to 24-GB VRAM but slow (Pascal architecture, poor quantization support). A100/H100 are out of hobby budget. Consumer GPUs (3090/4090) are today's sweet spot.
How much RAM do I need?
For pure GPU inference: 32 GB is enough. If you load models to system RAM first then push to GPU (or use CPU offloading): 64+ GB. For training or datasets: 128+ GB. RAM is cheap, err on the side of more.














































