/ Hardware / HYDRA
HYDRA: a 4×RTX 3090 inference rig that runs Gemma-4-31B at 76.9 tok/s
Built mostly on the used market. Serves a production agentic workload while doubling as the reference benchmark target on this site. Bill of materials below — each row links to the marketplace where we actually sourced that part.
Build cost
~$6,200
Power
~800 W under load · ~170 W idle
measured at the wall
Performance
76.9 tok/s decode · Gemma-4-31B-AWQ · single-stream
Serves
Maple agentic workflows + LLM Monitor bench rig
Bill of materials
| Component | Part | Price | Buy |
|---|---|---|---|
| GPU | NVIDIA RTX 3090 24GB Used market only. Look for Founders Edition or EVGA FTW3 for cleaner thermals. Each card runs at PCIe 3.0 x16 in this build (see motherboard note). | ~$890 each used | Find on eBay → |
| Motherboard | Asus Pro WS X299 Sage II (LGA 2066) CEB workstation board. PLX chips on the board feed all four PCIe slots at full x16 PCIe 3.0 to each GPU — no bandwidth split. The reason this specific board was chosen. | ~$475 | Find on eBay → |
| CPU | Intel Core i9-7920X (LGA 2066, 12-core / 24-thread) Skylake-X. Socket LGA 2066 (Socket R4). 44 PCIe 3.0 lanes. CPU compute barely affects single-stream inference — the platform is chosen for the lane count. | ~$185 used | Find on eBay → |
| RAM | 16 GB DDR4 NON-ECC stick Eight matched 16 GB sticks for 128 GB total — full quad-channel population on the X299 platform. ~$224 across the eight slots. | ~$28 each | Find on eBay → |
| PSU | 1600W 80+ Titanium PSU (dual) Dual-PSU rig. Even though measured load is ~800 W at the wall, dual 1600 W units keep each PSU well below its peak-efficiency knee and tolerate a single-PSU failure. Headroom is cheap insurance on used parts. | ~$320 each | Find on eBay → |
| Case | 6U open-frame mining/AI chassis (dual-PSU support) Open-frame 6U chassis sized for 4–6 full-length GPUs with built-in dual-PSU mounting. The same form factor crypto rigs use; for inference duty it gives you the GPU spacing and airflow that a standard ATX case can't provide. | ~$200 | Find on eBay → |
| Storage | Samsung NVMe SSD 1TB Model weight cache + OS. Bumping to 2 TB worth considering once you cache more than two large GGUFs. | ~$85 | Find on eBay → |
| Cooling | Thermalright 240mm AIO CPU loop. The 7920X is happy under any 240mm class AIO at inference loads. | ~$75 | Find on eBay → |
| Cooling | Noctua NF-F12 / NF-A12 industrialPPC 120mm Case + GPU stack airflow. Industrial PPC variant for the static pressure needed across packed 3090s. | ~$30 each | Find on eBay → |
| Risers | PCIe Gen5 60mm risers Lets the GPU stack physically clear in an open-frame layout. Gen5-rated only because Gen4/Gen3 ones drop signal at 4090 / 3090 power loads on long runs. | ~$45 each | Amazon |
| Cabling | DisplayPort 1.4 extension cable 3ft (8K-rated) One per GPU. Lets you mount the rig in a way where the cards face into the case but the DP outputs route cleanly to a switch or breakout panel. | ~$12 each | Amazon |
| Misc | DisplayPort headless ghost dongles (3-pack) EDID emulators that spoof a connected monitor on each GPU. Required if you ever want full clock-rate behaviour on cards running headless — some workloads (and some driver versions) downclock GPUs that have no display attached. | ~$14 per 3-pack | Amazon |
| Software | NVIDIA driver 580 + CUDA 13.0 + vLLM 0.17 Stack as of v15. Pin versions until benched. | free | — |
eBay rows pull a daily median across the cheapest Buy-It-Now listings sampled. Amazon rows show our actual paid price at time of build — Amazon's pricing is stable enough that a daily probe isn't worth the API overhead. Click through to verify current pricing before purchasing.
In the flesh
A few photos from the build and testing phases. Front to back, lit up running, then back through the build.
Eagle-eyed readers will notice an RTX 3090 Ti in the testing photos. The Ti was a stand-in during bring-up; HYDRA's production lineup is four 3090s, which is what the bill of materials and the benchmark numbers reflect.
Not ready to commit $6,200?
Rent the equivalent on cloud GPUs first
Test your workload on a single A100 or H100 before deciding whether a 4×3090 build pays back. Both platforms below offer per-second billing.