X570 · Ryzen 9 5950X · Dual RTX 3090

The primary homelab machine. Built for AI inference and local model serving, but it runs Proxmox and handles general homelab duties too.

Specs

Motherboard:  MSI MAG X570 TOMAHAWK WIFI
CPU:          AMD Ryzen 9 5950X  (16 cores, 32 threads, 3.4 GHz base / 4.9 GHz boost)
RAM:          64 GB DDR4
GPU:          2× NVIDIA RTX 3090  (24 GB GDDR6X each, 48 GB total)
Storage:      2× M.2 NVMe
              2× 1 TB SATA HDD  (local backup)
              3× WD Red 8 TB NAS HDD  (NAS pool, ~24 TB usable with redundancy)

Why This Platform

The Ryzen 9 5950X was the highest core-count AM4 consumer CPU AMD shipped before the move to AM5. 16 cores and 32 threads on a mature, affordable platform made it a better value than chasing AM5 at the time this was built.

The dual 3090s are the reason the machine exists for AI work. Each card carries 24 GB GDDR6X. With both installed, large models that do not fit on a single card can be split across 48 GB of combined VRAM using tensor parallelism or multi-GPU inference in llama.cpp or vLLM.

PCIe Topology

The X570 platform is the first consumer platform with native PCIe 4.0. The Ryzen 5000 series (Zen 3) ships 24 CPU-direct PCIe 4.0 lanes.

Ryzen 9 5950X (24 CPU-direct PCIe 4.0 lanes)
  |
  ├── x16  PCIe 4.0   GPU #1 (RTX 3090)   ← full bandwidth, 32 GB/s
  ├── x4   PCIe 4.0   M.2_1 NVMe          ← CPU-direct, 8 GB/s
  └── x4   PCIe 4.0   → X570 Chipset link (upstream)
 
X570 Chipset (PCIe 4.0 x4 upstream to CPU = ~8 GB/s shared ceiling)
  |
  ├── x4   PCIe 4.0   GPU #2 (RTX 3090)   ← chipset-attached, ~8 GB/s
  ├── x4   PCIe 3.0   M.2_2 NVMe          ← chipset-attached
  ├──      SATA        5× HDDs
  └──      USB / WiFi

The critical detail: the second RTX 3090 is not in a bifurcated CPU lane arrangement. The MSI MAG X570 TOMAHAWK WIFI routes its second physical x16 slot through the chipset at x4 electrical. That means GPU #2 runs at PCIe 4.0 x4 — about 8 GB/s — not the x16 or x8 a dual-CPU-lane platform would provide.

For gaming this would be a hard constraint. For inference it is workable. The GPU's internal memory bandwidth (936 GB/s on the 3090) dominates inference performance. The PCIe link is primarily used for loading model weights and transferring intermediate tensors — operations that happen in bursts, not sustained.

The Dual-GPU Reality

Running two GPUs for inference is not plug-and-play. A few things to understand:

Single GPU inference:
  Model fits in 24 GB  →  runs fully on GPU #1  →  fast
  Model exceeds 24 GB  →  layers split to CPU RAM  →  slowdown at PCIe boundary
 
Dual GPU inference (tensor parallel):
  Model split across both cards
  GPU #1 ↔ GPU #2 communication: NVLink or PCIe
  RTX 3090: no NVLink bridge (consumer card)
  Inter-GPU traffic goes: GPU #1 → PCIe → CPU → PCIe → GPU #2

The RTX 3090 does not have NVLink on consumer configurations (unlike the RTX 3090 Ti or A-series). Inter-GPU communication falls back to PCIe. With GPU #2 on a x4 chipset lane, that path is further constrained.

The practical outcome: tensor-parallel inference on this setup is slower than it would be on a server platform with NVLink or a dual-socket board with both GPUs on CPU-direct lanes. It still runs. For large models that simply cannot fit in 24 GB, splitting is the only option. For models that fit on a single card, running on GPU #1 alone is faster.

Storage Topology

M.2_1 (CPU-direct, PCIe 4.0 x4):  OS drive, Proxmox, VM primary storage
M.2_2 (Chipset, PCIe 3.0 x4):     Overflow VM storage, scratch
 
SATA pool:
  2× 1 TB HDD  →  local backup target (rsync, Proxmox backup jobs)
  3× WD Red 8 TB  →  ZFS RAIDZ1 pool (~16 TB usable) for NAS / bulk storage

The SATA pool is fully chipset-attached. All five drives share the X570 chipset's upstream bandwidth to the CPU. For the workloads involved — backup writes and NAS reads — this is not a bottleneck. The HDDs themselves are the limiting factor, not the PCIe path.

Use Cases

• Large model inference via llama.cpp and Open WebUI
• Proxmox hypervisor running multiple VMs and LXC containers
• NAS storage for the wider homelab network
• Local backup target for other machines

Related Pages

• GPU AI Overview
• Common Misconceptions — why GPU #2 at x4 is not as bad as it sounds
• PCIe Devices — the x8 vs x16 GPU performance numbers