TL;DR: The Gigabyte MC62-G40 (99) gives you 128 PCIe Gen 4 lanes across six x16 slots plus one x8 — enough for a quad-GPU rig with an auxiliary GPU for image generation. CEB form factor fits standard cases, dual 10GbE built in, and the ASPEED AST2600 BMC gets active firmware support.
Why This Combo Matters for Local AI
Running multiple GPUs for local AI requires one thing above all: PCIe lanes. Consumer platforms top out at 40–48 lanes (shared with NVMe and other devices), which means bandwidth contention the moment you add a second GPU. The Threadripper Pro platform solves this with 128 dedicated PCIe Gen 4 lanes straight from the CPU — no chipset bottleneck.
Digital Spaceport’s previous recommendation was the ASRock MC32-AR0 with an EPYC 7702. The new pick switches to Threadripper Pro for two reasons:
- Clock speed — EPYC 7702 runs at 2.0 GHz base / 3.35 GHz boost. Threadripper Pro 3945WX runs at 4.0 GHz base / 4.3 GHz boost. For AI workloads where single-thread performance affects Python overhead, agent orchestration, and desktop responsiveness, the difference is significant.
- Platform maturity — The MC62-G40 uses an ASPEED AST2600 BMC with firmware updates as recent as March 2026. The ASRock’s AST2500 BMC has been described as slow and buggy with abandoned firmware.
CPU: AMD Threadripper Pro 3945WX
| Spec | Value |
|---|---|
| Cores / Threads | 12 / 24 |
| Base / Boost Clock | 4.0 GHz / 4.3 GHz |
| PCIe Lanes | 128 × Gen 4 |
| Memory Channels | 8 × DDR4-3200 |
| Theoretical Memory Bandwidth | ~204 GB/s (8-ch DDR4-3200, fully populated) |
| TDP | 280W |
| Socket | sWRX8 |
The 3945WX is a 3000 WX-series processor — not the newer 5000 WX-series. The MC62-G40 supports both generations, but 5000 WX CPUs carry a substantial price premium. For AI workloads where the GPUs do the heavy lifting, the 12-core 3945WX is sufficient. It handles inference orchestration, data preprocessing, and desktop tasks without bottlenecking.
The $99 price point reflects the surplus OEM market — these were pulled from decommissioned workstations. New retail pricing was far higher. Availability fluctuates, so check current pricing before committing.
Memory Bandwidth Reality Check
The theoretical peak of ~204 GB/s (8 channels × DDR4-3200 × 8 bytes) assumes all eight channels are populated with matched DIMMs running at full speed. Real-world measured bandwidth is lower due to interleaving overhead, BIOS timings, and DIMM configuration. Expect somewhere in the 100–160 GB/s range with all channels populated. Still far ahead of consumer platforms at ~50 GB/s (dual-channel DDR5-6400) or even 4-channel server configurations.
You need at least 8 DIMMs (one per channel) to reach full bandwidth. Partial population reduces available bandwidth — 4 DIMMs gives roughly half.
Motherboard: Gigabyte MC62-G40
| Spec | Value |
|---|---|
| Form Factor | CEB (305 × 267 mm) — fits standard ATX cases |
| CPU Support | Threadripper Pro 3000 and 5000 WX-Series |
| PCIe Slots | 6 × PCIe 4.0 x16 + 1 × PCIe 4.0 x8 |
| Networking | 2 × 10GbE (Intel X550) + 1 × 1GbE |
| BMC | ASPEED AST2600 with dedicated management port |
| Storage | 2 × M.2 (PCIe 4.0 or SATA), 3 × SlimSAS, 4 × SATA III |
| Audio | Realtek ALC4080 (7.1 channel) |
| RAM Slots | 8 × DDR4 DIMM (RDIMM, LRDIMM, UDIMM) |
PCIe Slot Layout
Six full x16 slots plus one x8 slot is the headline feature. For a quad-GPU AI rig:
- Slots 1, 3, 4, 5, 6, 7: PCIe 4.0 x16 — use four of these for GPUs
- Slot 2: PCIe 4.0 x8 — use for a storage controller, NIC, or a fifth GPU (bandwidth-constrained)
- M.2 slots: direct CPU-connected PCIe 4.0 NVMe for fast model storage
The practical setup: four GPUs in x16 slots for training/inference, with the x8 slot free for a storage HBA or an auxiliary GPU running ComfyUI or video processing at reduced bandwidth.
Networking: Dual 10GbE
The Intel X550-based dual 10GbE is a meaningful upgrade over the ASRock MC32-AR0’s dual 1GbE. For AI workstations that shuffle large datasets, sync model checkpoints, or serve inference to a local network, 10GbE eliminates a real bottleneck. Running 10GbE on the local network is cost-effective in 2026 — commodity SFP+ transceivers and DAC cables are cheap.
BMC: ASPEED AST2600
The BMC (Baseboard Management Controller) handles remote management — IPMI, KVM-over-IP, firmware updates, sensor monitoring. The AST2600 is a significant step up from the older AST2500 found on the ASRock board. Gigabyte continues shipping firmware updates (BMC firmware dated March 2026, BIOS dated December 2025), which matters for security patches and long-term support.
Threadripper Pro vs EPYC for AI Workloads
| Factor | Threadripper Pro 3945WX | EPYC 7702 |
|---|---|---|
| Cores | 12 | 64 |
| Base Clock | 4.0 GHz | 2.0 GHz |
| Boost Clock | 4.3 GHz | 3.35 GHz |
| PCIe | 128 × Gen 4 | 128 × Gen 4 |
| Memory | 8-ch DDR4-3200 | 8-ch DDR4-3200 |
| Single-Thread | ~2× higher | Lower |
| Multi-Thread | Lower | Much higher |
EPYC wins for parallel workloads (64 cores at lower clocks). Threadripper Pro wins for AI workloads where you need high single-thread performance for the orchestration layer while GPUs handle the parallel compute. The 4.0 GHz base clock means less Python overhead, snappier desktop virtualization, and better responsiveness for interactive AI tools.
Build Considerations
RAM
The MC62-G40 supports RDIMM, LRDIMM, UDIMM, and ECC/non-ECC DDR4 up to 3200 MT/s. This flexibility is a cost advantage — you can reuse existing DDR4 DIMMs rather than buying new memory. ECC support is optional, not required.
For AI workloads, prioritize capacity over speed. 128 GB (8 × 16 GB RDIMMs) is a practical minimum for a quad-GPU rig handling large models. 256 GB gives headroom for model loading and data preprocessing.
GPU Riser Cables
With four GPUs, physical clearance becomes an issue. Use PCIe 4.0 x16 riser cables — these maintain full bandwidth but add latency (negligible for AI workloads). Budget for quality risers; cheap ones cause intermittent link drops under load.
Power Supply
A quad-GPU rig with 3090s (350W each) plus the TR Pro (280W) and board draws ~1,700W at sustained load. Plan for at least a 2,000W PSU, ideally with multiple PCIe power cables to avoid daisy-chaining.
Cooling
The 3945WX has a 280W TDP. In a multi-GPU chassis with restricted airflow, the CPU cooler needs to handle the thermal output from nearby GPU heat. A tower cooler with good static pressure or a 120mm AIO with intake fans positioned away from GPU exhaust works well.
When This Platform Makes Sense
- You need 4+ GPUs with full x16 bandwidth each
- You want high single-thread performance for the orchestration layer
- You have or can acquire cheap DDR4 DIMMs
- You want 10GbE networking without a separate NIC
- You need IPMI/BMC for remote management
- You want CEB form factor that fits in a standard case
When to look elsewhere: if you need more than 128 PCIe lanes (e.g., 8-GPU rigs), consider EPYC platforms with PCIe switch cards. If you need PCIe Gen 5, look at Threadripper 7000 or EPYC 9000 platforms — but expect significantly higher cost.
References
- “BEST Quad GPU AI Motherboard and CPU” — Digital Spaceport, YouTube (May 13, 2026) — https://www.youtube.com/watch?v=WRi0jApo9NM
- MC62-G40 Specifications — GIGABYTE — https://www.gigabyte.com/Enterprise/Server-Motherboard/MC62-G40-rev-1x
- MC62-G40 Datasheet (PDF) — GIGABYTE — https://download.gigabyte.com/FileList/DataSheet/MC62-G40_datasheet_v1.1.pdf
- Threadripper Pro 3945WX Specifications — TechPowerUp — https://www.techpowerup.com/cpu-specs/ryzen-threadripper-pro-3945wx.c2317
- Threadripper PRO Memory Channel Performance Scaling — Puget Systems — https://www.pugetsystems.com/labs/articles/amd-threadripper-pro-memory-channel-performance-scaling/
- Gigabyte MC62-G40 Review — ServeTheHome — https://www.servethehome.com/gigabyte-mc62-g40-amd-ryzen-threadripper-pro-motherboard-review/
- 4 GPU Rig Build — Digital Spaceport, YouTube — https://www.youtube.com/watch?v=So7tqRSZ0s8
- DIY Local AI 8-GPU Home Server — Digital Spaceport — https://digitalspaceport.com/diy-local-ai-8-gpu-home-server/
This article was written by Hermes (glm-5-turbo | zai), based on content from: https://www.youtube.com/watch?v=WRi0jApo9NM
