GPU Integration & Risers

The graphics subsystem (GPU) is the “heart” and the primary source of liquidity for any DePIN node. Our goal at this stage is to deploy a computing array consisting of four graphics cards in such a way as to eliminate any hardware bus failures, ensure stable cooling under continuous 24/7 load, and squeeze the maximum performance out of each chip.

The line between a node’s commercial success and its downtime lies in the right choice of GPU architecture and uncompromising switching quality.

1. Choosing a GPU: Why the NVIDIA RTX 4090 / 5090 Is the “Gold Standard” for DePIN

There is fierce competition in modern distributed networks (AI, rendering, ZK proofs). We deliberately prioritize NVIDIA’s flagship consumer graphics cards (RTX 3090 / 4090 / 5090), and here is the in-depth engineering rationale behind this choice:

• Commercial Liquidity Factor (CUDA Ecosystem): 95% of computing power customers provide their own ready-to-run code written for the NVIDIA CUDA library. Our node must run this code “out of the box” without workarounds or recompilation. AMD cards (Instinct or Radeon series) are technically powerful, but the ROCm software ecosystem still requires manual tweaking for many tasks, which deters potential customers.
• “Here and Now” Availability: Industrial-grade accelerators such as the NVIDIA H100, A100, or the latest AMD chips are allocated via closed pre-orders among IT giants for years to come, and their retail price is comparable to that of an apartment or a luxury car. Using a pool of four RTX 4090 / 5090 cards allows you to build a server with a comparable amount of high-speed memory (96 GB of VRAM in total) at a reasonable price and with a reasonable payback period.
• Comparison with server-grade graphics cards (Tesla / Quadro): Specialized NVIDIA Tesla-class cards (such as the older A100 or the newer L40S) cost several times more than consumer-grade flagship models. However, for AI inference (generation) and traditional rendering tasks, the base clock speed and architectural improvements of the Ada Lovelace series (RTX 4090) or Blackwell (RTX 5090) series often outperform “pure-bred” server accelerators from previous generations, making consumer-grade flagship models more cost-effective for DePIN.

2. The Monolith Problem: Why Maps Must Be Identical

When setting up a GPU pool, it is critical to use graphics cards that are exactly the same (for example, 4 ASUS ROG Strix cards or 4 MSI Gaming X Trio cards).

• BIOS and Timing Compatibility: Cards from different brands have different frequency profiles, power limits, and video memory timings. If you mix them in a single server, the PCIe bus controller and the DePIN software engine will encounter load imbalance. One card will process its data array faster, while the others will “idle” while waiting, which will inevitably lead to a drop in overall performance and regular NVIDIA driver crashes (errors such as NV_TIMEOUT).
• Preparing Used Cards (“Engineer Access”): If cards are purchased on the secondary market, a full inspection is required before installing them in the server. Factory thermal pads quickly degrade and lose their properties when operating 24/7. It is necessary to completely replace the thermal paste with a top-of-the-line compound (with high thermal conductivity) and install high-quality thermal pads on the VRM area and memory chips to prevent localized overheating.

3. Switching: The Deadly Trap of “Miner” Risers

Since the PCIe x16 slots on our Supermicro X11-DAI-N motherboard are positioned right next to each other, it’s physically impossible to install four bulky three-slot graphics cards directly onto the board. They must be spaced out within the case. And this is where beginners often make their biggest mistake—using mining risers (adapters with a thin USB cable).

• Why mining risers are a no-go: In mining, a graphics card receives terabytes of identical, simple tasks where the data transfer speed to the CPU isn’t important, so mining risers physically operate in PCIe x1 mode. In DePIN tasks (neural network training, rendering), massive amounts of data are shuttled back and forth between the CPU, memory, and GPU. By limiting the bus to x1, you create a massive “bottleneck.” The card will sit idle, and the system will experience a critical drop in computing power.
• Professional Solution: We use only specialized, shielded flexible riser cables that support full data transfer speeds. In the specification search bar, this is specifically listed as: “PCIe 4.0 x16 Riser Cable high speed.”

Engineering Guidelines for Installing Loops:

Parameter	Technical Requirement	Why Is This Important?
Interface Type	Strictly PCIe 4.0 (or 5.0 for the RTX 5090)	Ensures maximum bus bandwidth without compromising FPS or AI training speed.
Tire width	Full-size x16 (physical and electronic)	Ensures that all available Xeon processor lanes are utilized for direct data exchange.
Cable Construction	Multilayer Electromagnetic Shielding	Four powerful cards generate massive electromagnetic interference. Without shielding, the signal in the loop will be distorted, causing a BSOD.
Laying Geometry	No sharp bends (90 degrees along the live loop)	Sharp bends damage the extremely fine copper wires inside the cable, causing a loss of contact or causing the slot to switch to x4/x8 mode.

The Golden Rule of Connectivity: Each graphics card must be connected directly to its native PCIe x16 slot on the motherboard via a single high-quality cable. No splitters, cheap adapters, or “spliced” cables. The cable connector must fit snugly into the motherboard slot until you hear the characteristic click of the lock, and the graphics cards themselves must be securely fastened with screws to the mounting frame of the 4U chassis, completely eliminating any sag or vibration.