Understanding Verification

​

How Verification Works

Verification is fully automated and powered by proprietary algorithms that continuously evaluate each machine’s operational health and performance. Only machines that meet the platform’s defined reliability and performance thresholds are marked as Verified. This process involves no manual intervention, ensuring consistent, scalable, and objective verification across all systems.

---

To qualify, a machine must pass minimum baseline, and health/stability checks. Beyond that, the system evaluates four primary criteria (order not indicative of priority):

​

1) Reliability

Definition: Stable, uninterrupted operation over time (uptime, resilience under continuous workloads). Do

• Maintain consistent uptime with minimal downtime.
• Keep network connectivity stable; avoid jitter and drops.
• Manage thermals and power to prevent throttling.
• Proactively monitor hardware health and perform maintenance.

Avoid

• Frequent restarts or unplanned outages.
• Overheating, undervolting, or unstable power delivery.

Note: Higher reliability greatly improves verification likelihood. Sustained ≥99.99% (up to 99.9999%+) uptime is typically favored.

---

​

2) Infrastructure Configuration

Definition: Hardware, network, and software readiness to meet operational standards.

​

Hardware

• GPU: Type, memory, and count matter. Newer datacenter/workstation GPUs are prioritized (e.g., B200 > H200 >> 5090 > 4070).
• VRAM: More VRAM improves performance profiles.
• GPU Count: For the same GPU type, more GPUs increase verification likelihood (e.g., 8×5090 >> 2×5090 > 1×5090).
• PCIe Bandwidth: Adequate throughput is essential; bottlenecks depress DLPerf and overall performance.
• CPU: Favor strong, server-grade CPUs; actual measured performance matters.
• Storage: Both capacity and bandwidth (e.g., NVMe) impact responsiveness and reliability.

​

Network

• High-speed, symmetric, stable bandwidth is favored.
• Ensure required ports are open and accessible; a static IP helps.

​

Virtualization

• Enabling VM support significantly improves verification likelihood.

​

Software

• Drivers/CUDA must be correctly installed and compatible (use stable latest releases).
• Keep systems clean; run workloads via Create Job only.

​

System Optimization & Upgrades

• Balanced scaling matters (CPU/RAM/PCIe/bandwidth commensurate with GPU tier).
• Do not reduce hardware after creation (e.g., fewer GPUs/RAM) – this will trigger Deverified.
• Upgrades (adding GPUs/RAM) are allowed but may take time to reflect across the platform.

Do

• Verify GPU PCIe connections provide full bandwidth and are not throttled.
• Keep the latest drivers/CUDA aligned with workloads.
• Confirm required ports and end-to-end reachability.

Avoid

• Pairing high-end GPUs with under-provisioned CPU/RAM.
• Letting hidden background services consume resources.

---

​

3) DLPerf Score

Definition: Estimated GPU performance on typical deep-learning tasks (e.g., CNN/Transformer training) for cross-hardware comparison. Higher DLPerf improves verification odds. Read more Do

• Use the latest compatible drivers/CUDA.
• Eliminate PCIe, thermal, and power bottlenecks to maintain sustained clocks.

Avoid

• Misconfigurations that suppress benchmark performance.

---

​

4) Supply & Demand Analysis

Definition: Ongoing evaluation of market trends and renter behavior to surface configurations most likely to be rented. Implication: Machines aligned with active renter preferences—popular GPUs, sufficient VRAM, strong reliability, fast internet—are prioritized for verification to maximize utilization and profitability. Do

• Offer in-demand GPU models with adequate VRAM and balanced resources.
• Maintain strong reliability to remain attractive once listed.

Avoid

• Niche/mismatched configurations with low renter interest.

---

​

Quick Reference