In the fast-moving world of Web3 and AI infrastructure, innovation is everything. But every now and then, someone skips the hard work and simply copies. It’s an easy approach — which is unfortunately, far too common. Today, we’re calling out a project that crossed the line from inspiration into outright theft.

The platform in question is realium.xyz, a self-proclaimed “AI-powered real-world asset tokenization protocol” on Polygon. It appeared out of nowhere in June 2025 with a sleek interface and bold claims about tokenized sovereign bonds, streaming yields, and decentralized governance. But underneath the glossy UI lies something disturbingly familiar — because almost every architectural and conceptual element appears to be lifted from GPUAI.me, a decentralized GPU rental platform that’s been building in the open for over a year.

This is not a vague resemblance, “inspiration,” or a shared category — but rather a direct, strategic copy-paste plagiarism of our project, and today we’re going to expose it to our global community and the Web3 world.

The Evidence: Mirror Websites, Stolen Ideas

Plagiarism in Web3 isn’t just unethical — but rather corrosive.

The GPUAI.me project has been around for a while, offering decentralized GPU rentals, token-based computation rewards, staking for node operators, and a clear roadmap. The decentralized team built a platform for democratizing access to GPU resources — crucial in a time when AI demand is exploding and centralized GPU access is increasingly gated.

Then comes Realium.xyz. At first glance, it pretends to offer “tokenized real-world assets.” What’s particularly egregious is that Realium simply swapped out the word “GPU” with “asset,” rebranded the compute layer as “AI-powered risk engine,” and added some trendy ZK-flavored jargon. But the scaffolding is unmistakable. It sounds innocent, even unrelated. But dig one layer deeper — and you’ll see the rot.

Here’s what they’ve copied:

Identical core concepts: Decentralized infrastructure, tokenized compute power, staking mechanisms, DAO governance.

• Reworded but clearly lifted copy: Marketing phrases that sound eerily familiar. Realium didn’t just get “inspired” — they took our project’s structure and rephrased it like a lazy school assignment.
• Whitepaper structure: Multiple sections in Realium’s documentation are direct analogs of GPUAI’s docs — same flow, same concepts, just repackaged.
• Roadmap mimicry: Realium’s “upcoming features” list seems to shadow GPUAI’s roadmap almost exactly — AI task scheduling, federated compute models, and token unlock timelines.

To sum things up, this isn’t a coincidence. It’s a bold theft.

On the left: a sleek, thoughtfully crafted original.

On the right: a rushed, derivative knock-off.

Spot the difference? Neither can we.

Realium’s Shell Game: A Shiny Mask Over Plagiarism

The Realium.xyz domain was registered in June 2025. That’s over a year after GPUAI.me’s first public testnet and nine months after their roadmap was first published. Even the metadata on Realium’s site suggests it was built using clone site templates commonly used by fast-fork scams.

The worst part? Realium wraps itself in a glossy interface, throwing around buzzwords like “AI-powered Risk Engine,” “ZK Proofs,” “Streaming Yield” and “Tokenizing Real Assets” to confuse audiences. However, under the hood, the mechanics mirror GPUAI’s original architecture almost line-by-line.

Imagine stealing someone’s car and spray-painting it a different color. Even more disturbing, Realium is actively misleading users into thinking it is pioneering tech. In reality, it’s offering nothing new — just recycled infrastructure from GPUAI with a new font and some AI-flavored frosting.

Why It Matters

This is not just a petty branding issue but about trust. Web3 depends on open innovation and fair contribution. When projects like GPUAI put in the work — writing code, running test nets, gathering community — they do it to move the space forward. And when someone like Realium comes along, rips that work, and tries to cash in without attribution or contribution, it undermines the entire ecosystem.

Worse still, this isn’t a victimless act. Realium’s attempt to pass off this stolen framework as their own:

• Confuses investors
• Misleads developers
• Disrespects open-source builders
• Creates a false market narrative
• And undermines the reputation of serious protocols.

Do you need more points? Time to act.

Final Verdict

Realium.xyz is not an original project. It is, by every observable metric, a repackaged derivative of GPUAI.me — launched with a time delay, stripped of credit, and rewrapped in misleading marketing.

We’ve seen this before. Projects try to leech off others’ work to make a quick buck. But if we let it slide, it becomes normal. If you’re part of the AI, DePIN, or Web3 infrastructure space, this is a line in the sand.

• If you’re an investor: Do your due diligence. Ask Realium’s team about their “inspiration.”
• If you’re a builder: Speak up. Don’t let authentic work be buried under fakes.
• If you’re a user: Don’t fall for clones. Support the originals.

Builders like GPUAI put in real work. They define specs, run devnets, fix bugs, and engage with the community. Their reward? A lookalike competitor with more aggressive marketing and none of the technical chops.

For users, it creates confusion. For investors, it creates false signals. And for the broader ecosystem, it’s a signal that scammers can copy-paste their way into serious capital flows.

Realium isn’t a “rival project.” It’s not even a fork. It’s a synthetic overlay, a mimic — designed to siphon off attention, capital, and developer time that should rightfully belong to the original creators.

By every observable metric, it’s a repackaged derivative of GPUAI.me — launched with a time delay, stripped of credit, and rewrapped in misleading marketing. It might fool casual users, but it won’t fool the builders paying attention.

We see you, Realium. And now, so does everyone else. The Web3 world doesn’t need more hype or more copycats. It needs builders, ethics and call out grifters — loudly.

Meanwhile, to learn more about GPU.Ai & the latest project updates, and also support the team, visit the website, follow on X, join the Telegram community, and more socials. Stay vigilant, stay safe.

1. Multi‑Channel Idle GPU Discovery

Node Self‑Registration & Declarative Probing

• Local Client Registration
  Users install the GPUAI node agent (CLI or GUI) and declare their machine’s GPU specs — model, VRAM, driver version, network bandwidth — and desired availability windows.
• Declarative Probing
  Administrators configure thresholds (e.g. GPU utilization < 20%) or time slots (nights, weekends). When conditions are met, the agent automatically advertises those idle periods to the network.

Enterprise & Research Partnerships

• Bulk API Onboarding
  GPUAI integrates with university clusters, corporate HPC, and private labs via secure REST APIs, importing batch inventory and real‑time idle metrics.
• Containerized Proxy Deployment
  A lightweight Docker/Sandbox proxy can be deployed in existing on‑prem or private‑cloud environments, streaming idle GPU data without requiring infrastructure changes.

Cloud Instance Scanning

• User‑Authorized Scans
  With explicit permission, GPUAI agents invoke cloud provider APIs (AWS, GCP, Azure) to query GPU instance utilization and metadata tags. Idle instances are elastically added to the pool.

2. Real‑Time Health Monitoring & Management

Heartbeat Protocol

• Periodic Telemetry
  Every few seconds, nodes emit heartbeats containing GPU utilization, core temperature, power draw, memory usage, and network latency.
• Decentralized Event Streaming
  Heartbeats are published over a libp2p PubSub mesh or light on‑chain event logs, visible to authorized schedulers but encrypted at rest.

Off‑Chain Logs & On‑Chain Summaries

• Merkle‑Root Anchoring
  Detailed telemetry is stored in decentralized storage (IPFS, Arweave). At regular intervals, a Merkle root is committed on‑chain for tamper‑proof auditing.
• Dynamic Performance Profiles
  Each node maintains a historical profile — uptime ratio, disconnect frequency, task completion time — used both for reputation scoring and scheduling decisions.

Fault Detection & Automatic De‑Listing

• Threshold Alerts
  If temperature or network jitter exceeds safe limits, the scheduler de‑lists the node in real time to avoid compromised tasks.
• Re‑Onboarding Post‑Recovery
  Once metrics return to nominal, nodes undergo multiple health checks and ZK‑proof validations before re‑entering the active pool.3. Intelligent Scheduling & Load Balancing

Demand Forecasting & Time‑Slot Planning

• Machine‑Learning Demand Models
  Historical job volumes and node health data feed demand‑prediction models, which generate optimized scheduling windows to minimize queue times.
• Preemption & Reservation
  High‑priority inference or latency‑sensitive tasks can reserve slots in advance, with lower‑priority jobs backfilled as capacity allows.

Multi‑Dimensional Load Distribution

• Geographic Sharding
  The network is partitioned by latency and regional proximity. Low‑latency inference runs locally; training jobs span shards based on bandwidth and cost trade‑offs.
• Heterogeneous Hardware Orchestration
  A hardware abstraction layer profiles NVIDIA, AMD, and specialized AI accelerators. Data‑ and model‑parallel jobs are decomposed dynamically to maximize throughput.

Energy‑ and Cost‑Aware Allocation

• Time‑Of‑Use & Carbon Signals
  Nodes can bid to run during off‑peak electricity rates or green‑energy periods. Smart contracts adjust token rewards to favor low‑carbon compute.
• Efficiency Bonuses
  Compute providers who deliver high FLOPS-per‑watt receive bonus $GPUAI tokens on top of base rewards.4. Roadmap & Future Directions
• Predictive Maintenance
  Integrate additional sensors (fan speed, voltage fluctuation) and anomaly‑detection algorithms to preempt hardware failures and cut maintenance costs.
• Cross‑Protocol Interoperability
  Collaborate with DePIN storage and bandwidth networks to offer unified, on‑chain compute-plus-storage marketplaces.
• Edge & IoT Integration
  Support micro‑GPU devices and embedded accelerators at the network edge for real‑time inference in autonomous vehicles and smart‑city applications.

Conclusion：

GPUAI’s technical stack — from multi‑channel resource discovery and heartbeat‑driven health management to ML‑powered demand forecasting and energy‑aware scheduling — forms a robust, fully decentralized “find‑manage‑schedule” loop. As this architecture matures and expands, millions of idle GPUs worldwide will power AI workloads at unprecedented scale, cost efficiency, and environmental sustainability, ushering in the truly shared‑compute era.

For decades, centralized computing has been the backbone of the digital world, and cloud computing has been dominated by big tech companies like AWS, Google Cloud, and Microsoft Azure. These giants control over 65% of the cloud market, setting high prices, limiting access, and putting businesses at risk of sudden outages.

But there’s a shift happening. A new paradigm is — decentralized computing is a faster, cheaper, and more secure alternative — and it’s set to change everything. Let’s learn how CPU AI tackles GPU shortages and high costs better than centralized cloud providers!

What’s Wrong with Traditional Cloud Computing?

Centralized institutions dominated the market for far too long. And their monopoly comes at a cost:

● It’s expensive. Renting an A100 GPU from AWS can cost over $4 per hour, making long-term AI training unaffordable for many startups.

● It’s unreliable. In 2023, AWS experienced 27 major outages, leaving businesses stranded and scrambling for alternatives.

● It’s centralized. A handful of companies control the majority of computing power, creating single points of failure and privacy concerns.

As AI, Web3, and 3D computing continue to grow, this outdated model just won’t cut it anymore. These limitations slow down innovation when decentralized computing is expanding at record speed.

A new approach is needed: decentralized computing. Powered by distributed networks, decentralized computing allows users to share, rent, and access computing power without relying on a single provider. This shift is reshaping industries, offering solutions that are more accessible, scalable, and resilient than traditional cloud services.

But with every innovation comes challenges. Decentralized computing has the potential to revolutionize AI, Web3, and high-performance computing, but it also faces technical and adoption hurdles that need to be addressed.

Enter Decentralized Computing

Decentralized computing works differently. Instead of relying on a single company’s data centers, it connects users to a global network of independent GPUs and CPUs. It’s like Airbnb for computing power — people with extra GPU capacity can rent it out, and businesses can access high-performance computing at a fraction of the cost.

Why Decentralized Computing is Taking Off:

● It’s 50–70% Cheaper

1) Cloud providers charge huge markups for GPU power.

2) Decentralized networks use idle GPUs from contributors worldwide, slashing costs.

3) AI model training can be 50% cheaper, freeing up budgets for innovation.

● No More Downtime

1) When AWS crashes, millions of businesses go offline.

2) Decentralized networks distribute workloads across thousands of nodes, meaning there’s always backup power available.

● Scale Instantly

1) Need more power for AI training or blockchain transactions? No problem.

2) Decentralized computing automatically scales, providing more GPU power when you need it — no waiting in line.

● Better Privacy & Security

1) No centralized storage = no massive data leaks.

2) With decentralized encryption, your data stays yours — no corporate tracking, no surveillance.

The Numbers Don’t Lie

The AI market is expected to reach $407 billion by 2027 and the Web3 industry is projected to grow at a CAGR of 43.7%! At the same time, the metaverse is on the rise — expected to reach $936 billion by 2030!

With demand for computing power skyrocketing, decentralized computing is stepping in to fill the gaps left by traditional cloud services.

So, what’s the catch for us here? Okay, decentralized computing isn’t perfect (yet).

There are still challenges to overcome:

�� Adoption barriers — Many businesses are still tied to centralized platforms.

✔ Solution? Easy integrations and developer-friendly tools.

�� Performance variability — Some nodes may be slower than others.

✔ Solution? AI-driven load balancing for optimized speeds.

�� Regulation & compliance — Some industries need strict data controls.

A possible solution? Hybrid models that balance decentralization with legal compliance.

Final Thoughts? The Future is Decentralized

The centralized cloud monopolies are holding back progress in AI, Web3, and high-performance computing. Decentralized computing breaks these barriers, offering a faster, cheaper, and more secure alternative.

Cloud computing isn’t disappearing, but the days of total reliance on big tech are over. AI, Web3, and metaverse applications need faster, cheaper, and more secure computing options — and decentralized computing is the answer.

Startups, developers, and enterprises are already making the shift. The question is: will you be ahead of the curve, or left behind? Platforms like GPU AI are proving that decentralized GPU networks can power AI, 3D rendering, and Web3 applications at scale.

Unlike centralized cloud providers that control GPU supply and set high prices, GPU AI taps into a global network of idle GPUs, unlocking underutilized computing power and making it available on demand at significantly lower costs. Instead of waiting for cloud providers to release limited GPU batches at premium rates, GPU AI instantly connects users with available resources, eliminating bottlenecks and ensuring consistent, affordable access to high-performance GPUs.

By decentralizing computing power, GPU AI is breaking the cloud monopoly, offering a scalable, cost-effective, and resilient solution for AI developers, 3D artists, and Web3 innovators. Learn more by checking our cutting-edge website and following us on socials!

The demand for high-performance computing is skyrocketing as artificial intelligence (AI), 3D rendering, and Web3 applications become increasingly complex. Traditional cloud computing platforms, dominated by centralized providers like AWS, Google Cloud, and Microsoft Azure, have long been the go-to solutions for computational needs. However, their limitations — high costs, lack of accessibility, and scalability bottlenecks — have created an urgent need for a new approach.

Decentralized GPU computing is emerging as the next revolution in AI and Web3 infrastructure. Platforms like GPU AI leverage a distributed network of GPU nodes to offer efficient, scalable, and cost-effective computing power, enabling AI researchers, blockchain developers, and 3D creators to unlock new possibilities.

Understanding Decentralized GPU Computing

Decentralized GPU computing operates on a distributed network, where individuals and organizations contribute their idle GPUs to form a global supercomputer. This model allows users to access GPU power on demand, often at a fraction of the cost of traditional cloud services.

The key advantages of decentralized GPU computing include:

● Cost Efficiency: Users can rent GPU power at lower rates compared to centralized providers.

● Scalability: The network dynamically adjusts to demand, ensuring availability of resources.

● Democratization of Compute Power: Anyone with a GPU can contribute and monetize their idle hardware.

● Resilience and Redundancy: Unlike centralized services that can experience downtime or regional failures, decentralized networks remain operational even if some nodes go offline.

By 2027, the AI market is expected to surpass hundreds of billions of USD in total value, while the Web3 industry is projected to grow at a CAGR of 43.7%. Grand View Research estimates that the global Web 3.0 market size, valued at $2.25 billion in 2023, is projected to reach $33.53 billion by 2030, expanding at a CAGR of 49.3% during the forecast period. Meanwhile, MarketsandMarkets forecasts that the global Web 3.0 market is projected to grow from $0.4 billion in 2023 to $5.5 billion by 2030. These technologies require massive computational power, yet traditional cloud services remain expensive, centralized, and often inefficient.

How Decentralized GPUs Power AI & Web3

AI and Web3 projects require immense computational resources, making decentralized GPU computing a game-changer in several ways:

1. AI Model Training: Cut Costs & Speed Up Execution

Training deep learning models requires thousands of GPU hours. With centralized cloud services, businesses can spend over $100,000 annually on AI processing. GPU AI’s decentralized model distributes workloads across multiple nodes, slashing training costs by up to 50% and boosting efficiency.

2. Real-Time AI Applications Without Delays

From chatbots like ChatGPT to self-driving cars, AI applications demand real-time processing. Decentralized GPU networks provide low-latency, high-speed execution, ensuring seamless AI performance without the risk of cloud provider throttling or outages.

3. 3D Rendering & the Metaverse: Accelerating Creative Workflows

The metaverse industry is expected to experience significant adoption as the global market size is projected to grow at a compound annual growth rate (CAGR) of 46.4% from 2025 to 2030, reaching approximately $936.57 billion by 2030. High-quality 3D rendering and VR applications need serious GPU power. Instead of buying expensive hardware, creators can rent decentralized GPUs for a fraction of the cost, reducing rendering times by up to 40%.

4. Web3 & Blockchain: Powering a Decentralized Future

Decentralized applications (dApps), NFT marketplaces, and blockchain games require constant computational power. Decentralized GPU networks help process smart contracts faster, validate blockchain transactions, and support scalable Web3 applications without relying on centralized entities.

Leading the Decentralized Compute Revolution

GPU AI is at the forefront of decentralized GPU computing, offering an efficient and scalable platform that supports AI model execution, 3D workloads, and Web3 applications. With over 320 GPU nodes globally and 97% efficiency in AI model execution, GPU AI provides a powerful alternative to traditional cloud services.

Key Features of GPU AI:

● Global Network: Access to a distributed infrastructure for seamless computing.

● Optimized AI Workflows: Reducing model training time by up to 30%.

● Cost-Effective Solutions: Competitive pricing compared to centralized cloud providers.

● Decentralized Ecosystem: Enabling contributors to earn by sharing their GPU power.

The Future of Decentralized GPU Computing

The shift toward decentralized computing is inevitable as the demand for AI and Web3 applications grows. By eliminating reliance on centralized providers, decentralized GPU platforms like GPU AI empower developers, businesses, and researchers with a more flexible and affordable alternative.

As AI innovation accelerates and Web3 adoption expands, decentralized GPU computing will become the backbone of next-generation digital experiences. Whether you’re training AI models, rendering complex graphics, or building blockchain applications, embracing decentralized GPU power is the key to unlocking new levels of efficiency and scalability.

Are you ready to join the future of decentralized computing? Explore GPU AI and experience the power of decentralized GPU technology today.

Introduction: The Rising Demand for AI Compute Power

AI is advancing rapidly, driving progress in healthcare, finance, automation, and language processing. However, these innovations require enormous computing power, especially GPUs, which are both costly and in short supply.

Major cloud providers like AWS, Google Cloud, and Azure dominate the GPU market, but they come with significant challenges:

• Expensive Pricing: High costs make large-scale AI training out of reach for many startups and independent researchers.
• Limited Availability: GPU shortages slow down development and stall innovation.
• Scalability Issues: Expanding in-house GPU infrastructure demands large investments and technical expertise.

To solve these issues, AI developers are moving towards decentralized GPU compute networks — a new model that uses blockchain and AI-driven optimization to make computing power more accessible and affordable.

The Challenge: Centralized Systems Have Limits

Despite the growth of cloud computing, traditional systems are struggling to keep up with AI’s increasing needs. Key problems include:

1. High Costs of GPU Compute

Cloud GPU services are expensive, making AI development difficult for smaller players. Usage-based pricing can quickly add up, especially for deep learning models requiring long training periods.

2. Scalability Constraints

Cloud providers impose limits on GPU availability, making it hard to scale AI projects. Building an in-house GPU setup is another option, but it requires a lot of money and resources.

3. Inefficient Resource Use

Current AI compute platforms often have inefficient scheduling, leading to:

• GPUs sitting idle, wasting resources.
• Long wait times due to GPU shortages.
• Big companies dominating access, leaving smaller players with fewer options.

4. Lack of Real-Time Optimization

Cloud GPU providers struggle to adjust workloads in real-time, leading to:

• Static allocation models that don’t adjust to changing needs.
• Poor use of resources due to inefficient workload distribution.
• Higher costs due to unnecessary over-provisioning.

There is a clear need for a cheaper, more flexible, and efficient way to access GPU power for AI development.

The Solution: A Decentralized GPU Compute Network

A decentralized GPU network offers a new way to access computing power. By connecting users to a global network of GPUs, AI developers can get resources when they need them without relying on a single provider. This system uses blockchain technology and AI-driven management to ensure efficiency, lower costs, and better performance.

Key Features of a Decentralized GPU Network:

1. Real-Time GPU Availability & Performance Insights

Unlike traditional providers, this model provides:

• Live updates on available GPUs.
• Performance details, including processing power, memory, and latency.
• Smart workload distribution to match users with the best GPUs.

2. AI-Powered Compute Optimization

AI algorithms improve efficiency by:

• Distributing workloads based on GPU performance and network conditions.
• Predicting demand and adjusting resource allocation automatically.
• Reducing training time and costs while maximizing performance.

3. Scalability for AI Workloads

AI projects need flexible scaling, which traditional cloud providers struggle to offer. A decentralized model:

• Allows instant scaling as demand increases.
• Supports parallel processing to speed up training.
• Prevents resource shortages, ensuring continuous operations.

Why This Model Works for AI Compute

🚀 Lower Costs: Users pay only for what they use, avoiding high cloud provider fees. 🚀 More Availability: A decentralized system ensures GPUs are always accessible. 🚀 Better Performance: AI-driven allocation improves efficiency. 🚀 Flexible Scaling: AI workloads can grow without hardware limits.

Technology & Architecture

This decentralized GPU compute network is built to connect AI developers with a global pool of GPU resources. The system is designed with several key components:

1. Pooling GPU Resources

By gathering GPUs from individual contributors, data centers, and cloud providers, the system:

• Avoids GPU shortages.
• Enables cost-effective sharing of computing power.
• Provides global access to high-performance GPUs.

2. AI-Driven Optimization

AI-based scheduling ensures:

• Efficient workload balancing.
• Smart predictions for resource needs.
• Lower costs and faster processing times.

3. Easy Integration for Developers

The platform supports AI development through:

• APIs and SDKs for quick setup.
• Popular ML tools like TensorFlow, PyTorch, and JAX.
• Containerized execution for flexibility.

4. Real-Time Monitoring & Alerts

Unlike cloud providers, this system offers full visibility with:

• Live tracking of GPU performance and energy use.
• Instant updates on workload distribution.
• Alerts for performance issues or GPU availability changes.

5. Automatic Scaling

The system ensures AI applications scale smoothly by:

• Adjusting GPU availability in real-time.
• Enabling parallel processing for faster AI execution.
• Adding resources on demand to prevent slowdowns.

Applications & Use Cases

This decentralized GPU compute network supports a range of AI-driven applications, making advanced computing more accessible.

🔹 Deep Learning Model Training

• Faster training times through parallel processing.
• Affordable access to high-powered GPUs.

🔹 AI-Based Image & Video Processing

• Scalable computing for tasks like image recognition and video analysis.
• Real-time AI applications in areas like medical imaging and self-driving cars.

🔹 Enterprise AI & Research

• Cost-effective AI compute resources for companies and research institutions.
• Scalable AI solutions for fraud detection, drug discovery, and automation.

The Future of AI Compute: Expanding Access & Scalability

As AI continues to grow, this decentralized network aims to:

• Increase the number of available GPUs by partnering with more providers worldwide.
• Improve AI-powered predictive scaling to better manage demand.
• Enhance compute analytics for better workload balancing and problem detection.

The long-term goal is to build a self-sustaining, decentralized AI compute network that: 🚀 Supports large-scale AI workloads with ease. 🚀 Provides on-demand compute power at a fraction of traditional cloud costs. 🚀 Uses AI-driven resource management for peak efficiency. 🚀 Expands into edge computing and real-time AI applications.

Conclusion: The Shift to Decentralized AI Compute

The move to decentralized AI computing is inevitable. This platform helps bridge the gap between AI demand and computing power, creating an efficient and scalable solution for AI developers.

🌍 Making AI Compute More Accessible — Removing barriers to high-performance computing. 💰 Lower Costs, Greater Efficiency — Making AI compute affordable for all. 🚀 Faster AI Development — Giving innovators real-time access to GPU power.

Be part of the future of AI compute. The change is happening now! 🚀

We welcome you to join our thriving community and explore the endless possibilities that GPUAI has to offer. Let’s build a more accessible and powerful decentralized future together!

website：https:/gpuai.me/
Telegram Groups：https://t.me/Official_GPUAI
Twitter：https://x.com/GPUAI_Coin
document：https://docs.gpuai.me/