Tirami: A Distributed LLM Inference Protocol with Computing Power as Currency

Tirami is a distributed LLM inference protocol implemented in Rust, with the core concept of "Computing Power as Currency". By quantifying computing into Compute Units (CU, each unit equals 10^9 FLOPs of verified inference), it builds an open infrastructure without tokens or ICOs, aiming to solve the problem of computing power centralization and promote computing power democratization.

Background: Urgent Need for Computing Power Democratization

With the rapid development of large language models, the demand for inference computing power has grown exponentially. However, the current computing power market faces severe centralization issues: a few tech giants control most of the high-performance GPU resources, while ordinary developers and small teams often face the dual dilemma of difficulty in obtaining computing power and high costs.

At the same time, there are a large number of idle computing resources worldwide—from personal workstations to GPUs in small data centers. These resources cannot be fully utilized due to the lack of an effective coordination mechanism. How to integrate these scattered computing resources to build an open, fair, and efficient inference network has become an urgent problem for the industry to solve.

Tirami Project Overview and Core Mechanism

Tirami is an open-source distributed LLM inference protocol developed by the clearclown team, with the core concept of "Compute is Currency". Implemented in Rust, the project provides an interface compatible with the OpenAI API, aiming to build a decentralized model inference infrastructure.

The project's most distinctive feature is its token-free design, using Compute Units (CU) as the medium of value exchange instead. Each CU represents 10^9 verified floating-point operations (FLOPs) and has the following characteristics:

• Verifiability: The computing volume of inference requests can be verified via cryptography
• Standardization: A unified measurement standard allows comparison and exchange of computing power across different hardware
• Practicality: Directly linked to actual workload, avoiding price fluctuations of cryptocurrencies

Technical Architecture and Implementation

Tirami is developed in Rust, leveraging its high performance and memory safety features. The project is fully compatible with the OpenAI API, allowing existing LLM applications to migrate at almost zero cost.

The protocol architecture includes key components:

• Inference Node: A service node run by computing power providers, responsible for executing inference tasks
• Verification Layer: Verifies the computing volume of inference tasks to ensure accurate CU measurement
• Routing Layer: Intelligently schedules user requests to appropriate nodes, optimizing latency and cost
• Settlement Layer: Completes value settlement between computing power supply and demand parties based on CU

Reasons and Advantages of Token-Free Design

Tirami explicitly states "No Token, No ICO". This design is based on the following considerations:

1. Eliminate Speculation: Token economies are prone to speculation, and price fluctuations affect user experience; CU is anchored to actual computing power, leaving no room for speculation
2. Reduce Regulatory Risks: Avoid strict regulatory restrictions on cryptocurrency issuance and focus on technology itself
3. Lower Participation Threshold: Users do not need to manage crypto assets, only focusing on actual computing power costs and benefits

Application Scenarios and Value

Tirami's distributed architecture brings value to multiple scenarios:

• Computing Power Demanders: Obtain more cost-effective inference capabilities, especially suitable for long-tail needs and sudden traffic
• Computing Power Providers: Convert idle GPU resources into revenue, improving hardware utilization
• Developer Community: Open protocols promote competitive innovation, break computing power monopolies, and drive AI democratization

Conclusion

Tirami represents a new idea combining technological and economic model innovation in the field of AI infrastructure. By quantifying computing power into verifiable and exchangeable standard units, it lays the foundation for an open distributed inference network.

Although the project is in its early stages, its "Computing Power as Currency" concept and practical token-free design provide a noteworthy alternative for the industry. With the development of distributed AI infrastructure, Tirami is expected to play an important role in the process of computing power democratization.