DeepSeek V3 Update: The Dance of Computing Power and Algorithm

Recently, DeepSeek has made significant breakthroughs in the field of artificial intelligence models, launching the DeepSeek-V3-0324 version with a parameter count of 685 billion. This update significantly enhances the model's performance in areas such as coding capabilities, UI design, and reasoning abilities.

At the recently held 2025 GTC conference, a senior executive from a well-known technology company highly praised DeepSeek's achievements. He emphasized that the market's previous belief that DeepSeek's efficient model would reduce the demand for high-performance chips was incorrect. In fact, future computing demands will only continue to increase.

DeepSeek, as a model of algorithm innovation, has sparked extensive discussions in the industry regarding its relationship with high-performance computing hardware. This article will delve into the profound impact of Computing Power and Algorithm on the development of the artificial intelligence industry.

The Synergistic Development of Computing Power and Algorithm

In the field of artificial intelligence, the improvement of Computing Power provides a running foundation for more complex Algorithms, enabling models to handle larger-scale data and learn more complex patterns. At the same time, the optimization of Algorithms can utilize Computing Power more efficiently, enhancing the utilization efficiency of computing resources.

The symbiotic relationship between computing power and algorithms is reshaping the landscape of the artificial intelligence industry:

1. Technical route diversification: Some companies are dedicated to building ultra-large computing power clusters, while others focus on optimizing algorithm efficiency, forming different technical schools.

2. Industry Chain Restructuring: Some companies have become leaders in artificial intelligence Computing Power through ecosystems, while cloud service providers have lowered deployment thresholds with elastic Computing Power services.

3. Resource allocation adjustment: Enterprises seek a balance between hardware infrastructure investment and efficient Algorithm development.

4. The Rise of Open Source Communities: Open source models like DeepSeek and LLaMA enable the sharing of innovations in algorithms and achievements in Computing Power optimization, accelerating technological iteration and diffusion.

Technical Innovations of DeepSeek

The rapid rise of DeepSeek is closely linked to its technological innovations. Below is a simple explanation of its main technological innovations:

Model Architecture Optimization

DeepSeek adopts a combination of Transformer and MOE (Mixture of Experts) architecture, and introduces a Multi-Head Latent Attention (MLA) mechanism. This architecture functions like an efficient team, where the Transformer handles routine tasks, and MOE acts like a group of experts, each with their own specialization. When faced with specific problems, the most skilled expert handles it, significantly improving the model's efficiency and accuracy. The MLA mechanism allows the model to flexibly focus on different important details, further enhancing performance.

Methodology Innovation

DeepSeek has proposed the FP8 mixed precision training framework. This framework acts like an intelligent resource allocator that can dynamically select the appropriate computing precision based on the requirements of different stages during the training process. It uses higher precision when high-precision computation is needed to ensure the accuracy of the model; it reduces precision when lower precision is acceptable, thereby saving computing resources, improving training speed, and reducing memory usage.

Improvement in Inference Efficiency

During the inference stage, DeepSeek introduces Multi-token Prediction (MTP) technology. Traditional inference methods proceed step by step, predicting one Token at a time. The MTP technology can predict multiple Tokens at once, significantly speeding up the inference process while also reducing inference costs.

Reinforcement Learning Algorithm Breakthrough

DeepSeek's new reinforcement learning algorithm GRPO (Generalized Reward-Penalized Optimization) optimizes the model training process. It's like equipping the model with an efficient coach, guiding the model to learn better behaviors through rewards and penalties. Compared to traditional reinforcement learning algorithms, the new algorithm is more efficient, able to reduce unnecessary computing power while ensuring an improvement in model performance, achieving a balance between performance and cost.

These innovations form a complete technological system, reducing the Computing Power requirements across the entire chain from training to inference. Now, ordinary consumer-grade graphics cards can also run powerful artificial intelligence models, significantly lowering the threshold for AI applications, allowing more developers and businesses to participate in AI innovation.

Impact on High-Performance Computing Hardware

There is a view that DeepSeek bypasses certain software layers, thus reducing reliance on specific hardware. In fact, DeepSeek performs algorithm optimization by directly manipulating the underlying instruction set. This approach allows DeepSeek to achieve more precise performance tuning.

The impact of this approach on high-performance computing hardware manufacturers is twofold. On one hand, DeepSeek's deeper binding with hardware and the ecosystem may lower the threshold for AI applications and potentially expand the overall market size. On the other hand, DeepSeek's algorithm optimization may change the market demand structure for high-end chips; some AI models that originally required top-tier GPUs to run may now operate efficiently on mid-range or even entry-level graphics cards.

Significance to the Artificial Intelligence Industry

DeepSeek's algorithm optimization provides a new path for technological breakthroughs in the artificial intelligence industry. In the context of limited supply of high-end chips, the idea of "software compensating for hardware" alleviates the dependence on top imported chips.

Upstream, efficient algorithms reduce the pressure on computing power demand, allowing computing power service providers to extend hardware usage cycles through software optimization and improve return on investment. Downstream, the optimized open-source model lowers the barrier to entry for artificial intelligence application development. Many small and medium-sized enterprises can develop competitive applications based on the DeepSeek model without needing a large amount of computing power resources, which will lead to the emergence of more artificial intelligence solutions in vertical fields.

The Profound Impact of Web3+AI

Decentralized AI Infrastructure

DeepSeek's algorithm optimization provides new momentum for Web3 AI infrastructure. The innovative architecture, efficient algorithms, and lower Computing Power requirements make decentralized AI inference possible. The MoE architecture is naturally suited for distributed deployment, where different nodes can hold different expert networks without the need for a single node to store the complete model, significantly reducing the storage and computation requirements of a single node, thereby enhancing the flexibility and efficiency of the model.

The FP8 training framework further reduces the demand for high-end computing power, allowing more computing resources to be added to the node network. This not only lowers the threshold for participating in decentralized AI computation but also enhances the overall computing capability and efficiency of the network.

Multi-Agent Systems

1. Intelligent Trading Strategy Optimization: By analyzing real-time market data, predicting short-term price fluctuations, executing on-chain transactions, and supervising trading results through the collaborative operation of multiple agents, it helps users achieve higher returns.

2. Automation of smart contract execution: The collaborative operation of agents such as smart contract monitoring, smart contract execution, and execution result supervision enables the automation of more complex business logic.

3. Personalized Portfolio Management: Artificial intelligence helps users in real-time to find the best staking or liquidity provision opportunities based on their risk preferences, investment goals, and financial situation.

Conclusion

DeepSeek is breaking through under the constraints of Computing Power, seeking innovations through Algorithm to open up differentiated development paths for the artificial intelligence industry. Lowering application thresholds, promoting the integration of Web3 and AI, reducing reliance on high-end chips, and empowering financial innovation, these impacts are reshaping the digital economy landscape. In the future, the development of artificial intelligence will no longer be just a competition of Computing Power, but a competition of collaborative optimization between Computing Power and Algorithm. On this new track, innovators like DeepSeek are using wisdom to redefine the rules of the game.