AI Technology Advancement

DeepSeek's AI Training Acceleration Technologies

DeepSeek has open-sourced two AI training acceleration technologies, DualPipe and EPLB, designed to reduce the cost and time of large model training.

• DualPipe enables simultaneous forward and backward computations and overlaps computation with communication, reducing idle time and improving data transfer efficiency during training.
• EPLB balances the workload between GPUs by dynamically adjusting expert assignments, optimizing the uneven resource allocation in the MoE architecture. DeepSeek-V3's pre-training cost was significantly lower than models of similar scale due to the adoption of DualPipe technology. The open-source training and inference framework also discloses performance analysis data, helping developers understand communication-computation overlap strategies and underlying implementation details to improve training speed and resource utilization.

Technology

DeepSeek R1 Model Integrated into ByteDance's Wukong Browser

ByteDance's Wukong Browser has officially integrated the DeepSeek R1 model.

AI & Tech

DeepSeek Advancements

• DeepSeek has open-sourced three algorithms aimed at improving computational performance.

AI Search Behavior

• An author noted that DeepSeek's search engine used their Geek Park article as a key reference when answering the question "What are good AI headphones like?". The AI adopted the article's interpretation of user needs, particularly highlighting its satisfaction of essential needs in diverse immigrant communities.
• In contrast, Tencent Yuanbao, which uses DeepSeek, suggested a startup brand, WISHEE AI, due to positive reviews in its referenced material.
• Experiments show that AI searches favor recent, relevant articles with clear judgments, even from personal accounts.

NVIDIA Financials and Future Outlook

• NVIDIA reported $39.3 billion in revenue for fiscal year 2025, Q4, a 77.9% year-over-year increase, though gross margin was slightly below expectations.
• Data center business growth drives revenue, up 93.3% YoY, while gaming declined by 11%.
• The company forecasts $43 billion in revenue for the next quarter, with a further decrease in gross margin.
• The market is focused on the progress of Blackwell, the impact of DeepSeek, and capital expenditure of major cloud providers.
• Blackwell contributed $11 billion this quarter, but slow GB200 scaling affected revenue and gross margin.
• DeepSeek's low-cost strategy may affect market expectations regarding NVIDIA's competitive advantages but could also stimulate application breakthroughs.
• Cloud service giants are maintaining high capital expenditure growth, but changes need to be monitored.
• The market will closely watch the progress of the next-generation Blackwell Ultra.

AI and Semiconductor Industry

Nvidia's Performance and Future Prospects

Nvidia's latest financial report reveals a significant revenue increase, driven by surging orders for the H20 chip and a promising outlook for the Blackwell platform. Demand for AI infrastructure is robust, with GPUs remaining the preferred choice. The company has addressed initial production defects in the Blackwell chip and is now in mass production, preparing for Blackwell Ultra. Market demand for Blackwell is exceptionally high, fueled by accelerated inference needs and large-scale post-training and model customization, necessitating greater computational acceleration.

AI Data Center Investment

Meta plans to invest over $200 billion in constructing next-generation AI data center parks.

Scaling Law and Architectural Advancements

The direction of Scaling Law is shifting towards Test-Time Scaling, with Nvidia's Hopper and Blackwell architectures poised to enhance model inference efficiency. The Blackwell series is particularly anticipated for its capabilities.

Impact of Export Controls

Affected by export controls, Nvidia introduced a special version of the H20 chip for the Chinese market, which has seen significant order growth.

No news to summarize.

AI Chip Market and Nvidia's Performance

Nvidia's Optimism Amidst AI Innovation

Nvidia CEO Jensen Huang expressed optimism about the company's future despite the emergence of DeepSeek's R1 model. He sees R1 as an "excellent innovation" driving demand for compute power in AI, benefiting Nvidia due to the high compute consumption of reasoning models.

Nvidia's Sales Surge

Nvidia's sales continue to surge, reporting record-breaking revenue of $39.3 billion. The company projects further growth in the next quarter, estimating revenue of around $43 billion. Data center sales nearly doubled in 2024 to $115 billion. Huang highlighted strong demand for Nvidia's Blackwell chip, custom-built for reasoning, and expects strong growth in 2025.

Robust AI Chip Market

The AI chip market remains robust, with major companies like Meta, Google, and Amazon investing heavily in AI infrastructure.

AI Industry

DeepSeek Technology Breakthrough

Shenwan Hongyuan Securities reports that the breakthrough of DeepSeek technology may mark an explosive growth for the AI industry similar to the mobile internet in 2010, akin to the "iPhone 4 moment." The report suggests that the AI industry is in an early "incubation" stage, and DeepSeek's technological advancements will accelerate the application and popularization of AI.

AI Inference Model Competition

The inference model war triggered by DeepSeek R1 is intensifying, with Alibaba, Anthropic, ByteDance, and others entering the fray. DeepSeek is accelerating the launch of the R2 model, and Shen Xiangyang points out that the industry focus has shifted to the Reasoner mode. Various manufacturers are vying to launch inference models, aiming to seize users. Alibaba released the inference model QwQ-Max-Preview, but it is only a preview version. Anthropic, on the other hand, released a complete model, Claude 3.7 Sonnet, with performance surpassing OpenAI and DeepSeek R1. Companies are eager to release products to get ahead of DeepSeek, leveraging multimodal differentiation for competitive advantage.

Artificial Intelligence

Nvidia's Situation

Nvidia is facing challenges.

Artificial Intelligence

Text Generation in Video

The technology to generate text within videos has emerged.

DeepSeek's DeepGEMM for FP8 GEMM Optimization

DeepSeek has open-sourced DeepGEMM, an FP8 GEMM library implemented with only 300 lines of code. It focuses on optimizing matrix multiplication, particularly on NVIDIA Hopper GPUs. DeepGEMM achieves speeds up to 1350+ FP8 TFLOPS on Hopper GPUs and up to 2.7x acceleration in small-batch matrix multiplication. It employs optimizations such as warp-specialized kernels, Tensor Memory Accelerator (TMA), and specialized PTX instructions to concurrently execute data movement, tensor core MMA instructions, and CUDA core enhancements, thus improving computational efficiency and throughput. The design is simple, avoiding complex templates to lower the learning curve, and it outperforms expert-tuned kernels across various matrix shapes. DeepGEMM is suitable for V3/R1 training and inference, with excellent performance in AI model real-time inference and batch data processing.