Veryken Frontier Industry News|AI, Semiconductor, Consumer Electronics, Power Battery, Energy Storage, New Energy Vehicles, Photovoltaics
Artificial Intelligence (AI)
[Google's new technology enables large models to handle "infinitely long" text inputs]
In its latest research, Google has proposed a mechanism called "Infini-attention", which greatly expands the ability of Transformer models to process long text sequences. With this innovation, Google's 1B parametric model can now handle contexts equivalent to the length of 10 novels, i.e., 1 million tokens, while excelling in Passkey retrieval tasks. Even more impressively, its 8B parametric model has set a new SOTA (state-of-the-art) record when processing book summarization tasks up to 500K tokens.
At the heart of the "infinite attention" mechanism is its compressed memory system, which allows the model to retain and reuse previous contextual information when processing new inputs. This mechanism stores and recalls information by a fixed number of parameters, rather than increasing the number of parameters as the length of the input sequence increases, significantly reducing the memory footprint and computational cost. This technology allows the model to process 114 times more information with the same memory size, which is equivalent to expanding a library of 100 books to 11,400 books.
In practice, "infinite attention" divides the input sequence into shorter subsequences, and uses a local attention mechanism to process the contextual information within each subsequence. This approach not only avoids the challenge of processing the entire long sequence at once, but also allows the model to process the input step-by-step in a streaming fashion. In addition, Infinite Attention differentiates two different types of attention heads, specialized heads and hybrid heads, which work together to process long-term and short-term contextual information, further improving the performance of the model.
VeryKen: This breakthrough is of great significance for the future development of large models, which not only improves the ability of models to process long texts, but also opens up new possibilities for multimodal data processing and the further development of Transformer models. With the further application and refinement of this technology, it is expected to see more innovative AI solutions in the future, especially when it comes to processing discontinuous data. (qubits, tiger sniff)
semiconductor
[Intel releases a new generation of AI chips, or it may become a strong competitor of NVIDIA products]
Intel has made significant progress in the field of AI accelerators, and its subsidiary Habana Labs showcased the latest generation of high-performance AI accelerator Gaudi 3 at the Vision 2024 conference, claiming that its performance will surpass that of Nvidia H100 and will ship to customers in the third quarter of this year. Gaudi 3 has significant performance improvements, especially FP8 with 1835 TFLOPS of computing throughput, which is expected to take a share of the AI computing field.
The design of the Gaudi 3 partially outperformed Nvidia's H100 in internal benchmarks and had an advantage in energy consumption. Gaudi 3 has demonstrated excellent performance when handling critical tasks such as large language models, and is expected to train large language models such as GPT-3 40% faster than the NVIDIA H100 and 50% faster than the NVIDIA H100 on the inference end. In addition, the release of Gaudi 3 is also a reflection of the change in the positioning of Intel's AI accelerator products, and the Gaudi series has been upgraded to Intel's flagship AI accelerator.
The Gaudi 3 is a direct evolution of the Gaudi 2 hardware, using TSMC's more advanced 5nm process and expanding on the computing hardware. The Habana team revealed the total throughput of the Gaudi 3 chip's FP8 precision, which makes Gaudi 3 twice as powerful as Gaudi 2 when using 8-bit floating-point calculations, and quadruple as powerful as BFloat 16.
Intel also highlighted the advantages of the Gaudi 3 in terms of energy efficiency, which is expected to be 100% faster than the H220 on the Llama model and 230% faster on the Falcon model. The dual-core configuration of the Gaudi 3 chip, connected by a high-bandwidth link, provides a unified memory address space, similar to NVIDIA's Blackwell.
VeryKen Smart Review: The release of Gaudi 3 comes at a time when demand for AI accelerator cards is high, and its strong performance and scalability make it a strong competitor to Nvidia H100, which may open up new opportunities for Intel in the AI accelerator market. Intel CEO Pat Gelsinger showed off the performance of the Gaudi 3 at the conference and said that its performance will be comparable to the Nvidia H200, or even better in some areas. The launch of Gaudi 3 is not only an important technology showcase for Intel in the AI field, but also part of its strategy to build a broad AI ecosystem alliance and promote enterprise AI innovation. (Heart of the Machine, Qubits, The Paper)
[Meta launches a new version of its self-developed AI chip, with 3 times higher performance than the previous generation]
Social giant Meta recently announced the latest progress of its next-generation self-developed AI chip, the new generation of AI training and pushing integrated accelerator MTIA v2. This chip is a family of custom chips designed by Meta specifically for AI training and inference work, with a 3x increase in computing performance, a 3.5x increase in memory bandwidth, and a doubling of memory capacity over the first-generation MTIA v1 launched last year.
The MTIA chip is designed to provide a more efficient ranking and recommendation system for Meta's social software. The chip uses TSMC's 5nm process and has more processing cores and larger on-chip memory at a frequency of 1.3GHz. Meta's tests show that the new chip far outperforms its predecessor in terms of performance for handling key models. To support the chip, Meta has also developed a large rack-mounted system capable of accommodating up to 72 accelerators, significantly improving operational efficiency. On the software side, the software stack of the new chip is very similar to MTIA v1, ensuring rapid deployment and compatibility. Meta has integrated the entire software stack into the chip, enabling developers to quickly take advantage of the new chip to process production models.
The new MTIA chips have been deployed in Meta's data centers and have shown positive results, proving that they are highly complementary to commercial GPUs in terms of delivering performance and efficiency. Meta's move not only reduces its dependence on external chip suppliers such as Nvidia, but also provides strong support for its expansion in the AI field. (Heart of the Machine, The Paper, Finance Associated Press)
[Led by Intel, giants are vying to lay out glass substrates, and advanced packaging will welcome change]
With the increasing demand for computing power and data processing speed in AI and high-performance computers, the semiconductor industry has entered the heterogeneous era, where multiple "chiplets" are widely used in packages. In this context, Intel introduced glass substrate technology to address the limitations of organic substrates in terms of signal transmission speed, power delivery optimization, design rule refinement, and package substrate stability.
As a new type of substrate material, glass substrate has excellent physical properties. Its main ingredient is silica, which is more stable at high temperatures and can handle higher temperatures more efficiently while effectively managing heat dissipation for high-performance chips. In addition, glass substrates enable higher interconnect densities, which are critical for power transmission and signal routing in next-generation packages, and have higher flatness, making packaging and lithography easier.
Intel, one of the leaders in the field of glass substrates, began investing in R&D a decade ago and plans to provide complete commercial application solutions in the next few years to drive the continued development of Moore's Law and meet the computing power needs of data-driven applications. Intel expects to use glass substrates for advanced packaging later this decade, with the first products being high-end HPC and AI chips. In addition, industry giants such as Samsung, LG, and Apple have also increased their layout in the field of glass substrates. Samsung set up a "legion" to increase research and development, LG Innotek began to prepare, and Apple actively explored the application of glass substrate technology to chip packaging.
The glass substrate market is promising, and according to MarketsandMarkets, the global glass substrate market is expected to grow from $7.1 billion in 2023 to $8.4 billion in 2028, at a CAGR of 3.5%. However, glass substrate technology is still in the early stage of development, and it also faces challenges such as immature technology and high cost. Glass substrates and organic substrates have different characteristics and advantages in their respective application fields, and in the future, they will not completely replace each other, but will play their respective advantages in their respective fields. Organic substrates still have a stable market for specific applications due to their cost-effectiveness and maturity, while glass substrates may play an advantage in areas such as high-performance computing. （cnBeta）
[Taking advantage of the AI east wind, TSMC's revenue growth rate in the first quarter hit a one-year high]
TSMC's recent financial report shows that its first-quarter revenue growth rate reached a new high in more than a year. According to statistics, TSMC's revenue in the first quarter increased by 16.5% year-on-year, or about 134 billion yuan, exceeding market expectations. In March, the monthly revenue increased by 34.3% year-on-year, equivalent to about 44.156 billion yuan. This strong performance reflects TSMC's strong demand in the high-end chip and server market, especially in the field of AI. TSMC's market capitalization has nearly tripled since its market trough in October 2022, and investors are confident in the advanced AI chips it produces, which are expected to offset the decline in global smartphone demand.
TSMC's customer base includes tech giants such as Apple and Nvidia, and their products are widely used in smartphones, tablets, high-performance computing, and artificial intelligence. Although the first half of the year is usually an off-season for business, this year's AI boom has kept TSMC's business demand hot. TSMC also said that its AI revenue is growing at a rate of 50% per year and is building new factories around the world to meet the demand for AI chips in data centers.
Although the market is optimistic about TSMC's long-term prospects, some investors have warned that the current demand for AI chips may be unsustainable. (Wall Street News, Finance Associated Press)
Power battery
[CATL's Q1 performance is outstanding, power battery and energy storage two-wheel drive]
CATL's financial report data for the first quarter of 2024 shows that its operating income in the first quarter was 79.8 billion yuan, a year-on-year decrease of 10.4%, but the company increased its gross profit margin to 26.42%, net profit margin to 14.03%, and net cash flow from operating activities increased by 35.25% year-on-year through effective cost control and the advantages of declining raw material prices.
CATL's battery capacity and market share increased in the first quarter, with the global installed capacity of power batteries reaching 35.5GWh, a year-on-year increase of 44.9%, and the market share increased to 38.4%. The market share of the domestic market also increased by 4 percentage points year-on-year to 48.93%. In addition, the company has also made progress in technology licensing, exploring cooperation with international automotive giants such as Ford, General Motors and Tesla, and providing new growth avenues for the company to explore overseas markets by providing battery manufacturing technology and services, and collecting royalties and service fees.
CATL's strong performance in the first quarter was mainly due to the following factors:
The new energy vehicle market continues to grow rapidly. In the first quarter of 2024, China's new energy vehicle sales reached 1.456 million units, a year-on-year increase of 141.1%. The rapid growth of the new energy vehicle market has driven the strong growth of power battery demand.
CATL's technological leadership in the field of power batteries. CATL has leading technical advantages in the field of power batteries, and its products have excellent performance and good cost control, and have strong competitiveness in the market.
CATL actively expands its energy storage business, providing new impetus for the company's performance growth. CATL has released a new Tianheng standard 20-foot containerized energy storage system, which has the characteristics of 4-dimensional true safety, 5-year zero attenuation, and 6MWh level, which increases the energy density per unit area by 30% and reduces the footprint of the whole station by 20%. The launch of this new product further strengthens CATL's leading position in the global energy storage market, especially in terms of long life and high safety performance. (Wall Street News)
energy storage
[Sunwoda liquid-cooled energy storage system is newly upgraded, and the revenue of energy storage business has skyrocketed]
Sunwoda continues to make efforts in the field of new energy energy storage, and its NoahX 2.0 liquid-cooled energy storage system was newly upgraded and unveiled at the 12th Energy Storage International Summit and Exhibition (ESIE 2024) held in Beijing. Sunwoda's new 5MWh liquid-cooled energy storage product has been put into use in a grid-side energy storage project in Jinhua City, Zhejiang Province, setting a record for the "most" in Zhejiang, which has been included in the 14th Five-Year Development Plan for New Energy in Zhejiang Province and is expected to provide strong support for the green development of the country.
Sunwoda's liquid-cooled energy storage system adopts advanced potassium iron phosphate battery technology, combined with high-efficiency PCS and intelligent management system, to achieve effective storage and release of power resources. The system has significantly improved energy density, cycle times and safety, and adopts the full life cycle safety boundary design to improve the safety performance of the system. In addition, Sunwoda has also launched a new product of Super industrial and commercial liquid-cooled energy storage system, further expanding its product and service range in the energy storage market.
Although Sunwoda's revenue fell for the first time since its listing last year, its net profit hit a new high. The company's revenue in the consumer battery and electric vehicle battery segments declined, while the energy storage system battery segment achieved revenue growth of 144%, although it accounted for less than 2% of total revenue. (Interface News)
New energy vehicles
[Tesla lays off 14,000 employees and promotes Robotaxi, with the ambition of subverting shared mobility with AI]
Tesla recently announced that it will lay off about 14,000 employees, or about 10% of its global workforce. The move is a strategic restructuring to optimize operational efficiency, reduce costs, and better adapt to market changes, and is expected to impact multiple divisions and business lines of the company globally. At the same time as the adjustment of human resources, Tesla is also making a strategic shift in focus, deciding to abandon the originally planned "Model 2" cheap model and turn to accelerate the Robotaxi (Robotaxis) project, which is expected to be released on August 8. In the face of fierce competition in the Chinese market and the rise of local brands, Tesla seems to be looking to maintain its market leadership through technological innovation. Robotaxi's business imagination is not only in the C-end hardware sales in the field of autonomous driving, but also in the software subscription revenue brought by FSD (full self-driving), behind which is the ambition to subvert shared mobility with AI, but Tesla must face the severe challenges of technology, cost and market competition. (Geek Park)
Solid-state batteries
[Solid-state battery concept stocks broke out in an all-round way, and more than 20 stocks rose sharply]
Solid-state battery technology has recently become a hot spot in the market again, driven by the news of the launch of 900V ultra-fast charging solid-state batteries by Zhiji Automobile, solid-state battery concept stocks have ushered in a full-scale outbreak, with more than 20 related stocks rising or rising by more than 10%. Overall, the commercialization of solid-state battery technology has achieved a key breakthrough from 0 to 1, and the market is optimistic about its prospects. The market performance shows that the medium-term upward trend of the solid-state battery sector has not been damaged by short-term differentiation, and the core high-standard stocks such as Sanxiang New Materials have achieved a return in the end of the market, laying the foundation for the follow-up capital return, and the lithium battery industry chain has also ushered in a collective rebound driven by solid-state batteries. (Finance Associated Press)
Photovoltaic industry
[LONGi Green Energy and GCL Technology lock in a large order of more than 20 billion yuan of granular silicon, and the polysilicon market may welcome changes]
LONGi Green Energy recently signed a high-profile procurement contract with GCL Technology, planning to purchase 425,000 tons of granular silicon from GCL Technology, with a valuation of about 22.1 billion yuan, becoming the first purchase order of more than 20 billion yuan in the photovoltaic industry this year. This cooperation not only reflects the competitiveness of granular silicon in the market, but also indicates that the polysilicon market may usher in new changes.
As a low-cost, high-quality differentiated product, granular silicon has gradually increased its market share in recent years, and is regarded as an innovative material with the potential to break the existing market pattern. As a leading enterprise in the production of granular silicon, GCL Technology's cooperation with LONGi Green Energy is expected to further increase the market penetration rate of granular silicon and promote its wide application in the photovoltaic industry.
LONGi's move is also related to the strategic layout of its new product, Terui silicon wafers. Terrig wafers use antimony-doped technology, which allows the use of a large proportion of granular silicon, thereby reducing costs and meeting the demand for high-efficiency cells. LONGi Green Energy has ensured the stable supply of granular silicon through the procurement method of locking quantity without locking price, and also shows the company's confidence in the market prospect of granular silicon.
Through this cooperation, GCL Technology has accelerated its transformation from the rod silicon market to the granular silicon market. The company said that the purchase contract will help promote stable sales of products, improve market rate, and be in line with the company's development strategy. GCL Technology plans to increase its granular silicon production capacity to 500,000 tons by the end of 2024 to meet the strong market demand.
VeryKen Smart Commentary: The cooperation between LONGi Green Energy and GCL Technology will not only help both parties improve their position in the photovoltaic industry chain, but may also lead to changes in the competitive landscape of the polysilicon market. The wide application and cost advantages of granular silicon may attract more enterprises to put into production, promoting technological innovation and market competition in the photovoltaic industry. With the continuous maturity of granular silicon technology and the improvement of market acceptance, the competition in the polysilicon market will become more fierce in the future, and cooperation and competition between enterprises will become an important driving force for the development of the industry. (Titanium Media)