AI Chip Company, Announced Its Dissolution

It has been revealed that AI semiconductor startup LeapMind will be disbanded on July 31, 2024.

0021-09760 Gas Box Dxz

0040-09095 Gas Box, Wcvd

Soichi Matsuda, the company's director and chief executive, said in an email to interested parties, "We believe that in order to actually use AI, we need to consider both software and hardware, and besides, there are very few companies like that." "There are such ideas all over the world, so we are constantly challenged to think that they are valuable, but we are very disappointed that we have not yet been able to prove their worth", explains that they decided to do so. Voluntarily dissolve a company while it still has cash and deposits to prevent the risk of default.

From August, the Company plans to commence normal liquidation proceedings, at which time Mr. Matsuda will act as the representative liquidator.

LeapMind begins development of new AI chips

According to the company's official website, society is constantly changing as technology advances. Rapid innovation in equipment and the proliferation of infrastructure have made it possible to collect and utilize large amounts of data. With the practical application of machine learning, the accuracy of analysis has increased, and the use of data has become more familiar. The flow of data cycles, i.e., better data generated by devices that become smarter through machine learning, will become faster than ever before, making people's lives more convenient.

LeapMind was one of the first companies to predict this future and has been running a machine learning-based business since 2012.

LeapMind said that the company's technical capabilities and vision have been highly praised by many companies and organizations, and it has participated in many projects that use machine learning, but unfortunately not many cases have been translated into social implementation. There are two problems. The first is to build practical machine learning models. In order to leverage machine learning to solve hitherto unsolvable tasks, we must develop high-quality machine learning models. The second is a computing environment that allows machine learning models to actually run. There is no clear device that allows it to actually work with limited use, such as at the edge.

In order to improve society through the use of machine learning devices, LeapMind believes that companies must overcome two challenges. By continuing to face customers and problems in good faith with colleagues with different skill sets, we were able to solve problems in two ways: "developing high-quality machine learning models" and "developing high-speed, efficient hardware IP" "I came up with the answer.

By working on both the software and hardware sides, we can make the impossible possible. This future is within reach. We believe that by providing the world with the key technologies of the future, we can create a more humane way of life.

Based on these backgrounds and considerations, LeapMind announced in October last year that it would develop new AI chips to accelerate the computational processing of AI models and pursue industry-leading cost performance.

They say that due to the recent increase in the size and computational complexity of AI models, including large-scale language models (LLMs), the cost of training cutting-edge AI models has increased significantly compared to 10 years ago. This rising cost is a major bottleneck in AI development.

In order to create a good AI model, a large number of processors are required for parallel computing. Providing a large number of processors requires a large budget. If you can use a cost-effective processor, you can develop better AI models even with the same budget. In other words, the processor characteristics required for AI learning are shifting from absolute performance to price-performance.

In light of these circumstances, LeapMind began to develop new processor semiconductors (hereinafter referred to as "AI chips") for AI learning and inference, applying the technology we have accumulated in the development of edge AI accelerators. The new AI chip focuses on AI model learning and inference, with a compute performance target of 2 PFLOPS (petaflops) and 10 times the price/performance of GPUs of comparable performance. The product is expected to begin shipping by 2025 at the latest.

 

According to reports, the new AI chip has the following characteristics: Designed for AI model learning and reasoning · Emphasis on low-bit expressions such as FP8 Open Source Drivers and Compilers They say that when AI model learning and inference are considered computational tasks, it has the following characteristics: · Matrix multiplication is a computational bottleneck, easy to parallelize, and has few conditional branches.

LeapMind emphasizes that the company does not aim to improve the performance of general-purpose computers, but uses the above features to be specifically designed for AI model learning and reasoning. For example, since there are few conditional branches in the program, the number of transistors can be reduced by omitting the branch prediction unit.

The reason for emphasizing low-level expressions such as fp8 is that, in their view, the computational bottleneck of AI models is matrix multiplication, which involves a lot of multiplication and addition. Multipliers tend to be large circuits, but by using a data type with a lower bit width than before, such as FP8, you can reduce the number of transistors required. In addition, because the data processed is small, it is possible to make effective use of DRAM bandwidth, which has become a bottleneck in recent years.

As for open-source drivers and compilers, this is because developing AI models requires an advanced software stack that cannot be provided by a single company. There is already an open source software ecosystem involving multiple companies, and in order to be a part of this ecosystem, it is important to join the community as open source software.

Under the LeapMind program, the company will expose hardware specifications as much as possible and release software such as drivers and compilers under an OSI-compliant license.