TRIPLE-1 Inc. (Headquarters: Hakata-ku, Fukuoka-shi, Fukuoka; Representative Director CEO: Takuya Yamaguchi; Henceforth: TRIPLE-1) has announced two new products: The ASIC “KAMIKAZE Ⅱ” for Bitcoin mining that uses a 7 nm process made by TSMC, and the AI processor “GOKU” for deep learning that uses the world’s leading 5 nm process.
In the “cryptocurrency mining market” that is driving the miniaturization of semiconductor processes in recent years, an extremely severe business model has been formed that directly links the power performance of semiconductors used for mining with mining proceeds. Starting in February 2017, TRIPLE-1 has been using core know-how involving the designability for cutting edge processes, the optimization of electrical consumption, and the improvement of the yield rate that have been fostered from the development of ASIC (application-specific integrated circuits) for cryptocurrency mining. Development has started on both “KAMIKAZE Ⅱ”, the latest high-performance ASIC for Bitcoin mining that uses a 7 nm process made by TSMC, and the AI processor “GOKU” for deep learning that uses the world’s leading 5 nm process.
— Next generation 7 nm Bitcoin mining ASIC “KAMIKAZE Ⅱ” – That design completed
By making use of the know-how fostered from the initial “KAMIKAZE” model, the design for the next generation model “KAMIKAZE Ⅱ” has been finalized. After tape outing to TSMC and the prototype has been inspected, we will start mass production from 2021.
Additionally, we have started to offer reference boards and reference designs for the first “KAMIKAZE” mining machine model, the system that can provide smoother installation support is ready.
< Feature > ASIC Boost is built-in, improving the mining efficiency rapidly
By loading the custom algorithm “ASIC Boost” that improves mining efficiency (it was not loaded on to the first “KAMIKAZE” model), a more highly efficient mining environment is provided. Additionally, by supporting ultra-low voltage drive (supplied voltage 0.2x V), we anticipate a drastic reduction in electrical consumption.
KAMIKAZE / KAMIKAZE Ⅱ Performance specifications comparison
|Process||TSMC 7 nm||TSMC 7 nm improved process|
|Die size||3.5 × 3.9 mm||3.0 × 9.0 mm|
|Package||7 × 7 mm FC-BGA||7 × 12 mm FC-LGA|
|Input voltage||0.3 V or less||Drives at a lower voltage than KAMIKAZE|
|Power consumption||1.2 W / chip||6.7 W / chip|
|Hash rate||23 GHs / chip||240 GHs / chip|
|Electrical efficiency||52 W / THs||28 W / THs|
*KAMIKAZE: Chip actual value, KAMIKAZEⅡ: Design target
— New challenges, enter the AI market
In recent years, various companies have entered the AI chip market, including not only giant IT companies but also fast-growing start-ups in China. As competition for edge AI chip development intensifies, oligopoly supply by several major companies continues in the deep learning AI processor market. Currently there are no mass-produced chips smaller than 12 nm process.
On the other hand, the electrical consumption of data centers is expected to excess 10% of the entire world’s electrical energy by 2030. TRIPLE-1 believe that AI processors for deep learning need not only high “computing power” to learn and process a huge amount of data, but also a severe viewpoint on “power consumption” from the viewpoint of energy conservation and cost reduction.
Therefore TRIPLE-1 has launched the project to develop “GOKU”, an ultra-low power consumption dedicated deep learning AI processor using the world’s most advanced 5nm process, utilizing our unique Low power / High performance technology and design know-how in the most advanced processes.
— “GOKU”, an AI processor for deep learning with the world’s leading 5 nm process
The first stage prototype has been completed, and we are inspecting the performance. Realizing the mass production in 2021, we will complete the prototypes closer to mass-produced products within 2020. It will be sold to the equipment and devices manufacturers for data servers or deep learning centers around the world, and we have already begun to discuss with multiple companies.
< Feature > Using the world’s leading 5 nm process, it has 10 times higher power efficiency than conventional
Our goal is to reduce power consumption by 1/10 of the equivalent performance compared to conventional products (process: 12 nm), and its feature is low-voltage operation that is inconceivable with existing products.
< Feature > Design to connect multiple calculation units
Like the structure of the human brain, where synapses connect neurons in a complex manner, great amount and very complex inter-core communication is important for deep learning AI processors.
Utilizing the unique circuit design technology of “GOKU”, which arranges many small arithmetic units on one large die as much as possible, it is designed to secure the communication (interconnect) band between the arithmetic units as much as possible.
It has been designed to get closer to the human brain.
< Reference > GOKU performance table (Design target)
|Power consumption||100 W|
|Peak performance (half-precision)||1 PFLOPS (1,000 TFLOPS)|
|Electrical efficiency (half-precision)||10 TFLOPS / W|
— TRIPLE-1 outline
Company Name: TRIPLE-1, Inc.
7F IT Bldg.Ⅱ,1-14-20 Hakataeki-Higashi, Hakata-ku Fukuoka-shi, Fukuoka, Japan
Representatives: Representative Director CEO Takuya Yamaguchi
The capital: 3662 million 895 thousand and 398 yen (including the capital reserve)