Semifive Begins Mass Production of AI Accelerator on Samsung's 4nm Process
TECHWORLD ·
✦ AI Summary
Semifive has begun mass production of Bertha, a data center AI inference accelerator based on Samsung Electronics Foundry's 4nm process.
This mass production is Semifive's first large-scale production case using Samsung Foundry's leading-edge 4nm process, and the company has now signed the initial production contract.
Semifive expects additional PO orders and expanding mass-production volumes as HyperAccel expands its services.
Semifive has begun mass production of a data center AI inference accelerator based on Samsung Electronics Foundry's 4nm process. On the 8th, Semifive announced the start of mass production for HyperAccel's data center AI inference accelerator, Bertha.
This mass production is Semifive's first large-scale production case using Samsung Foundry's leading-edge 4nm process. It has now signed the initial production contract.
As HyperAccel expands its services, Semifive expects to win additional PO orders. Accordingly, the volume of mass production is also expected to continue increasing in the future.
Bertha is an accelerator specialized in AI inference workloads such as LLMs. The product was implemented as a large-area Big Die chip of 500 mm² or more.
As semiconductor dies become larger, the number of chips that can be produced per wafer decreases and the likelihood of defects increases. Yield management also becomes more difficult as semiconductor dies grow larger. In addition, chips for high-performance AI computing require power delivery and thermal management, further increasing the complexity of design and manufacturing.
The significance of this project lies in the fact that Semifive handled the entire process from design to mass-production supply, connecting it to actual large-scale production. For this project, Semifive provided a turnkey solution that included front-end design and verification, packaging, testing, and mass-production supply. It also linked leading-edge process and large-area AI semiconductors from the design stage to mass production. This case is described as proving Semifive's technological capabilities.
This mass production also marks a case in which Semifive's business structure is shifting from NRE-centered operations to recurring mass-production revenue generation. Through this project, Semifive showed a transition from its existing NRE-heavy business structure to a recurring mass-production revenue model.
Semifive began mass production of AI ASIC Wisenet9 for Hanwha Vision security cameras in the third quarter of last year. It then mass-produced AI semiconductors for HPC in Japan in the second quarter of this year. It also added a mass-produced product lineup for data center AI inference accelerators in the third quarter of this year.
Through this trend, Semifive has expanded its applications from security to HPC and data center AI. It has also secured a mass-production lineup linked to customers' product roadmaps. This expansion and lineup buildup are being analyzed as having improved business continuity and earnings stability.
As Semifive's mass-production projects expanded, its order performance was also reflected. New mass-production orders in the first half of this year reached KRW 42.3 billion, about twice last year's annual order total of KRW 21.2 billion. By quarter, orders came to KRW 15.6 billion in the first quarter and KRW 26.7 billion in the second quarter, with second-quarter orders up 71% from the first quarter.
In particular, overseas orders accounted for 45% in the second quarter. As a result, the company appears to have entered a phase of full-scale expansion of its global customer base.
Semifive said it has established a structure in which development project orders are linked to actual mass-production revenue. In addition, OEM volumes secured with end demand in place were reflected, and the addition of high-value-added ASIC projects for data centers improved the quality of orders.
Semifive CEO Jo Myung-hyun said that AI is shifting its center of gravity from training to inference and that demand for ASIC accelerators in the data center market is surging. He also said the success of this mass production was the result of combining the customer's innovative architecture with Semifive's turnkey capabilities. He added that the company successfully carried a large-area chip of more than 500 mm² from design to mass production and described the achievement as a milestone.
The speaker said the company plans to continue pushing the mass-production growth momentum demonstrated in the first half into the second half as well.
In the AI data center market, GPU-centered computing infrastructure is expanding, and demand for ASICs optimized for specific workloads is also growing. Against this backdrop, attention is focusing on the possibility that Semifive's experience in mass-producing large-area AI accelerators based on leading-edge processes could strengthen its competitiveness in the global ASIC market.
Source: TECHWORLD · Park Gyu-chan
Original: https://www.epnc.co.kr/news/articleView.html?idxno=406674
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.