高通采用台湾iSTART-TEK的EDA工具以提升AI驱动的良率
高通已采用台湾iSTART-TEK的电子设计自动化(EDA)工具,用于其Dragonwing系统级芯片(SoC)产品,该产品主要用于终端侧人工智能。
原文深度解读
本文报道高通(Qualcomm)已采用台湾iSTART-TEK公司的START™ v5 EDA工具,用于其面向终端侧人工智能的Dragonwing系统级芯片(SoC)产品,旨在通过AI技术提升半导体良率。文章涉及高通、iSTART-TEK、Cadence、Synopsys等主体,时间上提及9月30日行业消息、6月签署授权协议、此前10个月测试验证、合同期限超过10年。关键数字包括电路面积减少约50%、合同超过10年、测试验证10个月。与电子元器件供应链的关系在于:EDA工具用于SRAM的BIST/BISR电路设计,可提升量产良率、降低DPPM和制造成本,并可能扩展至Dragonwing之外的产品线,从而影响芯片制造端的良率与成本结构。
- 高通于6月与iSTART-TEK签署START™ v5 EDA工具的授权协议,此前经过10个月的多领域测试与验证,合同期限据信超过10年,这是首个确认的高通采用台湾公司EDA工具的案例。
- 行业消息人士称,高通选择START™ v5的原因是该工具内置AI算法功能及其专利的用户自定义算法(UDA)技术,并且相比竞争方案可将部分电路面积减少一半。
- 该EDA工具被高通内部WLAN、IoT和AI团队使用,应用于Dragonwing芯片;Dragonwing是面向工业、嵌入式IoT、边缘AI、智能网络和蜂窝基础设施的计算平台,高通近期推出了面向边缘AI的Dragonwing Q-2390和IQ-2390处理器,成本和尺寸均降低。
- Dragonwing芯片包含用于高速临时缓存存储的SRAM,START™ v5用于设计芯片内部的内建自测试(BIST)和内建自修复(BISR)电路,以识别和修复存储器缺陷;BIST自动检查芯片内嵌存储单元是否有缺陷,BISR自动修复测试发现的缺陷,有助于提高量产良率并降低DPPM。
- 该EDA工具在自动识别芯片设计中的存储器类型后,其AI引擎会推荐可最小化电路面积并优化性能的电路架构,电路面积相比竞争方案可减少约50%,尤其适合存储器密度增加时的高性能低功耗芯片设计,生成的电路会自动连接到芯片更高层架构;iSTART-TEK将START™ v5描述为支持硬修复和软修复功能以及UDA的存储器测试与修复电路开发环境。
- 即使半导体晶圆前端工艺已完成,测试方法仍可通过软件重新配置,这得益于UDA专利技术;此前一旦半导体电路形成,芯片内嵌的BIST即固定,无法更改测试方法。采用START™ v5预计将提高未来高通Dragonwing产品的量产良率并降低制造成本,鉴于长期授权协议,高通也可能将该工具的使用扩展到Dragonwing之外的产品系列。
供应链影响
- 若START™ v5确实能提升Dragonwing芯片的量产良率并降低DPPM,可能减少高通在晶圆制造和封装测试环节的废品损失,从而可能影响其芯片供应稳定性和单位成本。
- 电路面积减少约50%可能使同等晶圆面积下产出更多芯片,或使单芯片尺寸更小,进而可能影响高通对晶圆代工产能的需求结构,但具体影响取决于实际设计迁移范围和量产节奏。
- 合同期限超过10年且可能扩展至Dragonwing之外的产品系列,可能使iSTART-TEK在存储器测试与修复EDA细分领域的供应链地位增强,并可能影响其他EDA供应商在高通内部的份额分配。
- iSTART-TEK已向韩国设计公司及比亚迪供应解决方案,若其汽车半导体市场拓展持续,可能使其EDA工具在汽车芯片供应链中的采用范围扩大,但取决于车规验证进度和客户导入情况。
- 由于该工具支持晶圆前端工艺完成后通过软件重新配置测试方法,可能为芯片设计公司提供更灵活的测试调整能力,从而可能影响测试环节的返工和修复策略,但实际效果取决于UDA技术在不同工艺节点上的适用性。
系统解读,仅作为判断线索,不构成备货、出货、涨价或投资建议。
高通Dragonwing工业计算平台。(照片:高通)
高通已采用台湾iSTART-TEK公司的电子设计自动化(EDA)工具,用于其Dragonwing系统级芯片(SoC)产品,该产品主要用于终端侧人工智能。此举旨在通过基于AI的技术提高半导体良率。
据9月30日行业消息,高通于6月与iSTART-TEK签署了其START™ v5 EDA工具的授权协议。该交易此前经过了10个月的多领域测试与验证。合同期限据信超过10年。这是首个确认的高通采用台湾公司EDA工具的案例。
“高通选择START™ v5是因为该工具内置了AI算法功能,并且拥有专利的用户自定义算法(UDA)技术,”一位行业消息人士表示。“该公司还取得了相比竞争方案将部分电路面积减少一半的成果。”
该EDA工具正被高通内部的无线局域网(WLAN)、物联网(IoT)和AI团队使用。它正应用于Dragonwing芯片。Dragonwing是一个面向工业、嵌入式IoT、边缘AI、智能网络和蜂窝基础设施解决方案的计算平台。高通近期推出了面向边缘AI的Dragonwing Q-2390和IQ-2390处理器,成本和尺寸均有所降低。
Dragonwing芯片包含静态随机存取存储器(SRAM),用于高速临时或缓存存储。START™ v5用于设计芯片内部的内建自测试(BIST)和内建自修复(BISR)电路,以识别和修复存储器中的缺陷。
BIST自动检查嵌入半导体芯片中的多个存储单元是否有缺陷。BISR自动修复测试发现的缺陷,有助于提高量产良率并降低每百万机会缺陷数(DPPM)。
一旦该EDA工具自动识别出芯片设计中使用的存储器类型,其AI引擎就会推荐一种能够最小化电路面积同时提供优化性能的电路架构。相比竞争方案,电路面积可减少约50%。随着存储器密度增加,该技术特别适合高性能、低功耗芯片设计。生成的电路会自动连接到芯片的更高层架构。iSTART-TEK将START™ v5描述为支持硬修复和软修复功能以及UDA的存储器测试与修复电路开发环境。
即使半导体晶圆的前端工艺已经完成,测试方法仍可通过软件重新配置。这得益于UDA专利技术。此前,一旦半导体电路形成,芯片内嵌的BIST即固定,无法更改测试方法。
采用START™ v5预计将提高未来高通Dragonwing产品的量产良率,同时降低制造成本。鉴于长期授权协议,高通也可能将该工具的使用扩展到Dragonwing之外的产品系列。
一位韩国EDA行业消息人士表示,高通已经使用Cadence和Synopsys等主要EDA公司的解决方案,但这些公司工具中的AI功能作为昂贵选项提供。“从定价角度看,iSTART-TEK的工具必定具有显著优势,”该消息人士表示。
iSTART-TEK向韩国半导体设计公司eASIC LAND、UnoSilicon和SilverChips供应其解决方案。该公司还在拓展汽车半导体市场。它曾为中国比亚迪U8车型的量产供应解决方案。
Qualcomm’s Dragonwing industrial computing platform.(Photo: Qualcomm)
Qualcomm has adopted an electronic design automation (EDA) tool from Taiwan-based iSTART-TEK for its Dragonwing system-on-chip (SoC) products, which are primarily used for on-device artificial intelligence. The move is aimed at improving semiconductor yields through AI-based technology.
According to industry sources on September 30, Qualcomm signed a licensing agreement with iSTART-TEK in June for its START™ v5 EDA tool. The deal followed 10 months of testing and validation across multiple areas. The contract is understood to run for more than 10 years. It is the first confirmed case of Qualcomm adopting an EDA tool from a Taiwanese company.
“Qualcomm chose START™ v5 because AI algorithm functions are built into the tool and because of its patented User-Defined Algorithms (UDA) technology,” an industry source said. “The company also achieved results that reduced the area of some circuits by half compared with competing solutions.”
The EDA tool is being used by Qualcomm’s internal wireless local-area network (WLAN), Internet of Things (IoT) and AI teams. It is being applied to Dragonwing chips. Dragonwing is a computing platform for industrial, embedded IoT, edge AI, intelligent networking and cellular infrastructure solutions. Qualcomm recently introduced the Dragonwing Q-2390 and IQ-2390 processors for edge AI, with reduced cost and size.
Dragonwing chips contain static random-access memory (SRAM), which is used for high-speed temporary, or cache, memory. START™ v5 is used to design built-in self-test (BIST) and built-in self-repair (BISR) circuits inside the chip to identify and repair defects in the memory.
BIST automatically checks whether multiple memory cells embedded in a semiconductor chip have defects. BISR automatically repairs defects identified by the test, helping increase mass-production yields and reduce defects per million opportunities (DPPM).
Once the EDA tool automatically identifies the types of memory used in a chip design, its AI engine recommends a circuit architecture that can minimize circuit area while delivering optimized performance. Circuit area can be reduced by about 50% compared with competing solutions. The technology is particularly suited to high-performance, low-power chip designs as memory density increases. The resulting circuits are automatically connected to the chip’s higher-level architecture. iSTART-TEK describes START™ v5 as a memory testing and repair circuit development environment supporting hard- and soft-repair functions as well as UDA.
Even after front-end processing of a semiconductor wafer has been completed, the testing method can be reconfigured through software. This is enabled by the UDA patented technology. Previously, once semiconductor circuits had been formed, the BIST embedded in the chip was fixed, making it impossible to change the testing method.
The adoption of START™ v5 is expected to improve the mass-production yields of future Qualcomm Dragonwing products while lowering manufacturing costs. Given the long-term licensing agreement, Qualcomm is also likely to expand the tool’s use to product families beyond Dragonwing.
A South Korean EDA industry source said Qualcomm already uses solutions from major EDA companies such as Cadence and Synopsys, but AI functions in those companies’ tools are offered as expensive options. “From a pricing perspective, iSTART-TEK’s tool must have offered a significant advantage,” the source said.
iSTART-TEK supplies its solutions to South Korean semiconductor design companies including eASIC LAND, UnoSilicon and SilverChips. The company is also expanding its push into the automotive semiconductor market. It supplied solutions for mass production of BYD’s U8 vehicle in China.