Subscribe to the Byrne-Wheeler Report on YouTube for more content like this.

SEMIFIVE has begun mass production of Bertha, a data center AI inference accelerator developed for fellow South Korean company HyperAccel, the custom semiconductor firm announced on September 8, 2026....

SEMIFIVE has begun mass production of Bertha, a data center AI inference accelerator developed for fellow South Korean company HyperAccel, the custom semiconductor firm announced on September 8, 2026. The chip is manufactured on Samsung Foundry’s 4nm process node and marks SEMIFIVE’s first large-scale mass production project on that advanced node. Following the initial production contract with HyperAccel, mass production volumes are projected to expand steadily through follow-on purchase orders aligned with HyperAccel’s service rollout and expansion, the company said. A Big Die on Samsung’s 4Nm Node Bertha is an accelerator optimized for AI inference workloads, including large language models. It is a “Big Die” measuring more than 500 mm². SEMIFIVE said that as die size increases, managing power consumption, heat dissipation, and manufacturing yield becomes far more challenging, and that it brought the advanced-node project into mass production by delivering a complete turnkey solution spanning front-end design and verification through packaging, testing, and volume manufacturing supply. The design work behind the chip predates this week’s production start. At the Samsung Advanced Foundry Ecosystem Forum 2026 in San Jose, SEMIFIVE said it had successfully executed the comprehensive design and verification process for the Bertha Big Die, exceeding 500 mm², based...

SEMI FIVE has kicked off the mass production of HyperAccel's AI inference accelerator, 'Bertha,' utilizing Samsung's cutting-edge 4nm process technology, signaling immense growth.

SEMIFIVE announced today the mass production of a data center AI inference accelerator for HyperAccel, marking its first large-scale project utilizing Samsung Foundry’s 4nm process node. Following an initial production contract, mass production volumes are projected to expand with follow-on purchase orders aligned with HyperAccel’s rollout.

To read the full post, click through. Once at http://xpu.news, you can also comment and submit your own posts to appear in this feed.

HyperAccel is about to launch its Bertha 500 chip, an LLM inference accelerator designed for economical token generation in the data center.

**Abstract:** This paper introduces a novel framework for automated design verification and optimization of Micro-Electro-Mechanical System (MEMS) accelerometers leveraging a hyper-dimensional evaluation pipeline and reinforcement learning (RL). Current accelerometer design relies heavily on iterative finite element analysis (FEA) simulations and manual optimization, which is time-consuming and computationally expensive. We propose a system, "HyperAccel," that integrates numerical simulation, logical consistency verification, original concept detection, and impact forecasting to provide a hyper-score representing the potential of a given accelerometer design.

Южнокорейският стартъп HyperAccel, според EETimes, се готви да пусне Bertha 500 — специализиран чип, предназначен за ИИ-инференс. Благодарение на уникалната си архитектура, се твърди, че чипът генерира пет пъти повече токени в секунда от решенията, базирани на графични процесори, при същото ниво на TOPS. Bertha 500 поставя силен акцент върху ефективността на разходите. За тази цел той използва LPDDR памет вместо скъпата HBM. Чрез елиминиране на традиционната йерархия на паметта, той постига 90% използване на честотната лента на LPDDR. Допълнителни подобрения в ефективността се постигат чрез оптимизиране на архитектурата специално за задачите за инференс.

Корейски изследователи от Училището по компютърни науки KAIST, в сътрудничество с HyperAccel Inc, са разработили нов невронен процесор, който подобрява производителността на генеративните модели с изкуствен интелект средно с 60%. Съвременните модели на генеративен изкуствен интелект, като ChatGPT-4 и Gemini 2.5, изискват не само висока пропускателна способност на паметта, но и значително повече памет. В резултат на това компаниите, които използват облачен генеративен изкуствен интелект, включително Microsoft и Google, купуват стотици хиляди графични процесори от NVIDIA. Нов невронен процесор, създаден от корейски изследователи, подобрява производителността на генеративните модели с изкуствен интелект с 60% и намалява консумацията на енергия с 44%.