Yasuhiro I. Email & Phone Number

Tokyo, Jp

広域東京エリア

Tokyo, Japan

Accelerating customer's AI workload by Tensix architecture + its scale-out environment.

日本 東京

As a customer-facing engineer, I have been contributing these items. - Technical consultation for integrating Graphcore's scale-out AI accelerators into the customer's IT infrastructure.- Direct support for converting customers' AI workload to Graphcore's IPU using POPLAR SDK. -Hosting public presentations and seminars to introduce our technology.For the first item: I am leading discussions with academic customers considering our AI accelerator clusters in their shared computer systems. Their major issue was how these new hardware components work with job scheduler or container orchestration systems. By using a deep understanding IPU's software stack, I could show prototypes of how multiple types of workload can be managed on such clusters. For the second item: I lead a project proofing concept of an Ultra-high bandwidth CV inference system for anomaly detection systems for local manufacturers. Not limited to porting work from existing GPU code from IPU-specific codes, proposing optimized model structure which exploits maximum efficiency of HW resource. For the third item: I have made tech-speech in AI exhibitions to inspire advantage of our Hardware products & their SDK, customer's use-case, etc.

Tokyo, Japan

- Designing and managing high-performance GPGPU cluster systems that are optimized for accelerating machine learning (DNN and SVM). For DNN learning environment, I worked on tuning of intra- and inter-node communication in order to accelerate data-parallel learning. For SVM learning environment I worked on both MPI performance tuning for accelerating SVM learning and CUDA kernel optimization for accelerating feature extraction.- Evaluating non-GPU SIMD/MIMD accelerators that can be used for optimizing machine learning. - Designing a data management system by combining DB (MongoDB, Kibana & elastic search) that helps machine learning engineers to manage huge amount of learning data, models, evaluation results.

Yokohama, Kanagawa, Japan

Virtual hardware in the loop (vHIL) (2008-2015): Prove a concept for simulation based software testing environment system for automotive product. By using CPU simulation tools provided by Synopsys, we modified micro-controller and mechanic system’s model tuned for software testing usages. This simulation systems showed high timing accuracy enough to check correct software behavior. Additionally, by adding customized model which can overwrite simulator’s internal state, I achieved functional fault-injection test that can evaluate software’s reliability. This technology was applied to software product development of automotive supplier in Japan.

Farmington Hills, Michigan

Application engineer

Developing Demo tools,tuning performance of tools