Nominated: Cloud 7nm GPGPU Chip Product BI

Big Island (BI) features high performance, versatility, and flexibility. It supports mainstream GPGPU ecosystems in China and overseas and various mainstream frameworks for deep learning. Through standardized software and hardware interfaces, it helps users address the actual pain points, such as challenges in product use and high migration costs for development platforms. It enables customers to optimize the computing system continuously within its life cycle, improve computing capabilities, and reduce operational costs. Meanwhile, in response to the application environment in China, Iluvatar CoreX works with key partners to define and locally optimize the design, laying a solid foundation for large-scale commercial use in the local market in the future. With its outstanding performance, energy efficiency, and cost-effectiveness, Iluvatar CoreX's BI chips and chip cards can provide computing support for various high-load tasks such as AI training and reasoning, cognitive AI, high-performance data analysis, genome research, and predictive financial analysis after its mass production. It will serve a myriad of industries, including education, internet, finance, autonomous driving, healthcare, and security, thus empowering an AI-driven intelligent society.BI has the following key characteristics:

1. First high-end cloud training GPGPU chip developed in China

The General-Purpose Graphics Processing Unit (GPGPU) is a chip with an architecture similar to a GPU. It can wrap up general-purpose computing tasks previously handled by the central processing unit (CPU). GPGPU has powerful parallel processing capabilities and programmable pipelines. Especially when addressing the needs of single instruction, multiple data stream (SIMD), GPGPU has far better performance than the conventional CPU, as it plays a key role in many fields. The BI chip is the first GPGPU developed in China. It possesses proprietary core intellectual property rights, and leading-edge 7nm process nodes, and 2.5D chip-on-wafer-on-substrate (CoWoS) packaging, enabling it to have outstanding performance. It doesn’t have a dedicated graphics rendering module, leading to better programmability. It highlights general computing functions while optimizing AI training and inferencing.

2. Proprietary computing unit and instruction set

To support general computing, BI has a complete and diverse set of GPGPU instructions. The set of instructions, 64-bit length, supports various operations on vectors, scalars, and tensors. These operations include mathematical operations, logical operations, data reading and writing, and control, with rich functions compatible with mainstream ecosystems.

BI is the first high-end cloud training GPGPU chip developed in China based on the general GPU architecture. It adopts the industry-leading 7nm node and 2.5D CoWoS packaging technology. It features 24 billion transistors and supports mixed-precision training, including FP32, FP/BF16, and INT32/16/8. It is integrated with a 32GB HBM2 memory, with storage bandwidth up to 1.2TB. The single-core has the capability of crunching 147 TFLOPS at FP16 (147TFLOPS@FP16).

3. CUDA ecosystem compatibility

The indigenous CUDA compatible software layer and back-end compiler, combined with the driver program adapted to the BI chip, provides a comprehensive software and hardware solution consistent with the CUDA ecosystem from the API level. The existing CUDA-based applications only require the recompilation of the toolchain of BI chips for them to run on BI GPGPU, thus eliminating unnecessary secondary development.