Strategic Client Solutions Partner

[rifat28dddd]

6. NVIDIA launched its next-generation NVIDIA chip in January to commemorate mathematician NVIDIA. This chip has 100 million transistors and its features include: It is produced using TSMC's NVIDIA process. It is actually two chips connected together to ensure that they can run seamlessly as a whole. Through NVIDIA's high-bandwidth interface (-), it can be interconnected with bidirectional bandwidth to support higher NVIDIA cache bandwidth without memory locality problems and cache problems. The processing speed of AI-enabled models can be increased several times, including the training and reasoning stages. Nvidia said at the press conference that compared with previous chips, the super chip can provide a performance improvement of 100 times for large language model (LLM) reasoning loads and reduce costs and energy consumption by 100 times.


On January 2, 2017, Google Sundar Pichai announced that it poland whatsapp resource would launch a chip in cooperation with Nvidia in 2018. On January 2, 2017, Nvidia's Huang Renxun said that the chip has begun production. It will help push AI work beyond relatively simple tasks such as recognizing speech or creating images. Its emergence reflects NVIDIA's forward-looking judgment on future market demand and industry trends. For example, Moore's Law has made it increasingly difficult to improve performance, and chip iteration requires a combination of various technological innovations; data centers will be regarded as factories, and system-level performance, energy efficiency, and multiple combinations of "giant solutions" need to be considered; the scale of models and the amount of data continue to grow, and it is necessary to reduce computing-related costs and energy consumption; high-performance inference or generation is crucial, and methods must be found to parallelize model work on many nodes, etc.


NVIDIA has also built a supercomputer composed of Zhang, which uses copper cables to connect internal nodes to reduce power consumption. On the basis of a 10-fold improvement in training performance and a 10-fold improvement in inference capability, the cluster integrates multiple systems driven by NVIDIA into a liquid-cooled rack, providing unprecedented computing power for data centers, and can speed up the training of large language models by 10 times, and provide a 10-fold real-time speed improvement for inference of large language models with trillions of parameters. In addition, NVIDIA's world's first high-speed interconnect technology provides a key foundation for processing the largest visual computing workloads and unleashing the full potential of exaflops and trillion-parameter artificial intelligence models.