Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Tencent + likely Alibaba & ByteDance have placed major H20 orders to secure enough compute for DeepSeek and likely before AI chip ban comes into place.
Apparently, Tencent & ByteDance were the 2nd and 3rd largest Nvidia customers last year by number of chips (after Microsoft), but I guess they were getting the less power H20 chips, so the total compute power would be ranked lower
Sesame's open source release was a bust. The company only released a 1B model (their original on the website was 8B) and not the rest of the pipeline code for creating their demo, either. The community is not receiving it well, so this is a great opportunity for a Chinese company to come in & leap frog on market share.
More likely, CIA had a talk with them
Also means they are getting desperate.... its reeks of desperation to openly ban Open AI, for a company called OpenAI, will saying that IP theft is okay because the ends justify the means, all to beat China at all costs
This is how Amerikkka will lose it all
Tsinghua team open-sources large model inference engine "Chitu" to help domestic chips break through the cost and efficiency problems of FP8 model deployment and DeepSeek deployment
Professor Zhai Jidong's team from the Institute of High Performance Computing at Tsinghua University and Tsinghua-affiliated science and technology company Qingcheng Jizhi jointly announced the open source of the large model inference engine "Chitu", which is the first engine to natively run FP8 precision models on non-NVIDIA Hopper architecture GPUs and various domestic chips, bringing new breakthroughs to the widespread application and ecological construction of domestic AI chips.
Breaking the "hardware binding" dilemma, FP8 model deployment is no longer restricted
The development of DeepSeek has promoted the FP8 precision model to become the mainstream in the industry. With the continued popularity of DeepSeek, the demand for private deployment of large models in enterprises has also shown a blowout trend.
However, the current world-leading FP8 model has long relied on NVIDIA's H-series high-end GPUs, which has limited domestic companies in deploying large models due to the limitations of AI chips. On the one hand, the import of NVIDIA's H-series chips is restricted, making it difficult for domestic companies to obtain high-performance hardware support; on the other hand, most domestic chips do not support the FP8 data type and cannot fully utilize the performance of the new generation of AI models, making enterprise deployment costs high.
To break this dilemma, Tsinghua University and Tsinghua Unigroup jointly launched the open-source "Chitu" inference engine. Through the innovation of underlying technology, the engine has for the first time realized the efficient deployment of native FP8 models on non-H card devices (including GPU cards before NVIDIA's Hopper architecture and various domestic cards), getting rid of the dependence on specific hardware, and greatly reducing the threshold and cost of deploying AI models for enterprises.
Professor Zhai Jidong of Tsinghua University emphasized that Chitu condenses the team's years of accumulation of parallel computing and compilation optimization technology, and its goal is to "bridge the gap between advanced models and diversified hardware, so that domestic computing power can truly 'run' and provide key support for the implementation of China's large model industry." Qingcheng Jizhi CEO Tang Xiongchao said: "Chitu is positioned to become a bridge connecting diversified computing power and large model applications. We not only support NVIDIA's full range of GPUs, but also deeply optimize for domestic chips. In the future, we will gradually open source adaptation versions."
Professor Zhai Jidong's team from the Institute of High Performance Computing at Tsinghua University and Tsinghua-affiliated science and technology company Qingcheng Jizhi jointly announced the open source of the large model inference engine "Chitu", which is the first engine to natively run FP8 precision models on non-NVIDIA Hopper architecture GPUs and various domestic chips, bringing new breakthroughs to the widespread application and ecological construction of domestic AI chips.
Breaking the "hardware binding" dilemma, FP8 model deployment is no longer restricted
The development of DeepSeek has promoted the FP8 precision model to become the mainstream in the industry. With the continued popularity of DeepSeek, the demand for private deployment of large models in enterprises has also shown a blowout trend.
However, the current world-leading FP8 model has long relied on NVIDIA's H-series high-end GPUs, which has limited domestic companies in deploying large models due to the limitations of AI chips. On the one hand, the import of NVIDIA's H-series chips is restricted, making it difficult for domestic companies to obtain high-performance hardware support; on the other hand, most domestic chips do not support the FP8 data type and cannot fully utilize the performance of the new generation of AI models, making enterprise deployment costs high.
To break this dilemma, Tsinghua University and Tsinghua Unigroup jointly launched the open-source "Chitu" inference engine. Through the innovation of underlying technology, the engine has for the first time realized the efficient deployment of native FP8 models on non-H card devices (including GPU cards before NVIDIA's Hopper architecture and various domestic cards), getting rid of the dependence on specific hardware, and greatly reducing the threshold and cost of deploying AI models for enterprises.
Professor Zhai Jidong of Tsinghua University emphasized that Chitu condenses the team's years of accumulation of parallel computing and compilation optimization technology, and its goal is to "bridge the gap between advanced models and diversified hardware, so that domestic computing power can truly 'run' and provide key support for the implementation of China's large model industry." Qingcheng Jizhi CEO Tang Xiongchao said: "Chitu is positioned to become a bridge connecting diversified computing power and large model applications. We not only support NVIDIA's full range of GPUs, but also deeply optimize for domestic chips. In the future, we will gradually open source adaptation versions."
Alibaba revamped its quark app to incorporate latest Qwen advances.