Afterits investments in OpenAIto propel its AI ambitions to new heights, Apple has just unveiled a new partnership, and not the least. The firmjoins forces with the graphics processor giant(and now AI): NVIDIA, a company led by Jensen Huang who literallyexploded on the stock market at the start of the school year. Results on the financial markets which can be explained precisely by its expertise in artificial intelligence.
With its high-performance GPUs, Huang's company has become THE ultimate provider for data centers, research labs and companies developing AI applications. Suffice to say that this new partnership with NVIDIAcould give Apple a nice boost, she who accusesa certain delay in this sectorin the face of competition.
An unprecedented hardware and software synergy
The technical collaboration between Apple and NVIDIA is based around the integration of ReDrafter (a new method of generating text using large language models developed by Apple) in the TensorRT-LLM environment. It is a Python library optimized by NVIDIA that allows you to run large language models (LLM)extremely fast on GPUs.
To enable this integration, NVIDIA has developed new specialized operators in its TensorRT-LLM framework. These software components ensure optimal communicationbetween Apple's algorithm and NVIDIA graphics processors. The adaptation focused on the efficient management of the data structures generated by ReDrafter, notably during two phases.
The first isbeam search, a heuristic method used in AI toefficiently explore a large search space, focusing on the most promising paths. The second isdynamic tree-based attention. This is a mechanism that allows an AI model to focus its attention on the most relevant parts of information, adapt this attention according to the context and the evolution of the latter and prioritize it.
The engineers implemented synchronization mechanisms allowing efficient distribution of calculations across the different GPU cores. This multi-level optimization takes full advantage of the tensor computing units of NVIDIA processors, while maintaining a modular architecturecapable of evolving with future developments.
Already tangible performances
This integration into the NVIDIA ecosystem is already bearing fruit. In the laboratory, tests carried out on massive neuronal architectures, equipped with tens of billions of parameters, reveal a spectacular increase in performance.The greedy algorithmrepresents the classic method of text generation by language models, where the algorithm selects at each step the token (text unit) with the highest probability.
The improvement brought by ReDrafter makes it possible to speed up this process considerably: where the traditional system generated a certain number of tokens per second, the integration of ReDrafterallows us to produce 2.7 times morein the same time interval.
Beyond the raw numbers, this technical advancement will transform the end-user experience. Application response time is significantly reduced, while the hardware and energy footprint is reduced. Apple researchers, aware of the practical issues,underline this double improvement: «Language models power more and more applications in production, and improving inference efficiency can impact computational costs while reducing latency for users».
No doubt, this partnershipwill give wings to Appleand will strengthen its position in the AI market. All we can do is hope thatApple Intelligence will one day benefit from this synergy.
- Apple and NVIDIA are teaming up to bring an advanced method of text generation to highly optimized GPUs.
- This technical collaboration significantly improves the speed and energy efficiency of AI models.
- The performance of this synergy will strengthen Apple's position in the race for artificial intelligence.
i-nfo.fr - Official iPhon.fr app
By : Keleops AG
