Chris LATTNER, the father of llvm: why should we rebuild AI infrastructure software

translate ｜ Shen Jiali 、 Jia Chuan

What people once imagined AI The vision is good , The current situation is unsatisfactory .AI On autopilot 、 The original prediction has not been realized in the daily application of new drug research and development , The common roast is , The global technology giants have gathered a large number of the smartest brains , But more focus on thinking about the precise delivery of advertising 、 Credit rating and not very intelligent “ intelligence ” On the speaker .

In theory, , As long as there are correct algorithms and sufficient computing resources ,AI It can solve all the problems represented by any available data , And now the data 、 Algorithm and hardware resources are rich enough ,AI All the conditions for benefiting the society have been met . We see that AI Its wide application and preliminary effect , But actually , The application of technology is not deep , It is far from realizing all the potential of existing machine learning research .

Why did this happen ？ In fact, compared with the world's technology giants and media AI The research update of . Compiler Daniel Chris Lattner It has been pointed out that ,AI The simplification and fragmentation of systems and tools is the root cause of this problem .

To solve this problem ,2022 year 1 month , Compiler Daniel Chris Lattner Announce to start a business , Same as Tim Davis Jointly established Modular AI, The goal is to rebuild the world ML infrastructure , Including compiler 、 Runtime , Heterogeneous computing 、 Edge to data center , And focus on usability , Improve the efficiency of developers . at present ,Modular AI The team has been involved in building from TensorFlow、TF Lite、XLA、TPU、Android ML、Apple ML、MLIR And most of the production machine learning infrastructure in the world , And has deployed production workloads to billions of users and devices .

In recent days, ,Modular AI Announced a complete 3 Ten million dollars seed round financing , from Google Venture Lead investment . stay Chris Lattner In the latest official blog post released by et al , Issued “ Soul three asked ”：AI So important , Why is software so bad ？ Why did the technology giant not solve AI problem ？ How to solve this problem ？ Of course , They also gave an answer .OneFlow The community compiled and sorted the original .

1

AI So important , Why is software so bad ？

AI Software was originally built for AI Full stack researchers of Technology 、 Designed by engineers and architects , It has never been defined as a product , therefore ,AI The software has defects in the underlying design .

This software is built by large technology companies to solve their own problems , And other enterprises are “ Drip type （trickle down） infrastructure ” Use these software on . Thus, such a phenomenon occurs ： Only the largest and most commercially influential AI Applications are built and deployed in practice , even so , It can only be achieved when the needs of enterprises are highly consistent with the internal needs of large technology companies .

Why is that ？ Because of the present AI The software is simple , Heavy research attribute , Mainly used to meet the needs of technology giants （ The developers of these software ） Development plan of . These software are created by researchers for research , and AI The rapid development of has given researchers no time to stop and rebuild .

contrary , as time goes on , We add more and more complexity , This makes it difficult for the industry to maintain and expand the fragmented customized tool chain , These tool chains are in research and production 、 Training and deployment 、 There are differences between the server and the edge .

Artificial intelligence system has now become a vast ocean of incompatible technologies , Only those integrated technology giants can use AI To achieve their goals .

2

Why did the technology giant not solve AI problem ？

AI Researchers and developers work together , To deploy AI It was a success , Technology giants use their huge computing and financial resources to promote the priorities of their products and core businesses , Including their own cloud 、 Telephone 、 Social networks and AI hardware .

Although they have made outstanding contributions to this field , But from a business perspective , They can't put AI Spread all over the world （ Cover all hardware 、 Yun He ML frame ）, And the rest of the world cannot expect them to do so . however , This unfortunate fact restricts the use of this technology in other countries in the world , Unable to solve problems outside the areas of concern of large technology companies , Including some of the most important socio-economic and environmental problems facing the world . But this is not the future we want .

Although giants have made great contributions to the development of artificial intelligence , But let AI give full play to its potential , We also need an independent company , This company doesn't have to prioritize its own hardware 、 Cloud infrastructure 、 The development of mobile phones or their own research ; At the same time, we need a neutral company , Do what best suits the interests of global users and enterprises . We need to integrate the knowledge learned from the rapid growth of AI software into the next generation of Technology , So as to provide available solutions and common standards for all kinds of problems faced by all organizations .

today , The most urgent problem faced by small and medium-sized technology companies is , How to break through ability 、 cost 、 Time and talent constraints will AI Put into production .

Considering the opportunity cost , Their innovative technology is difficult to market , Poor product experience , Will eventually have a negative impact on their development . For the whole society , This means that we still have to wait a long time to use AI To solve some major problems in the world .

We don't have time to wait for technology giants to launch trickle down AI Software .AI Can change the world , But the premise is that the fragmentation problem must be solved , And the world AI The developer community does not need to be plagued by high-quality infrastructure .

3

Who will solve this problem ？ How to solve ？

Modular Building the next generation AI Developer Platform , It will be more practical 、 High speed and flexibility .

Our platform unifies the popular AI The front end of the frame , It also strengthens the access and portability of various hardware backend and cloud environments . We are rebuilding workflow tools for core developers , Make it more expressive 、 Usability 、 Debuggability 、 reliability 、 Extensibility , Achieve superior performance . Our tools can be easily deployed to existing workflows , Users do not need to refactor or rewrite code , Then the work can be completed seamlessly , And achieve productivity and performance improvements at a lower cost . We will speed up the excavation AI value , And bring it to the market as soon as possible , Benefit the majority of users .

When AI When it can penetrate into various applications more subtly , Its potential will also be fully displayed —— When the , You won't have to surround AI To define your application . Our platform is modular 、 Built from composable infrastructure components , Support re collocation and extension to realize various use cases . meanwhile , Even without knowing how the whole system works , Experts in various fields can also innovate through our platform . We have seen firsthand how the modular approach unlocks new use cases , And this is something we never thought about before .

In order to really repair AI infrastructure , We have to solve “ Hard technology ” problem （ Such as compiler for heterogeneous computing technology ）, We should also establish an end-to-end developer workflow that can be seamlessly connected .

4

from “AI Research era ” enter “AI Production era ”

Our success means that developers around the world will be truly available 、 Portable and extensible AI Software .

In the new world , Developers who lack sufficient budgets or top talent can also work as efficiently as global technology giants ;AI Hardware efficiency and total cost of ownership (Total Cost of Ownership,TCO） Will be optimized ; Enterprises can easily insert customized ASIC To meet its use ; Deploying to the edge is as easy as deploying to the server ; Enterprises can use any one that best meets their needs AI frame ;AI The program can be seamlessly extended in hardware , Will be up to date AI It's easy to deploy research into production .

We will see ：AI The development of the industry is no longer limited by the timetable determined by the technology giants according to their own needs ;AI The development of the industry will be faster 、 More concentration ; Innovation thrives at all levels of the stack , Developers focus on bringing new innovations to market in their professional fields , And build a better future for all of us ; The industry is developing rapidly , Lead us from “AI Research era ” enter “AI Production era ”.

（ original text ：

https://www.modular.com/blog/the-case-for-a-next-generation-ai-developer-platform）

Everyone else is watching

• OneFlow v0.7.0 Release

• LLVM The father of ： The golden age of compilers

• The illustration OneFlow Learning rate adjustment strategy of

• LLVM The father of ： Modular design decisions AI future

• Reading Pathways： The next step forward is OneFlow

• Hinton： My 50 years of in-depth study career and Research on mental skills

• from MLPerf Talking about ： How to lead AI Accelerator the next wave

Click on “ Read the original ”, Welcome to download experience OneFlow v0.7.0