From [GeForce] reach artificial intelligence [AI]: why [NVIDIA] decision about [CUDA] change the company future?
Introduction: company fit leave today profit to build tomorrow computing world?
One company spend plenty money on capability wey customers no dey ask for immediately. To add am increase production cost, but customers no ready to pay extra. Still, the company no turn back. E no just wan sell new product; e wan build computing platform [Computing Platform] wey other people fit use create their own work for future.
Na this thinking dey behind [NVIDIA] decision to bring [CUDA] enter [GeForce] graphics processors [GPU]. Jensen Huang explain say the decision no bring immediate success. Instead, e bring costs, pressure on profit and long waiting. But later, e become one of the foundations for deep learning [Deep Learning] growth.
1. From accelerator for one narrow task reach broad computing platform [Computing Platform]
For beginning, [NVIDIA] na company wey dey make one particular task faster. That special capability give am strength. But to depend only on narrow market fit also limit research and development [Research and Development] capacity.
Two things dey pull different directions here. If design too specialised, general computing work fit hard. If company try do everything, e fit lose the original special capability. Huang approach na to move step by step between the two.
E explain say the first step na programmable pixel shader [Programmable Pixel Shader]. Next, IEEE-compatible 32-bit floating-point calculation [IEEE-compatible FP32] enter shaders [Shaders]. Researchers begin ask whether processor wey dem build for graphics fit do other calculations too.
Then [Cg], wey base on (C) programming language, come. That journey eventually reach [CUDA]. So no be magic solution wey appear for one day. Na continuous journey wey join one technical capability to the next.
2. Where dem carry [CUDA] go matter pass only creating am
To create excellent technology no mean say people go automatically use am. E need people wey dey write programs [Developers]. E also need plenty people wey fit use those programs.
Computing platform [Computing Platform] strength no dey only for fine design; e dey for how many people already get am too. Huang explain this as the importance of installed base [Installed Base].
E use [x86] architecture as example: despite plenty criticism, e survive because plenty people dey use am. E also say some reduced instruction set architectures [RISC Architectures], even with elegant design, no get the success wey dem expect. This no mean say every [RISC] design fail. The point na say fine technical design alone no guarantee platform success.
That time, [GeForce] already dey sell millions of units every year. So [NVIDIA] get ready route to carry new computing capability reach people.
3. [CUDA] for every [GeForce] processor: foundation before users ask for am
Dem fit keep [CUDA] capability only for few expensive research machines. If dem do that, fewer people go fit access am.
Instead, the company decide to use [GeForce] bring the capability enter personal computers [PC] wey plenty people fit get. Product wey person buy for gaming also give research opportunity. Student, university laboratory or researcher wey assemble computer by imself fit try new calculations with equipment wey already dey available.
This no mean say [CUDA] dey run for every computer. The point na say dem use [GeForce] installed base as starting foundation to make the capability accessible to more people.
Huang say the work no stop for supplying equipment. E include teaching for universities, writing books and running classes. To build platform no end with selling product; e need knowledge and community wey fit use am too.
4. The immediate price: higher costs and pressure on profit
The economic burden no small. Huang remember say to add [CUDA] capability increase processor cost by about 50%, while the company dey operate with gross margin [Gross Margin] of about 35% that time.
Gamers no go pay extra for computing capability wey dem no dey use. So the company need carry the cost. E describe am as big enough to swallow the company whole gross profit dollars [Gross Profit Dollars].
We no fit just subtract these two percentages to judge the company financial position. One base on production cost; the other na profit margin wey dem calculate from revenue. Im point no be simple arithmetic, but the hard position of sacrificing today earnings for tomorrow platform.
E also give rough figures from memory: the company market capitalisation [Market Capitalisation] na about 6 to 8 billion US dollars that time, before e later fall to about 1.5 billion US dollars. These na approximate figures wey e remember. Dem no prove say na only [CUDA] cause the whole change.
Still, im point say the expected benefit take about ten years to arrive matter well. Dem no make this decision for next quarter earnings.
5. Why this decision become foundation for deep learning [Deep Learning]?
One serious barrier to new research na the cost and trouble of getting equipment. If new experiment fit start on equipment wey plenty people already get, part of that barrier reduce.
[CUDA], wey spread through [GeForce], lay foundation for researchers to experiment inside common programming environment [Programming Environment]. Knowledge, programs and experience wey dem build there fit help later work too.
When deep learning [Deep Learning] begin grow, equipment wey fit handle the large-scale parallel computing [Parallel Computing], and people wey sabi use am, already dey available. Na important advantage be that.
The whole history of deep learning [Deep Learning] no depend on one company or one tool alone. But this decision get important place for making the computing platform widely available. Na why Huang emphasise say [NVIDIA] na the house wey [GeForce] build.
6. Trust wey dem build before the technical decision
Future fit clear for leader mind. But that clarity alone no enough if board, executives, workers, customers and partners no accept am.
Huang describe im approach as explaining cause and effect small small. Anytime new information, research progress or engineering achievement show, e explain to people around am how e connect to the future direction.
For im view, the day wey dem announce big decision no suppose shock everybody. Dem suppose don prepare the foundation so well that people go instead wonder why e take so long.
This kind long-term trust and patience also dey behind the [CUDA] decision. Apart from choosing technical path, dem need prepare the company wey go bear the cost for the journey.
7. No be product alone, but platform wey partners fit build upon
The way [NVIDIA] see itself as computing platform company [Platform Company] matter too. The main idea na to leave room for other companies to contribute and build together for every layer.
This no mean say all im programs na open source [Open Source]. To give partners chance to build together different from making source code public.
Good platform no grow only because of the company wey create am. E also grow through applications wey researchers, programmers, companies and schools build on top am. The [CUDA] decision open door for that growth.
8. Today lesson: to make one processor faster no enough
For large distributed workloads [Distributed Workloads], to speed up only one part no make the whole system faster by the same amount. Amdahl law [Amdahl's Law] explain this.
Make we assume say na only half of the time for one task we fit improve. No matter how fast we make that half, the unchanged half still dey. Even if the improved part no take any time again, the speed of the whole task fit only double at most.
Na why graphics processor [GPU], central processor [CPU], networking [Networking], memory [Memory], power supply [Power] and cooling [Cooling] need design wey make dem work together. Huang describe this as extreme co-design [Extreme Co-design].
We fit see connection with the early [CUDA] decision here. One tool capability alone no enough. The whole environment wey dey use am must grow along with am.
Conclusion: the decision wey create future use
The importance of introducing [CUDA] no stop for new programming capability. The deeper matter na say dem join am with [GeForce], wey already dey sell in large numbers, and put am early for the hands of people wey go use am for future.
If person look only at the costs of that time, the decision fit look wrong. But [NVIDIA] understand say computing platform value also dey inside applications wey other people go build on top am.
When technology, widespread use, education, developer community and company patience join together, new opportunity come. That opportunity later connect with artificial intelligence [AI] growth. Na this lesson the [CUDA] decision teach us: tomorrow big change need foundation wey plenty people fit start build on today.
Powered By K7