Globalization and the CPU wars
In order to get rid of some frequent prejudices I am proposing three heresies and two questions about the CPU wars. I am greatly indebted for the technical approach to Satoshi Matsuoka and the group of researchers from Riken Center for Computational Science, Kobe, Japan and 瑞士苏黎世联邦理工学院 , Switzerland who authored "Myths and Legends in High-Performance Computing", full publication here. Clearly, any errors in the following are mine and mine only.
This is the second part of a big issue about the technology wars, with special detail given to the CPU wars. You can find the first part here.
Heresy I - High performance computing (HPC) is strategic for the defence industry but has still little impact on everyday's life
An exascale computer is capable of processing at least a billion billion (10^18) 64-bit multiplications and/or additions per second. It is estimated that this is the same processing power of the human brain at the neural level. According to the Financial Times, the US has three exascale systems in the pipeline, while China aims to have up to 10 operational systems by 2025.
The race is still on: the US Department of Energy has put a tender for a class of supercomputers that will be up to 10 times faster than the recently inaugurated Frontier exascale system, which ranks the first in the world, and was also the first to reach the exaflops ceiling.
Among the first applications of HPC were nuclear physics and the simulation of nuclear reactors and nuclear weapons. Other real-world applications are about modelling:
They are also used by the scientific community for the processing of spacecraft data, urban systems planning and the production of graphics effects for the entertainment industry. Contrary to common thinking, HPC has little to do with Machine Learning: most notable exceptions are very specialized issues like astrophysics modelling.
Heresy II - Core computational performance is not a strategic asset in HPC
Technologists and their outreach managers are constanly on the look for communication opportunities. Computational performance data are easy to publish and there is no agreement in how to measure them, making them the perfect vehicle for airing political undertones. Usually, Chinese scientists tend to overestimate their results, while their US counterparts like to raise the Sputnik factor, declaring that China is on the brink of achieving technical supremacy.
Fact is, only gamers are interested in brute-force computing performances, which is easy to understand as gaming prizes are presently reaching 60$ million.? Processing workloads can be roughly divided into three classes: compute bound, memory bandwidth bound, and memory latency bound.
The hyperscalers and cloud providers just want more cores on their servers. As cores are their main invoice items, the more they are, the more revenue they can get out of a single server.
On the contrary, the HPC community has discovered memory bandwidth and memory latency to be far more critical, and is relying on parallel computing to optimize the performance-to-cost ratio. Among the hindering factors there are the established FORTRAN-based software packages, and the objective difficulty in refactoring/rewriting legacy code. The Climate Modelling Alliance (CLiMA), established at Caltech in 2018 has selected the statistical modeling language Julia to build a low-resolution digital twin of our planet.
Heresy III - Deep Learning supremacy in AI will become a thing of the past
Many of the architectures presently used in machine learning are based on a general-purpose classification paradigm. Question is, are these models just a complicated way of performing numerical interpolation??After all, Machine Learning is heavily dependent on the word size used for computations. Last generation of GPUs use a 19-bit floating point format, which allows faster computations than the 32-bit and eliminates the numerical instability problems that plagued 16-bit GPUs.?
领英推荐
When simulation science originated as a branch of dynamical system theory, main trend was toward developing domain-oriented, highly specialized models. These first attempts floundered as they required computational resources ahead of their time, and interest shifted to more standard architectures in order to ensure computability. Now, a more synthetic approach is getting popular, leveraging deep learning as an approxymation of the more precise (and computationally expensive) results achieved by domain-specific modelling.
Many of today's Deep Learning techniques will be soon implemented in the Internet-of-Things. The idea of voice-operated appliances has survived the demise of Alexa-type boxes and its slowly gaining traction in the IOT sector. As every house appliance needs to support the buttons, slides and switches we use for operating it, a voice-operated user interface could lead to substantial savings in manufacturing costs. This may contribute to the diffusion of Basic English and to the loss of national linguistic identities.
Question I: Are subsidies any good for semiconductor industry?
The CHIPS act in the US will move amounts in the order of 100$ billion, while quantum computing just got 10$ billion, etc. Are these amounts any good for the industry, or do they reflect some general uncertainty about the future of computing?
As always, money-based incentives may distort the semiconductor sector, especially as that market is very cyclical and prone to overproduction. During a price crash, the bigger the subsidies received for establishing a facility, the bigger will be the inducement to close it down.
As an example, in 1995 Siemens proposed the UK government to build a 1￡ billion fabrication facility in North Tyneside. In order to seal the deal, Siemens was offered public support at an amount of about 64￡ million. The facility opened in 1997, taking just a little more than a year to build, but it was closed after just two years' operation: the memory chips it produced earned very little margin from the start, and faced fierce competition from Asian producers. At first, the chips were still sold at one seventh of their original price but, in 1998, a further crash brought their market value down to 1 dollar each, well below break-even.
Sometimes the sheer amounts of money involved act as an inducement to waste them in dubious activities. In 2014, China established a 20$ billion fund (known as The Big Fund) aimed at gaining 70% in national chip production by 2025. Plans for establishing new chip fabrication plants were soon developed and launched all over the country.
Wuhan Hongxin Semiconductor Manifacturing (established in November 2017) managed to float a plan for creating a new 20$ billion fabrication plant in Wuhan. Strapped for liquidity, they got land from the municipal authorities and loans from the same subcontractors building the facility. They also bought expensive equipment and mortgaged it soon afterwards. As they escalated their strategy, they successfully targeted provincial and regional funding. In the end, they were prevented from asking a loan to The Big Fund by a bankruptcy request for a mere $5 million, originating from one of their building subcontractors.
At least two former managers in the semiconductor industry and a former Minister for Industry and Information Technology have been detained by the Central Commission for Discipline Inspection, the Party's corruption watchdog.
Question II: What about Europe?
Europe has established a 43€ billion plan aimed at producing 20% of the world's semiconductor ouput by 2030. However, individual states have conflicting priorities, as shown in the case of the Intel European mega-fab.?
A year ago, Intel announced they had selected Magdeburg (Lower Saxony) for their European chips production facility. A grateful German government undertook to cover about 40% of its cost of at about $7.2 billion. Over last year, estimated construction costs levitated from $18 billion to about $21.7 billion, and the Germans signaled their reluctancy at covering the difference.
As Intel is hoping to get more state-funded aids, they are now considering to split the fab into two. Italy is very interested in hosting the packaging and assembly plant for German-produced chips, offering again) to cover about 40% of a deal worth at least $5 billion. Last December, Reuters reported that Italian Prime Minister Giorgia Meloni was seeking a meeting with Intel as she considered theirs investment as “highly strategic”.
Stay tuned for the third part of this issue!
It's wonderful that you're diving into the complexities of #globaleconomy and #technology wars! ???? Remember what Nelson Mandela said - Education is the most powerful weapon which you can use to change the world. Keep exploring and sharing your insights to inspire change! ???