“The number of transistors and resistors on a chip doubles every 24 months.” – Gordon Moore.
It seems that Moore’s Law has finally become redundant. The size of transistors is now so prohibitively small (the smallest on record is 5nm), that it is becoming less and less efficient to manufacture them to successfully continue along the path predicted by Gordon Moore back in 1965. At this microscopic scale, it is almost impossible for transistors to function normally without also interfering with basic physics.
The above diagram illustrates transistor count over the past few decades and the corresponding CPU power and clock speeds. Towards the early 2000s, right about the time Moore’s Law starts faltering, clock speed and CPU power start collapsing because of transistors’ diminishing size and their associated issues (at 90nm the transistor gates became too thin to prevent current from leaking out into the substrate).
So what do we do now?
The Exponential Growth of Data
Walmart processes about 267 million transactions in their 600 stores worldwide. FICO (the fraud detection system) manages 2.1 billion accounts all around the world. LSST (the Large Synoptic Survey Telescope), records approximately 30 trillion bytes of data per day. Now that Moore’s Law has come to an end, how do we continue to leverage technology to support these massive surges of data?
Distributed Processing Systems
Amazon and Netflix are pioneers of this system. Distributed systems allow you to channel the processing power of multiple CPUs to execute or run a single application without having to rely on the processing power of individual machines whose power cannot scale to requirement due to the physical limitations described above. However, it is not without its limitations. Security is a lot harder to manage in distributed systems, and there is currently very little support in terms of software. Troubleshooting and diagnosis can also be challenging.
Re-thinking Existing Structures
Chip Level Enhancements
In 2011, Intel introduced 3d transistors which are thin, 3-dimensional silicon fins that rise from the substrate and is enveloped by the gate on all three sides, instead of just one. This allows a higher level of control and enables as much transistor current as possible to flow during its ‘on’ state and as close to zero as possible when in the off state.
Photon transistors on the other hand aims to replace electrons with photons. The technology is currently still under research but aims to use PTNTs (Photon Triggered Nanowire Transistors) that contain long crystalline silicon segments connected by short porous silicon segments that are sensitive to light and can be used to control the flow of current.
Architectural Level Enhancements
GPUs were formerly used for 3D game rendering, but they are now used for processing more complex computational tasks such as financial modeling, cutting edge scientific research, and oil & gas exploration. Compared to a CPU, a GPU has hundreds of cores, can handle thousands of threads simultaneously, has more power, and is more cost efficient than the CPU. A GPU is also capable of solving incredibly challenging problems that are inherently parallel in nature, such as video processing, image analysis, and signal processing.
The GPU almost always works in tandem with the CPU. The input data is copied from the CPU memory to the GPU memory then the GPU program executes the caching data on the chip for performance. The results are then sent back to the CPU memory from the GPU memory.
Field Programmable Gate Arrays are used engineers to design a special form of integrated circuits. Compared to GPUs, FPGAs are more deterministic, has a latency in hundreds of nano-seconds, and the hardware implementation for algorithms are a lot faster. GPUs tend to be problematic in battery dependent scenarios, and their latency is in single digit micro-seconds.
Application Specific Integrated Circuits are similar to FPGAs, but are designed specifically for applications. ASICs are used for auto emission control, environmental monitoring, and for personal digital assistants.
TPUs or Tensor Processing Units were invented by Google to power neural network computations for a host of Google products behind the scenes. The TPU delivers 15–30X higher performance and 30–80X higher performance-per-watt than contemporary CPUs and GPUs allowing Google’s services to run state-of-the-art neural networks at scale and at an affordable cost. They enable the identification of objects and faces in photos, recognition of the commands you speak into your Android phone, and the translation of text from one language to another.
New Computing Paradigms
Quantum Computing will be a game changer. The basic difference between a quantum computer and a digital computer is that a digital computer uses binary states to encode their data, while quantum computing uses binary bits which can be in superpositions of states.
DNA Computing uses DNA, biochemistry and molecular biology hardware instead of traditional silicon based computer technologies. DNA molecules, the material our genes are made of have the potential to perform calculations many times faster than the world’s most powerful human built computers. DNA might one day be integrated into a computer chip to create a “bio-chip” which will push computer processing power to hitherto unimagined levels.
The above article is based off an Innovation Session conducted by Daham Pathiraja from the Ops Engineering team at Sysco LABS, titled “Sustaining Moore’s Law”.