In order to make chips faster, designers often need to make the cores bigger and bigger, but this also greatly increases the difficulty of manufacturing. The good news is that, with AMD leading the way, the industry is actively embracing “small chip” (chiplet) design. After a painful decision to exit the chip manufacturing business in 2009, the company has been trying to turn its fortunes around to re-enter the server chip market, which was dominated by Intel for years.
Unfortunately, AMD executives concluded at the time that the company did not have the resources to compete head-to-head with Intel in the broad server chip space.
For the smaller AMD, doing so would not only be extremely expensive and difficult but also would not allow the company’s server chips to stand out.
AMD senior vice president Samuel Naffziger recalled that they only had one bullet left in their hands that could be fired.
After some review and brainstorming, AMD engineers came up with the “small chip” design concept that would become a major form of chip design in the coming years.
Specifically, instead of trying to pack a lot of functionality onto a large silicon chip, they stitched together the flagship chip from four separate parts.
This new approach was a huge boost for AMD, and Naffziger considers it one of the greatest engineering achievements in the industry in recent memory because it solved so many problems in one fell swoop.
Although AMD initially invented the small-chip design approach to meet its own needs, it also managed to reduce manufacturing costs by 40 percent by breaking the chip into more and smaller parts.
"This also allows the company to easily plan a complete set of server chips compared to its competitors -- adding or subtracting as needed to provide SKUs that cover multiple performance/price bands. Even better, by transitioning to smaller chips, AMD can reuse two small server chips and design lower-cost models that are also suitable for desktops, thereby capturing a more lucrative market."
After Dr. Suthep took over the company, the small chip initiative grew AMD’s revenue from $4 billion in 2015 to $16.4 billion by 2021 and is expected to save the classic Moore’s Law.
"What AMD achieved years ago is now becoming a model for the industry. Even rival Intel is planning products based on small-chip designs. In addition, other companies in the industry are talking about building a set of standards that will one day allow silicon IP from different vendors to be mixed and matched in a single package."
The world is growing so fast to produce and process data that without small chips, the manufacturing process will become too expensive and difficult to continue to provide software developers with the annual leaps in computing power they need.
In the long run, older chip designs will consume too much power and sooner or later become less economically viable.
"We're going to be stuck in a situation where you can only keep buying box after box with the same performance and functionality package. In the end it's one or the other -- either expand and build more data centers and power plants, or suffer the slowdown in Internet and data growth."
Another fascinating aspect of the small-chip concept is that it could continue the “Moore’s Law” set forth in Gordon Moore’s seminal 1965 paper — that chips get faster and cheaper every two years as transistor density doubles.
At the same time, Moore described how the economics of breaking a single chip into smaller parts would one day make more sense. Mix and match components in a way that will provide system designers with greater flexibility, complemented by performance gains and other benefits.
"It may prove more economical to build large systems with smaller functions that are packaged and interconnected separately," Moore writes. The availability of large functions combined with functional design and construction should enable manufacturers of large systems to design and build a large number of different devices quickly and economically."
Moore’s Law (PDF screenshot)
In fact, IBM was already building conceptual systems containing small chips as early as 1964, when this was the only way to achieve the necessary computing power.
Over the decades, companies like IBM have continued down this route and have taken the loose concept of the small chip and applied it to the most complex and expensive systems such as supercomputers and mainframes.
"The problem is that small chips used to be complex and expensive, but in recent years, manufacturers are working to integrate more features onto a single silicon chip. In addition to smartphone SoCs, we're using the latest designs in some server processors, as well as in notebook/Mac hosts using Apple Silicon."
Manufacturing, with the increasing scale of the chip, even with the support of advanced processes, the larger silicon area will lead to lower yields.
However, Jefferies research shows that through the flexible use of small chips, manufacturers can easily server chip size, to five times the size of a typical PC chip.
As for the GPU, NVIDIA, AMD’s technology line and focus is not the same. In addition, Intel senior researcher Debendra Das Sharma said.
"By packaging memory into heterogeneous single-chip systems as well, SoCs can also reap huge benefits such as high memory bandwidth, low latency, and low power consumption. Mixing and matching small chips allows chip vendors to easily customize a variety of silicon IP for large customers with specific needs, such as accelerated computing commonly used for AI computing tasks. But if a customer needs a chip for a specific type of AI, Intel can also leverage general-purpose gas pedals instead of more specialized gas pedals."
Finally, the universal interconnect small chip is not yet a reality, but according to several industry observers, a new version of the standard may be available around 2025.
