Chiplet: What Is It And What Is It For?

What is a chiplet? surely it is one of the questions that you are asking yourself if you have arrived here. And it is that it is a term that is being heard more and more in the field of CPU technology and also GPUs mainly. It’s nothing new, but it has never been as important as it is today, and it will continue to be so for the foreseeable future. However, this technique already has very promising successors, since it is not without disadvantages.

What is a chiplet?

AMD Chiplet
AMD Chiplet

chiplet is nothing more than an integrated circuit block, or chip, that has been designed to be able to work with others similar or identical to it. In this way, a chip with a larger or more complex area is replaced by several smaller chips.

Motivation

The reason why chiplets began to be used is related to the growth of the chips and the yield or manufacturing performance. One of the first times they were used for commercial processors was in the AMD Zen, as announced by Dr. Lisa Su in 2017.

As foundries have upgraded manufacturing nodes to create smaller and smaller components, high-performance chips have increased in complexity and surface area, and costs have continued to rise. For this reason, making larger monolithic chips is becoming less economically viable. Hence, it was decided to use designs based on chiplets, dividing a complex system into several smaller and simpler ones.

Keep in mind that with a die size of approximately 18 x 20 mm in size, which would be a medium chip, it has a surface area of ​​360 mm², and that makes it possible to manufacture around 150 wafers in a 300 mm diameter wafer. chips. Dividing such a design into four 9.5×10.5mm chiplets yields approximately 99mm², resulting in up to 622 chips being fabricated on one wafer.

This will cause the yield to multiply in the second case, that is, it will result in more functional chips than in the first case. Therefore, costs can be reduced in this way. How? In the following image you can see it more intuitively:

Yield Wafer
Yield Wafer

In this image you can see the following:

  • The 3 wafers are of equal size, let’s say 300mm.
  • However, the sizes of the chips are different in the three cases:
    1. 40×40 mm in the first case.
    2. 20×20 mm in the second case.
    3. 10×10 mm in the third case.
  • The resulting yield is:
    1. 35.7% in the first case.
    2. 75.7% in the second case.
    3. 94.2% in the third case.
  • Good chips are marked in dark gray and are:
    1. 10x die in good condition.
    2. 103x die in good condition.
    3. 620x die in good condition.
  • Defective chips are marked in green:
    1. 18x defective die.
    2. 33x defective die.
    3. 38x defective die.
  • Total chips per wafer:
    1. 28x die/wafer.
    2. 136x die/wafer.
    3. 658x die/wafer.

Did you know that binning is a technique to try to take advantage of as many chips as possible from a wafer or wafer. To do this, once they have been manufactured, they are subjected to a series of tests. Defective ones that don’t work at all are discarded, but some chips that may have some of their core(s) intact are used. Those who cannot work at such high frequencies can also be highlighted. Why do you think Intel Celeron, Pentium, Core i3, Core i5 and Core i7 exist? All wafer chips are built to be i7, but not all make it…

As can be deduced, the smaller the chip size, the greater the productivity, wasting fewer chips and being able to lower costs by producing a greater quantity for each wafer. Also, you should consider that Moore’s Law predicts the growth of complexity, and the advent of multicores has made the situation even worse. Just by reducing the nodes, the chip sizes cannot be maintained, with the consequent price increase.

It seems that the chiplet trend will go beyond CPUs and GPUs, also reaching other sectors such as automotive chips, memories, etc., becoming a trend.

Differences with MCM

MCM Vs Chiplet
MCM Vs Chiplet

There is a type of packaging that has been used for years called MCM (Multi-Chip Module), that is, a type of assembly that uses several chips on a substrate and interconnected with each other, sharing pins or conduction pads.

IBM POWER processors have been using MCM packaging, and now so is AMD for its latest processors. And if you wonder about the difference, the truth is that MCM is a term to designate the form of packaging and chiplets refer to the chips into which a system is divided. In other words, each IC is a functional subset of the complete system and is called a chiplet. All chiplets bundled together will complete the complete set.

For example, an AMD EPYC is divided into chiplets, but in turn, uses MCM packaging. Intel has also used this type of packaging in some cases, such as when integrating its GPUs with the CPUs, a different strategy than the one followed by AMD in its APUs. You have more cases in the AMD Radeon Instinct MI200 GPUs, in the Intel Xe Ponte Vecchio, etc.

Regarding the parts that can be seen in this type of packaging:

  • Chiplets: are the chips into which the system has been divided. Subsystems can be of various types, such as memory, processing units, communication units, etc. For example, the IBM POWER6 in the previous image is divided into:
    • CPU cores
    • Off-die L3 cache chips
  • Interposer: It is the substrate or PCB that interconnects the integrated circuits or chiplets. It is often made of organic material. In the case of the IBM POWER6, it is the brown area.
  • Others – It’s also common to see surface mount components on top of the interposer, such as resistors, capacitors, etc. You can see in the IBM POWER6 all the SMD components that surround the chips.

It is interesting to note that chiplets and new packaging systems will allow the integration of such interesting chips as various processors, memories, programmable FPGAs to be used as dedicated accelerators, etc., that is, the so-called heterogeneous computing.

Advantages and disadvantages of a chiplet versus a monolithic chip

Advantages and disadvantages of a chiplet versus a monolithic chip
Advantages and disadvantages of a chiplet versus a monolithic chip

Not all are advantages in the case of MCM or chiplet designs, they also have their drawbacks compared to monolithic ones, where the entire system is integrated into a single chip. Among the advantages of chiplets are:

  • Minimize design time.
  • Reduce costs, increasing yield.
  • Possibility of using more memory or more processing cores. That is, more scalable as they are modular.
  • Possibility of using several manufacturing nodes in the different chiplets and even different logic families, and even different semiconductors (Si, SiGe, GaN, GaAs, InP,…).

In the case of the disadvantages of chiplets are:

  • Higher communication latency between units.
  • Need for more complex interposers.
  • Implementation of communication logic.
  • The failure in the connection of a die can mean that the entire device stops working.

Currently, some solutions are being developed and beginning to be commercialized that will try to eliminate these disadvantages of chiplets, while keeping the advantages that this division paradigm brings. I’m referring to the new 2.5D and 3D packages, where you stack the chips and achieve latencies similar to those of a monolithic, but you can still manufacture each stacked chip separately. You have an example in the 3D V-Cache, where chips with the largest cache memory are placed on the processing units.

The problem with this type of technology is that it still lacks standardization since each manufacturer is using its methods. Therefore, if chips from different manufacturers needed to be packaged in the same device in the future, they would not be compatible. Surely you have heard terms like Intel Foveros, Intel EMIB, TSMC LSI, TMSC CoWoS, etc., all of them with some differences…

Monolithic chips: advantages and disadvantages

Monolithic chips, just like chiplets, also have their advantages and disadvantages which I summarize here:

  • Advantages of monolithic integrated circuits :
    • Smaller packaging size, making it more positive for mobile or embedded devices where size is at a premium.
    • Less complexity in packaging and therefore cheaper.
    • Better reliability due to fewer parts involved (fewer leads, solder joints, etc. that can fail).
    • It can mean lower power consumption.
    • Data transfer between the different parties is faster and with less latency given their proximity and that all units usually work at the same frequency.
    • Eliminates the need for additional logic to coordinate the different chiplets.
  • Disadvantages of monolithic integrated circuits :
    • Higher manufacturing costs than chiplets due to lower yield.
    • Greater concentrated dissipation power.
    • If just one transistor or interconnect fails, the entire IC stops working.
    • Lower flexibility and scalability.

Conclusion about the chiplet

Although MCM chiplets and packaging have their advantages, some designs, such as SoCs, still try to stay as monolithic chips as speeds are improved. It cannot be said that chiplets are or are not the future of high-performance chips, rather the future will be a hybridization between different technologies.

Chiplets are not incompatible with other technologies such as 2.5D or 3D, in fact, organic substrate interposers can be used as in MCM, and also smaller silicon interposers for stacked chips.

On the one hand, we will see chiplets and MCM packaging, at the same time that these chiplets are made up of stacks of integrated circuits (2.5D/3D) to bring memory and logic closer to reducing latencies, which are currently a bottleneck between the CPU and RAM. In addition to all this, other technologies will also arrive that could help boost performance even more, such as optical interconnections, for much faster data transfers. Something similar to the change from copper cables to fiber optics in telephone lines.

And with the arrival of the ceiling of silicon technology, all these technologies will become more and more important. Although some challenges must be solved, such as power dissipation when the level of integration is so dense. The famous decades-old claim of “Smaller, Faster, and Cheaper” will cost more and more.

Read more about Analyzing The Architecture Of Apple Processors: M1, M2, And their Variants

Website | + posts

Abram left his e-business studies to devote himself to his entrepreneurial projects. In 2017, he created the company Inbound Media and wrote articles about high-tech products for his Chromebookeur site. In 2019, Chromebookeur was renamed Macbound and became a general purchasing advice site. Today, Abram manages the development and growth of Macbound, surrounded by a young and talented team.

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *