Intel is clear, the chiplet is the future, and has a very ambitious vision

Intel is clear, the chiplet is the future,

and has a very ambitious vision

Intel is clear, the chiplet is the future, and has a very ambitious vision

One of the most exciting news Intel offered during Architecture Day 2020 came from Brijesh Tripathi, vice president and chief technology officer of the chip giant's client computing group. The executive confirmed that the chiplet is the future, and gave us a fairly comprehensive view of the value that such architecture offers.

As many of our readers will know, Intel and AMD use two types of architectures in their microprocessors , and each has, as we saw in this article, its advantages and disadvantages . The monolithic core architecture used by the chip giant has, as main advantages, that it allows to achieve processors at higher frequencies, with a higher margin of overclock, lower latencies at the cache level and with a less dependence on memory speed RAM. Its main disadvantages are, however, very important:

  • It is more difficult to advance with it to lower manufacturing processes.
  • Its cost is much higher, both at the level of R&D and production.
  • The impact at the wafer level is much more marked, since it does not have the same degree of use of defective chips.

Intel is clear, the chiplet is the future, and has a very ambitious vision

In contrast, the chiplet-type architecture, also known as multi-chip module (MCM) , has important advantages that facilitate the design and creation of new chips , both in their early stages and in their final stages. These are its most important advantages:

  • Facilitates process jumps.
  • Improves the success rate per wafer.
  • Helps reduce production costs.
  • It allows a high degree of exploitation of defective chips.
  • It makes the design of chips with many cores possible while maintaining a low price.

the chiplet-type architecture, also known as multi-chip module (MCM)

To better understand these differences, we are going to give a simple example. Intel's Core i9 9900K is a processor with 8 cores and 16 threads that are integrated into a single silicon chip. This means that for Intel to be able to take advantage of the chips it obtains from a silicon wafer and use them to create this processor, the eight cores of that tablet must be fully functional. If only one of them doesn't work, you can't use it to build that processor.

AMD's Ryzen 7 2700X also has 8 cores and 16 threads, but they are divided into two small silicon chips . Each has four cores, and they communicate with each other through an interconnect system called the Infinity Fabric. This means that AMD does not need to source eight active-core silicon chips to make the Ryzen 7 2700X, but only needs to source two active four-core silicon chips. The complexity of a quad-core silicon chip is lower, and the risk of something going wrong as well, allowing us to understand why its wafer-level impact and costs are so much lower.

Why is Intel praising chiplet-based designs?

Why is Intel praising chiplet-based designs?

During the last years we have seen how Intel has gone from a total rejection of the MCM architecture that AMD uses to considering that the chiplet is the future. It is no coincidence, Intel already had to resort to this architecture with the Pentium D and the Core 2 Quad , the first was a chip with two Pentium 4 glued together, and the second the same but with two Core 2 Duo attached.

The multi-chip module architecture allows creating multi-core processors in a simple and economical way, since it eliminates the complexities, problems and complications inherent to monolithic core designs , and at a sacrifice in terms of performance that, today today, it is minimal.

The years of effort and resources that Intel has put into advancing its 10nm and 7nm processes have confirmed that the monolithic core design is not as viable as previously believed, both economically and technically. Its impact at the wafer level when moving in configurations with a high number of cores in very high, and the improvements it offers compared to the MCM architecture are not enough to compensate for that reality.

In the attached image we can see three different designs. The first corresponds to the monolithic core architecture and requires, according to Intel, a development period of between three and four years . In addition to this long period of time, it does not present any degree of reuse, and is more prone to errors. The second is the type of MCM design currently in use by AMD, requires two to three years of development, has fewer bugs, and has some degree of utilization (AMD uses the Zen 2 design in both its general consumer processors and its EPYC line ).

Intel is clear, the chiplet is the future, and has a very ambitious vision

Finally, we have a multi-chip module design based on a new generation chiplet architecture that reflects Intel's goal, that is, that fits with the idea of ​​chiplet as a solution for the future according to the chip giant. In this design, each of the elements of the chip can be divided into a large number of chiplets , which allows building solutions with a high degree of customization that also present a very low cost, both technically and economically, they present a low impact at the wafer level and have a high degree of reusability.

Create a highly scalable architecture, easy to design and go to the wafer, with a high degree of reliability and a good balance between performance, customization and reusability, that is where Intel sees the future of the chiplet, and does not rule out using solutions with different manufacturing processes , something that is certainly not new, since AMD has already taken that path with Zen 2, an architecture that combines 7 nm chiplets with 12 nm I / O units.

I remind you that, if all goes according to plan, the first general market chips based on such a design will arrive with the Alder Lake-S architecture, which will combine high-performance core blocks based on the Core series with cores of high efficiency based on the Atom series. They will use, in theory, the 7nm manufacturing process.

No comments