15-01-2021 | | By Robin Mitchell
Recently, AMD announced that it had filed a patent for its next generation of chiplet-based GPUS. What are chiplets, what advantages do they provide semiconductor designs, and what is AMD doing differently to other chiplet designs?
Recently, AMD has filed a patent for its next generation of GPUs that will utilise chiplets instead of large-scale monolithic semiconductors. According to the patent, AMD will be taking a radical path that has not been done with chiplets before; sharing of on-chip resources.
Most chiplet designs break a system down into core components such as the CPU and GPU. Still, with technology becoming ever more complex, AMD will be breaking the GPU itself into individual chiplets that will then be connected.
GPUs are large parallelised systems that require common resources with fast access times, and it is this fact that has seen GPUs always integrated onto a single piece of silicon. However, silicon technology is at its limit, and companies cannot simply make dies larger for a whole list of reasons. Furthermore, when circuits are separated, access times increase, and buses are often reduced in size.
According to the patent, AMD will solve this issue by utilising GPU chiplets that use their own L1 and L2 cache, but the L3 cache will be accessible by all chiplets on the design. The chiplets will be connected using a specialised bus called the HBX passive cross-link, allowing chiplets to access low-level cache on other chiplets as fast as they can access their own local cache.
The filed patent also describes the structure of the chiplets being oriented vertically instead of horizontally. Such a design decision makes sense when considering that silicon dies can only be so large. Even those designed using chiplets still face similar dimension challenges (i.e. packing and placement).
Recently, there has been a lot of talk surrounding chiplets, and how companies such as AMD have been using them to produce the next generation of devices. Simply put, a chiplet is a small piece of semiconductor that holds a functional circuit used to make a larger circuit. Traditional semiconductors integrate all circuit functions onto a single die, but a chiplet spreads the functions across multiple smaller dies.
Chiplets are often bonded to a substrate and then connected using wires or traces. Designs that use chiplets can also integrate SMD components such as capacitors and resistors for filtering and current limiting purposes. Chiplets are still a new technology, and only a handful of companies use them, but their advantages will quickly see them become the dominant technology in the semiconductor industry.
The advantages of chiplets are only realised when dealing with cutting-edge semiconductor technology. When fabricating a semiconductor, the probability of a single transistor malfunctioning is very small, and a device containing one million transistors has a relatively good chance of working.
However, modern devices with extremely small feature sizes can have billions of transistors, which significantly increases the failure rate of dies. Therefore, chip designers have two choices to increase yields from wafers; either reduce the total number of transistors (making the device less powerful) or increase each device's price.
Increasing the price can be done, but many modern devices are aimed at the consumer market, and thus increasing prices too much will make such technology unavailable for commercialisation. Instead, designers are looking towards chiplets to solve their problems.
Chiplets essentially spread out the high transistor count across multiple devices. If one transistor fails in a billion, then 99 out of 100 chiplets could be useable whereas a single device containing one billion transistors would fail. Since many defects in silicon wafers are point-defects (meaning specific areas of the wafer are unusable), the use of smaller dies can produce more usable devices and therefore increase yield.
Chiplets also allow customers to create a customised device using chiplets as building blocks. In a similar scenario to Apples M1 processor development, a customer could choose to combine a CPU with various functional blocks aimed at their application. The use of building blocks also reduces the price of such designs as semiconductor designers can focus on developing standard glue-logic parts akin to that of the 4000 and 7400 series.
Chiplets will become the key to future semiconductor technology as foundries face increasing difficulty reducing the size of transistors. Will chiplets be stacked vertically? Will they be easily customisable? The answers to these questions are unclear, but what is clear is that the chiplet is here to stay.