Interposers are not a novelty, but much of the research and development of NVIDIA, AMD, Intel or other companies in the sector have gone in this direction in recent years, since in the same way as the development of chiplet-based architectures is important, so is your intercom.
What is an interposer?
Technically an interposer is a type of PCB, so it is a board in which several chips are mounted on top of it. Interposers are traditionally so called due to the fact that they are a plate that is located between what is the main PCB and the chips that go on top of it. For example an MXM module in which the dedicated GPUs for laptops are mounted can be considered interposers.
This is also the case of some chips with external “embedded” memory, such as the L4 for the Intel Iris that came out a few years ago, where both the embedded memory and the SoC were mounted on the same interposer.
But, in recent years the use of interposers has been related to what are called 2.5D IC integrated circuits, which are based on the implementation of a 3D DRAM memory, usually HBM and therefore use vertical interconnections. These interposers are designed to provide communication as fast as possible with the lowest possible energy consumption, and if there is something that gives HBM memory an advantage over other memories such as GDDR6, it is its pJ / bit ratio.
Simply if you do the calculation of multiplying the pJ / bit by the bandwidth per second, in bits, you will be able to see the consumption and how the HBM2 is a memory with a higher performance for consumption. However, much of the magic is due to the interposer that communicates the different parts, which is essential in the construction of a 2.5DIC system
Moving data is expensive
If you ask any expert in the field who has been designing new architectures then you will see how the answer will always be the same. The problem is no longer in getting as many calculations as possible, but in getting as much data movement as possible within a fixed energy budget.
All chips have a series of data pins that serve to send and receive information between them, each sending of information has an energy cost, which to increase the bandwidth requires two standard solutions:
- Increase the number of pins, which leads to the increase in the size of the chip, this is less chips per wafer and the reduction of the available stock. Apart from creating additional latency problems due to the larger size.
- Increasing the clock speed of the external interfaces is not an increase in price, but rather in consumption, since consumption increases quadratically with the clock speed and this increases linearly with the voltage.
The solution that was reached was to place the pins vertically, in such a way that there can be many more without increasing the area of the chip. The idea is the same as what we can see in CPU sockets where we have hundreds of pins that communicate with the board. The idea with the interposers in 2.5DIC configurations they communicate vertically with all the chips and then it is the interposer itself that is responsible for moving this data.
So we are talking about a large number of interconnections that have to be carried out by the interposer, which means that it has to have enormous internal complexity and that these are totally necessary in these configurations.
Types of Interposers
There are different types of interposers, which we are going to talk about next, not mentioning specific brands or proprietary technologies, but rather giving a general explanation of the different types of interposers that exist. However, it must be clarified that all of them are designed for the implementation of 2.5DIC systems.
They are the most used and at the moment the only ones that exist in the large-scale industry, for all the images that accompany this entry are of this type of interposers.
They are called like that due to the fact that they are still another chip, but on a large scale. The problem? They are expensive to manufacture and can range from $ 30 to $ 100 in cost. In addition, they have the problem that they cannot offer clock speeds for communication beyond 4 GHz. Since the HBM memory requires less clock speed at the moment, it has not reached its limit.
They are those that are usually used in their construction organic elements such as epoxy resin. They do not have the same performance as a conventional silicon interposer, since it allows much lower clock speeds, but their cost of production is very low and is between 2 and 3 dollars.
The reason why they cannot reach the clock speeds of silicon interposers is mainly because they cannot tolerate high temperatures, so this conditions the design of the interfaces that are created on them, as well as their use.
Organic interposers do not use silicon pathways for intercommunication, but there are so-called 2.1D interposers that are a combination of silicon and organic interposers, where the high transfer speed of silicon at a lower speed is sought. cost.
Glass Optical Interposers
There are also interposers made of glass, these are not based on communication using electrons, but rather they communicate by photons. They are therefore the ones with the highest performance, but their manufacture is complex. Most likely, they will end up replacing the silicon interposers when necessary.
Being able to communicate by optical intercom eliminates the need to run tracks through silicon through the interposer. So it is a different paradigm, which we are going to take years to see.
We must also take into account that they will need a new type of manufacturing, since we are not talking about the same methodology to manufacture them as silicon and organic interposers. This can limit its availability to high-performance equipment such as servers and supercomputers for years.
Silicon bridges or interposers without TSV
The bridges are silicon interposers, but with a particularity, they do not use pathways through silicon. This solution will become famous in a few years, since it is the form of intercommunication chosen by Intel with its EMIB and a variant of it is specified in the patents of future systems based on AMD chiplets, so they are becoming the ideal alternative to silicon interposers for the domestic market.
The idea of bridges is to directly connect the chips instead of having an internal interconnection. It is an ideal solution for when there are few chips on top of the interposer and with a much lower cost.