As you may already know, a lot has been said about PAM-4 and the importance of this technique to achieve high speeds for standards such as GDDR6X. But as speeds grow and bandwidths get bigger, PAM-4 won’t be enough, and therefore the industry is debating where to go to keep efficiency and costs under control. What’s after PAM-4? Will it affect the new GDDR7 and DDR6 ?
PAM-8 is actively talked about as a supposed successor, but the reality is that the industry is ruling it out because it is not as complete as they want for certain buses. The trend seems to be to double the bandwidth with 30-50% more power , because using more energy the profitability is not adequate.
From parallel to serial connections?
You have to take into account where the market is heading. The 10GBASE-T connections use PAM-16 with four lanes to transmit 800 million symbols per second, but this avoids enormous speeds like the ones that have been seen of 200 GHz.
Here the problem is in the signal loss in dB and the energy efficiency. Going out of a chip to send data over a bus has a serious impact on energy efficiency, so manufacturers try to avoid that.
Therefore, there are two open fronts: which standard to use for buses outside the package and which one to use for those buses that are inside the package. As we say, we can use PAM-8 for synchronization, but the consumption skyrockets and the speed drops, or speeds like the aforementioned 200 GHz could be used , but at the cost of crazy consumption.
As we can see, the problem is the efficiency to achieve that the signal arrives as cleanly as possible, so possible hybrids between series and parallel in these connections are being studied.
Can’t use QAM like wireless networks
The impossibility of cleaning the signal with PAM-8 and PAM-16 leads researchers to look to another standard such as QAM, used in high-speed wireless networks. The theory is simple, we take several bit streams as if they were channels and apply PAM to each one, after which each stream changes phase and mixes on the transmission cable.
This would logically be a palliative measure and would end up being obsolete when the electrical options on boards and processors reach their limit. Researchers seem to have the solution: optical signals.
These signals would allow greater control of the modulation, being possible to use NRZ or even PAM-4, but there remains a basic problem: the price. It is true that when the cost was standardized it would drop a lot, but until then the price is exorbitant. That is why serial connections such as those used by Intel in its EMIB technology are in fashion, since greater bandwidth is achieved by accumulating more lanes than by increasing the speed itself.
But there are manufacturers in the industry who do not think this and allude to the fact that speed will always be the determining factor to achieve more performance, to the point that we could see up to 10 GHz in a few years, since the oscillators are getting better and better. . The question here is whether these standards to be developed will use PAM or QAM, because for example DDR6 cannot use NRZ , so it will have to adapt to one of these two.
Therefore and in summary, the future looks to the serial technology and its connections will dominate the market, since less cables are needed in the chips and PCBs. Although yes, it would have a higher latency, so it would be necessary to see where it really affects the system and especially the costs, since serial connections require more sophisticated and expensive circuits.