Although we do not see it, the interconnections of the future are currently being debated and analyzed. Artificial intelligence and large data centers are looking for a key feature that will define the future of the sector and communications, as well as data centers: coherence. And here are two great contenders: CXL and CCIX, and although they compete with each other, it so happens that AMD has one foot in each of them, but Intel only supports and develops the first, which is better and who will win?
Although we have already seen them separately, the industry is torn between supporting one or the other. Not surprisingly, many companies have chosen the AMD path and support both at the same time, but, although they are complementary, in the end, the sector wants to work with you because, as always, there are trends to follow.
The sector is at that point where everyone looks to see what the opponent and rival are doing, so it is a good time to understand where they can go and, above all, to know who will win the battle.
Is CXL vs CCIX a new stage of the war between Intel and AMD?
Not quite. Intel supports only CXL since it was the main developer and promoter, where now AMD has seen its benefits and has joined the consortium to have access to the technology and the decisions made with it.
This drive with CXL where the main companies in the world are located is not going to generate a battle between them, but rather Lisa Su’s are going to be neutral, letting companies choose their coherence systems and opting for the winner. Instead Intel has to lead and that is an advantage and a disadvantage.
The first problem is basically that CXL is a connection protocol for CPU, whereas CCIX is a much more open system. If we consider that more than 80% of the world’s servers integrate Xeon CPUs, we will quickly understand the virtues of working with CXL, since consistency falls on the processor, with the remaining chips or components connected to CXL being complementary and slave.
This allows a bidirectional coherence between the slaves and the CPU, where to put a simile, the processor is the conductor who gives coherence to the rest of the band. CCIX does not work like this, being more complex from digital logic, since it is completely symmetrical in terms of coherence between devices.
We can have several sockets connected to each other with CCIX, since the protocol does not distinguish or discriminate, since its coherence is symmetric, which involves a basic problem. If an accelerator of the type that is coherence fails, the entire system can reproduce the failure and eventually crash, whereas with CXL asymmetric consistency this would not happen, since the failure can be isolated.
The approach is different, even if they are based on the same principles
Both CXL and CCIX use PCIe to interconnect devices, both seek consistency with each other and with cache memory, but the way to do this is clearly not the same or the approach to achieve it.
Intel created a specific interconnect bus called the Flex Bus , which uses PCIe to interconnect the accelerators, and it does so in a very specific way and speed, which, although they have not been disclosed as such, we do know that it reduces the latency present in the port itself and also achieves a speed of up to 32 GT / s.
CCIX does not have these latency improvements and also in its last revision it achieves 25 GT / s . The problem logically is how each system is thought. With CXL you always have to resort to the CPU to get to the cache at every step you want to implement in terms of consistency, while CCIX progresses its way between the accelerators to finally get to the CPU and treat it like one more, not giving it priority.
In other words, sharing coherence without anyone being above anyone else in an open system has the advantage of being very parallelizable, but in terms of latency it is not very optimal due to the hierarchy itself.
Intel support is making a difference
If you have a giant like Intel behind these types of systems, you have their CPUs on your servers and on top of offering support, support and a series of complementary products such as Agilex, the result that is being obtained is that the CXL ecosystem is deployed already in the main technological giants in the sector such as Amazon or Microsoft.
Since CCIX is also open source, the implementation is looking good and it seems that there are already many gurus in the sector who affirm that, mainly due to its low latency, CXL will be the interconnection of servers and accelerators for coherence systems. with world dominant cache.
Does this mean that CCIX will die? No, nothing is further from reality. There are systems and companies that need symmetrical coherence, especially those dominated by ARM, and that is that neither of the two rivals is tied to a specific type of architecture. This is a very clear advantage that will be coupled to another, the disaggregation of resources.
In this section, Intel with CXL has an advantage, since many companies need to configure a large amount of resources through software and there latency is key again. CCIX symmetry is not enough to manage resources more or less automatically and with better timings. It would require in any case a longer optimization time for the software without having any clear advantage in latency, but to match the content at least.
The problem facing the industry, saving these commented details, is that it has to choose. CXL and CCIX are not complementary , they are self-exclusion and due to what has been mentioned above and in view of how companies are receiving Intel technology, everything indicates that CXL is going to take the cat to the water in record time.
CCIX can stay for extremely scalable architectures that have the time and resources to perfectly manage each die in the data center, but out of there the market will surely opt for the blue team’s commitment to its greatest advantages.