AMD's CCX Architecture: A Scalable and Efficient Design

AMD's latest Zen3 chips have taken their core complexity to new heights with the introduction of the Core Complex (CCX) architecture. This innovative design features multiple cores grouped together along with some cache memory, allowing for significant improvements in performance and power efficiency. One CCX consists of a group of cores, along with some cache memory, and can be found on a chiplet called a CCD (Complex Chip Die). The setup on Zen3 is straightforward, with one CCX per CCD and up to 2 CCDs per processor, resulting in a maximum of 16 cores. In contrast, Zen2 CPUs feature 4 cores per CCX.

Each CCD can instead hold two of these smaller CCX's, resulting in the same 16 core maximum but allowing for greater flexibility and scalability. This approach also enables AMD to reduce costs by only discarding the defective chiplet if needed, rather than wasting entire CPUs. The use of CCXs with Infinity Fabric is a key benefit of this design, as it allows for high-speed interconnects between cores or between CCDs. While Infinity Fabric offers faster bandwidth and lower latency compared to traditional bus systems, its serial connection can result in higher latency.

Infinity Fabric: A Scalable Interconnect Solution

AMD's Infinity Fabric is their secret sauce when it comes to connecting CCXs. This high-speed interconnect enables fast data transfer between cores or between CCDs, making it an ideal solution for large-scale computing applications. While not as fast as direct connections between cores, Infinity Fabric offers a significant advantage in terms of cost savings and scalability. By using smaller chiplets, AMD can reduce the overall size of the CPU and decrease production costs.

Moreover, the use of Infinity Fabric enables AMD to scale their designs more easily. Instead of being limited by the size of their silicon under layer, AMD can simply use multiple Infinity Fabric modules to build outward. This approach is similar to how tiles are arranged on a modern basketball court, snapping into place with ease. Despite its benefits, Infinity Fabric does have some limitations, particularly in terms of thermal constraints. As the number of CCXs increases, so do the heat generated by each chiplet.

The Limitations of Infinity Fabric

While AMD's Infinity Fabric offers significant advantages in terms of cost savings and scalability, it also has some limitations. One major drawback is its serial connection, which can result in higher latency compared to direct connections between cores. This limitation may become more apparent as the number of CCXs increases. Additionally, while Infinity Fabric can transfer data faster than traditional bus systems, its speed may not be sufficient for all computing applications.

Intel's Alternative: Chiplet-Based Design

In contrast to AMD's use of Infinity Fabric, Intel has opted for a different approach. Their chiplet-based design uses smaller chiplets connected by EMIBs (Electro-Mechanical Interface Bus), which offer higher bandwidth and lower latency compared to traditional bus systems. EMIBs are a more elegant solution than Infinity Fabric, allowing for parallel data transfer and reducing power consumption.

Intel's EMIB technology is also more efficient in terms of silicon usage. By stacking chiplets on top of each other, Intel can reduce the overall size of their CPUs while maintaining high performance. This approach also allows for greater flexibility in terms of scalability, as additional EMIBs can be added to increase the number of chiplets.

EMIBs: A Smaller Alternative

Intel's EMIB technology offers several benefits over traditional bus systems and Infinity Fabric. By using smaller pieces of silicon, Intel can reduce costs and improve scalability. The use of EMIBs also enables parallel data transfer, which reduces latency and power consumption compared to serial connections like Infinity Fabric.

Moreover, Intel's EMIB technology is designed to be more flexible than traditional bus systems. Instead of relying on a single channel for communication between cores or between chiplets, EMIBs allow for multiple channels to be used simultaneously. This approach enables higher bandwidth and lower latency, making it an ideal solution for high-performance computing applications.

Chip Stacking: A Future Direction

While Intel's EMIB technology has been successful in their current lineup of processors, the company is also exploring other approaches to chip design. Chip stacking, which involves layering multiple chips on top of each other, offers several benefits over traditional monolithic designs. By reducing the overall size of the CPU, Intel can decrease production costs and improve scalability.

However, chip stacking has its limitations, particularly in terms of thermal constraints. As the number of layers increases, so do the heat generated by each chip. This limitation may become more apparent as the industry moves towards more complex chip designs. Despite these challenges, Intel remains committed to exploring new approaches to chip design, including chip stacking.

A Future with Chiplets

While both AMD and Intel are experimenting with innovative approaches to chip design, it's unlikely that traditional monolithic designs will disappear completely in the near future. At the lower end of the market, simpler designs will continue to be the norm due to cost constraints and performance requirements. However, as the industry moves towards more complex computing applications, we can expect to see a shift towards more scalable and efficient chip designs.

The use of CCXs with Infinity Fabric or EMIBs will become more widespread in high-performance computing applications, where scalability and efficiency are critical. As these technologies continue to evolve, we can expect to see significant improvements in performance and power efficiency from both AMD and Intel.

"WEBVTTKind: captionsLanguage: enas much as amd and intel fanboys like to scream at each other online the average user would be hard-pressed to notice the difference between these companies cpus in everyday use unless you're like us and benchmark things religiously but on the inside modern chips from team red and team blue are built quite differently so let's take a look at how starting with amd ryzen cpus are made up of at least one core complex or ccx but don't worry despite the name it's not a complex thing to understand a ccx is just a group of cores along with some cache memory newer zen3 chips have cc axes with 8 cores while zen 2 cpus have 4 cores per ccx each ccx lives on a chiplet called a ccd the setup on zen 3 is pretty straightforward one ccx per ccd with up to 2 ccds per processor for a total of 16 cores on zen 2 each ccd can instead hold two of these smaller ccx's resulting in the same 16 core maximum but what links all these core complexes together well you've probably heard of amd's infinity fabric which is their secret sauce bus that can run between ccx's or between ccds although it's a high high-speed interconnect that works pretty well with ryzen beating out competing intel chips and a good number of benchmarks it's still not as fast as having the cores more directly connected as there's higher latency with infinity fabric and the connection is serial rather than parallel but there are real benefits to this approach one is that the smaller chiplets mean big cost savings due to better yield if one chiplet is bad you can discard just that one chiplet without wasting as much silicon from throwing away a whole cpu this ccx plus infinity fabric approach also scales quite well so you can just add more chiplets if you need more cores instead of spending more time and money designing a whole new chip is that how intel does it we'll tell you right after we thank brilliant for sponsoring this video brilliant is a website and app built around active learning that's accessible and fun trade boring long lectures for problem solving and interactive visuals there's over 60 courses to choose from including their new everyday math course where you can get a refresher on foundational math topics you might not have touched since high school join the community of 10 million learners and educators today the first 200 people who head to brilliant.org techwiki will get a 20 off an annual premium subscription so intel in contrast to amd still uses monolithic chip designs where cores are right next to each other at least for the time being this approach allows the cores to talk with each other more directly but it has the opposite problems as amd's chiplet design lower yield less scalability and higher cost but team blue isn't going to be left behind as the industry moves towards chiplets except they call them tiles instead because they just had to be different in 2023 we're expecting a lineup of mainstream processors called meteor lake based on tiles connected by something called emibs which unfortunately is a way less cool name than infinity fabric in fact we've already seen emibs in the kbla g processors from a few years ago which used emibs to connect the built-in radeon gpu to its hbm memory as well as in intel's ponte vecchio supercomputer chips intel thinks it's a more elegant solution to amds even going so far as to drop a not so subtle star wars reference to it on its website and to be fair to them they might be right instead of having a big old slab of silicon that the chiplets all sit on an emib is a much smaller piece of silicon that connects tiles on either side not only do you need less silicon which keeps costs lower but it's also even easier to scale instead of being limited by the size of your expensive silicon under layer you can just use multiple e-mibs to build outward it's a little bit like how the tiles that make up a modern basketball court snap into place and unlike infinity fabric e-mibs move data in parallel giving them higher bandwidth while at the same time offering lower latency with less power loss according to intel anyway but there are still probably limits on just how many of these tiles you can put together as a setup that's too large could start hitting a thermal ceiling intel also has another method to connect chiplets called foveros which involves stacking chips on top of each other but this hasn't seen much adoption in the consumer space yet only a couple of laptops that were released in 2020 feature photorose chips and that specific lineup called lakefield was discontinued due to poor demand however the idea of chip stacking is something that intel seems committed to for its future but even though both intel and amd are trying to break up the classic monolithic chip design don't expect it to go away completely these fancy methods of linking chipless together do cost money so at the lower end of the market simpler monolithic designs for systems that don't need tons of computing muscle will be with us for quite some time so if you find yourself in the market for something cheaper but you really want a chiplet there's always those broken pringles at the bottom of the can thanks for watching guys like dislike check out some of our other videos maybe comment with video suggestions and don't forget to subscribe and followas much as amd and intel fanboys like to scream at each other online the average user would be hard-pressed to notice the difference between these companies cpus in everyday use unless you're like us and benchmark things religiously but on the inside modern chips from team red and team blue are built quite differently so let's take a look at how starting with amd ryzen cpus are made up of at least one core complex or ccx but don't worry despite the name it's not a complex thing to understand a ccx is just a group of cores along with some cache memory newer zen3 chips have cc axes with 8 cores while zen 2 cpus have 4 cores per ccx each ccx lives on a chiplet called a ccd the setup on zen 3 is pretty straightforward one ccx per ccd with up to 2 ccds per processor for a total of 16 cores on zen 2 each ccd can instead hold two of these smaller ccx's resulting in the same 16 core maximum but what links all these core complexes together well you've probably heard of amd's infinity fabric which is their secret sauce bus that can run between ccx's or between ccds although it's a high high-speed interconnect that works pretty well with ryzen beating out competing intel chips and a good number of benchmarks it's still not as fast as having the cores more directly connected as there's higher latency with infinity fabric and the connection is serial rather than parallel but there are real benefits to this approach one is that the smaller chiplets mean big cost savings due to better yield if one chiplet is bad you can discard just that one chiplet without wasting as much silicon from throwing away a whole cpu this ccx plus infinity fabric approach also scales quite well so you can just add more chiplets if you need more cores instead of spending more time and money designing a whole new chip is that how intel does it we'll tell you right after we thank brilliant for sponsoring this video brilliant is a website and app built around active learning that's accessible and fun trade boring long lectures for problem solving and interactive visuals there's over 60 courses to choose from including their new everyday math course where you can get a refresher on foundational math topics you might not have touched since high school join the community of 10 million learners and educators today the first 200 people who head to brilliant.org techwiki will get a 20 off an annual premium subscription so intel in contrast to amd still uses monolithic chip designs where cores are right next to each other at least for the time being this approach allows the cores to talk with each other more directly but it has the opposite problems as amd's chiplet design lower yield less scalability and higher cost but team blue isn't going to be left behind as the industry moves towards chiplets except they call them tiles instead because they just had to be different in 2023 we're expecting a lineup of mainstream processors called meteor lake based on tiles connected by something called emibs which unfortunately is a way less cool name than infinity fabric in fact we've already seen emibs in the kbla g processors from a few years ago which used emibs to connect the built-in radeon gpu to its hbm memory as well as in intel's ponte vecchio supercomputer chips intel thinks it's a more elegant solution to amds even going so far as to drop a not so subtle star wars reference to it on its website and to be fair to them they might be right instead of having a big old slab of silicon that the chiplets all sit on an emib is a much smaller piece of silicon that connects tiles on either side not only do you need less silicon which keeps costs lower but it's also even easier to scale instead of being limited by the size of your expensive silicon under layer you can just use multiple e-mibs to build outward it's a little bit like how the tiles that make up a modern basketball court snap into place and unlike infinity fabric e-mibs move data in parallel giving them higher bandwidth while at the same time offering lower latency with less power loss according to intel anyway but there are still probably limits on just how many of these tiles you can put together as a setup that's too large could start hitting a thermal ceiling intel also has another method to connect chiplets called foveros which involves stacking chips on top of each other but this hasn't seen much adoption in the consumer space yet only a couple of laptops that were released in 2020 feature photorose chips and that specific lineup called lakefield was discontinued due to poor demand however the idea of chip stacking is something that intel seems committed to for its future but even though both intel and amd are trying to break up the classic monolithic chip design don't expect it to go away completely these fancy methods of linking chipless together do cost money so at the lower end of the market simpler monolithic designs for systems that don't need tons of computing muscle will be with us for quite some time so if you find yourself in the market for something cheaper but you really want a chiplet there's always those broken pringles at the bottom of the can thanks for watching guys like dislike check out some of our other videos maybe comment with video suggestions and don't forget to subscribe and follow\n"