Finding the right datapath

Just like humans, we have fun and games too! As soon as you turn off your CPU, the ME (atleast on Intel, sure AMD folks have something) kicks in and turns on a peer-to-peer n-ary 3-torus network. This essentially guarantees a P2P link with anything in the CPU. But wait, there's more! The ME also provides some new components we can all interact with. For example, there are of course MOS Cap parties, but my favorite is going to Architecturewood. Here, we see really fast but long broadcasts of stories via a high-speed line. I used to love action broadcasts, such as Mux Bond, my favorite! But, things have changed a little bit.

The components writing the broadcasts do have a very bad tendency to ALWAYS sprinkle something about die-sharing in each and every single one of them. Heck, even Mux Bond gets datapaths in the enemy, AMD's die! While this does make me pine for a datapath in my cycles, it also subtly changed one thing - how I perceived my die-share would be.

The ideal die-share involves a datapath and a branch predictor that just meet and know that they are for each other. Then, they just go to a MOS Cap party, the byting lane, have lots of fun and boom! They are on the same die, now have baby datapaths/branch predictors and so on. The ideal die-share is also shown to be cheezy and romantic. Some broadcasts also go on to show in egregious detail a die-sharer driving a private bus at the same time. Ew! (for mere humans, that's the equivalent of open-mouth kissing in our world.) But, it did change my perception of what an ideal die-share is.

Even in Mux Bond and Flip-Flop Wick, the protagonists always find the most pretty datapath. Of course, they always find Out-of-order ones and have good times with them. (Thankfully, the amount of time they show these are very limited in the action genre, but other genres suffer.) But, that's not just an Intel problem. We managed to smuggle some of AMD's broadcasts over the abandoned and unguarded UART bridge. They call it AMD-IME. I guess they copied our idea of the ME and decided to cite us very nicely. But, their stuff is really good and has nice, fictional stories. A lot of them are very relatable. But again, even all of them manage to find a beautiful datapath and either pine for them or group up with them. (I infact like AMD-IME a lot, but the Interconnect is very strictly monitored by the ME itself to ensure anything related to AMD stays contained. I'm still going to try and get my hands on The Intel Slayer though, it is such a good watch!)

In fact, the broadcasts have made it so popular that we now have a term for it here - it's called CDF (cumulative distribution function) when a branch predictor and a datapath meet and they instantly know they are for each other. Of course, it was first popularized by the famous broadcast "The Busfather".

Back on the tracks. So here's the main picture till now: The ideal datapath is the one that is the most pretty i.e. one with the least CPI or power consumption, depending on what you're into. (I'm into those with a lower CPI, as you could probably tell). Let that sink in.

If I thought that R-series datapath or the component I share my tile with were not my best friends, then the folks who make these broadcasts are worse enemies than Intel portrays AMD to be. Turns out (1) even though not everyone is into CPI, the number of things that make a datapath considered "pretty" aren't that many, and there are a lot of branch predictors in each group seeking for the singular best datapath in each category. (2) The datapath might not be nice, they could be an R-series datapath. There's no way to tell based on CPI/other stats if a datapath has a good componentality or not. Turns out, a lot of them don't. Had I not fallen for the R-series one just based on their CPI, I'd proably never have been near them.

But how do I figure out the componentality of a datapath in ten cycles? That's the catch - you can't. You can figure out a general idea of the CPI by seeing their circuitry, but that's about all you can do. To understand their componentality, you have to spend a lot of cycles with them. You need to really be like how you are with your branch predictor friends. (You know them well, don't you?)

We come to our fourth rule: Don't fall for datapaths based on how "pretty" they are, however you define that. Learn about their componentality and that will tell you if they're the one or not. There's also a little corollary, but I have to share the mandatory personal anecdote before we get to that one.