edit, please just read replies, this topic has changed. a lot, very fast.
This already exists and has nothing to do with SMT or why itās useful?
Minus the 7ghz boost. We canāt reasonably hit those clockspeeds on a single core, SMT or no, because physics. These thoughts you have, have been had by the important people, and are ubiquitous in modern CPUs. I know I see my 5950x at anywhere from 1.7ghz to 4.9ghz. 960T could be anywhere from 0.8 to 4ghz. 4690k, 0.8 to 3.9ghz. I see mobile processors get as low as 50mhz on parts that can boost as high as 2-3ghz.
SMT/HyperThreading is just a way to run two or more commands on a single core at the same time, because thereās a lot of the core being left unused at any given time otherwise.
Could you give some more details of your idea, Iām not getting the full picture of what youāre trying to say.
Though you need to keep in mind that modern CPUs are too easily stalled. Simple events like branch prediction failures or reading memory that isnāt cached will cause 10s-100s of cycles of stalling. SMT/HT is one nice technique to prevent this issue, allowing one thread to progress while another is stalled.
SMT/HT is also a nice way of adding more compute power to a CPU without needing a full core, making the CPU cheaper.
I donāt see how adding more FPU to a CPU is going to prevent task stalling, and remove the benefits of SMT/HT. Here is a rough flowchart of operations in a Zen 1 core (Source), there is far more in there than just an FPU.
you got hung up on the speed. [X]-Ghz to [Y]-Ghz is not the important thing, itās how fast each core can change that speed.
alsoā¦ I had no idea that maybe it could be possible to have both.
I was going on the generally excepted āfactā (note no one ever told me this was true, or ever told me it was fake, I just heard it being said) that: āDisable smt to overclock really fast on ln2ā *I have no idea if that is true.
bottom line: my idea is: Change the Ghz of a CPU faster per task to tighten the ātimingā on a CPU more like a RAM module and I thought (besides having no idea to leave smt on for this) that a FPU in the FSB could do this on a hardware level faster than eating up instructions on the CPU cycle to control something like this.
no no no, you are hung up on the wrong details both of you are, itās about how FAST you can change the clock speed.
Also, turns out, you can leave smt/ht alone while doing thisā¦ I had no idea you could do both.
Yeah, the thing is, Iām saying, they already do that. They change frequency these days many many times per second. Zen3 can change CPU frequency 500 times per second at stock settings. Iād be surprised if Zen4 wasnāt at least that fast at changing clocks.
The modest gains to overclocking headroom by disabling SMT arenāt for improving performance, or everyone would already do that on air, because CPUs already boost as high as they can go within their target power, and thereās nothing different about performance scaling at differing temperatures.
Itās the same as disabling extra cores, just so you can put more power budget into a single core, to hit a higher target frequency without crashing, and so you can claim that your mhz was higher than the other guyās mhz, not even that you got more arbitrary performance metric points in a given software, just that technically the number of clock cycles was higher and it didnāt crash.
This really is already the norm. Some engineers thought of this probably close to half a century ago, and itās been implemented and improved upon ever since. Your phone couldnāt function without it.
Like, I donāt want to sound rude or anything, itās cool that you thought of that on your own, and you should continue thinking about things and learn more about processors and stuff, but someone else beat you to this one by quite a long time.
Also, thatās not really how RAM works, but RAM is also very interesting stuff, so you should probably look into that and learn interesting things about it as well.
I recommend this video series on it.
1 Like
so, your saying Iām only as dumb as the smart people who came up with this first?
Iāll accept that.
thanks?
I donāt remember you talking about that in your original post. Itād be helpful if you could state your ideas in a more concrete way, itās hard to follow your thoughts and what conclusion youāre trying to reach.
If youāre talking about changing power state as fast as possible (since frequencies are direct consequences of power states), Intel CPUs have been controlling p-states, c-states and frequency from the CPU for quite a few generations (since skylake?), so this kind of stuff is already going on. On Intel CPUs itās called HWP.
Documentation for this is a little slim, this stuff might help:
https://docs.kernel.org/admin-guide/pm/intel_pstate.html
https://docs.kernel.org/admin-guide/pm/amd-pstate.html
With CPU/GPU design everything is a compromise on top of a compromise. (Just look at AMD Bulldozer or Intel Arc). Making one part faster may improve some workloads, but not others.
Chuckle and tell you to google āthc-p stateā you can leave safe mode on, this is not that kind of joke, itās clean, but still funny.
(only a joke, Iām just in shock by your answer, do you know how odd this is that no one talks about what CPU-P-STATE Controll is, sure EVERYONE tells you it exists for some āoverclockā guide. but never why we should care! What? it does that???)
Mobile computers: Phones, Tablets, laptops, have all pushed the power saving states forwards, and desktops have benefitted from it.
Even these āEfficiency coresā is like the big-little popularised by mobile (even if some desktops had them a while in niche)
Oh ha, I didnāt see that one. Reminds me of innocent programmer grads searching for āC Stringā on the internet, and getting undergarments instead of programming resources.
There are so many details that go into a modern computer, but theyāre all hidden behind abstractions. If you want your mind blown you should watch Ben Eater build computers from scratch. He builds a basic GPU, and talks to a USB keyboard, all with bread boards and very simple chips.
This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.