Discussion A possible point in future where Intel will no longer make Core i3.

[whm1974]

I think that Spectre or Meltdown (I can never keep them straight) will eventually spell the end to all hyperthreading.

Both of these vulnerabilities go all the back to first Pentium, or even the 486.

 

[ehume]

Both of these vulnerabilities go all the back to first Pentium, or even the 486.

Even for non-hyperthreaded CPU's? I had read that the threat in part rises from HT/SMT.

 

[Topweasel]

Even for non-hyperthreaded CPU's? I had read that the threat in part rises from HT/SMT.

Some of the later ones discovered a basically an offshoot where HT basically makes invalidates other patch work. But the core of what Spectre and Meltdown are, are from Intel's implementation of OoOE (Out of Order Execution). So this goes back nearly 30 years.

 

[ehume]

Some of the later ones discovered a basically an offshoot where HT basically makes invalidates other patch work. But the core of what Spectre and Meltdown are, are from Intel's implementation of OoOE (Out of Order Execution). So this goes back nearly 30 years.

So OOE is the culprit, not HT. Whew. I had thought that the demise of HT was coming, that i5 and 9700K were our future.

 

[Topweasel]

So OOE is the culprit, not HT. Whew. I had thought that the demise of HT was coming, that i5 and 9700K were our future.

It still could be. In the strictness sense HT isn't SMT as we know of it from everyone else, Intel has HT handled in windows as kind of a backup the have Windows recognize which core is a real core and which is Logical. They did this a long time ago because their implementation really can't handle 2 threads interweaving that well and almost across the board HT lead to reduced performance. So MS patched windows to use that to fill in the gap, this helped alleviate the windows overhead and small tasks interrupting programs but not all threads are created equal. The legacy of that was still in windows in 2017 when Ryzen hit and MS had to patch the scheduler again to get AMD's SMT implementation working by doing what they did originally, which is treat both cores equally and schedule as needed. It's one of the reason AMD's SMT is a lot more effective (30% boost vs. 10%).

Now add the core race. Extra Logical cores aren't as useful as core count goes up. This is made worse by worse by how HT works. Since the logical threads are only assigned jobs in an emergency when physical cores usage is high enough they get used less and less as core count rises. Then look at the Data center usage. 2 things that they do the most is either VM hosting or heavy scientific processing. The Later is going to peg the physical cores at 100% so little room for extra little slices to fit in and in settings where those really don't exist. The former you take a legitimate chance of some VM's impacting others by one being forced to utilize the logical core of another's physical core and it would run like poo anyways because HT wasn't really made for that. When it comes to web hosting, most don't even have it turned on for that very reason, the cores you purchase are cores.

So you have an outdated tech that's usefulness is diminishing, that isn't being used in some of your biggest markets, and is now in this VM hosting world has become a big security hole. They have 3 choices, one stick with it anyways, 2 decide that it's outlived its welcome and use those transistors for something better, or 3 redesign HT to be a more full blown SMT implementation.

We have to remember why Intel developed it in the first place. Windows back in the day had horrible issues with thread lock. Anything that used your CPU basically locked the system out. Simple IO access from windows would momentarily lockup games. Antivirus programs created a perpetual motion performance crashes scanning files. Any of these could lead to a full system lockup and crash. That's what HT was originally meant for. So I think there is a decent chance that as Intel develops new CPU's specially ones that aren't just tweaked Sandybridge, HT could go away. If anything it's the need of the feature on low core count laptops that will keep it alive.

 

[whm1974]

@Topweasel I was thinking that as core counts scale up, there would be less need for SMT, if any at all. What are main issues with having and using SMT?

 

[Topweasel]

Mainly that it's either nearly useless like with folding and other scientific research where the all cores are pegged against the wall or on VMs where one of SMT threads can impede on another servers CPU usage. The last is probably a bigger issue with AMD SMT implementation. Overall the Core gets more accomplished but since it's thread agnostic but neither thread is going to be working at full speed. Not exactly what cloud guys are paying for. For normal business VMs though they can still be useful if you don't over provision to much.

 

[Amol S.]

Mainly that it's either nearly useless like with folding and other scientific research where the all cores are pegged against the wall or on VMs where one of SMT threads can impede on another servers CPU usage. The last is probably a bigger issue with AMD SMT implementation. Overall the Core gets more accomplished but since it's thread agnostic but neither thread is going to be working at full speed. Not exactly what cloud guys are paying for. For normal business VMs though they can still be useful if you don't over provision to much.

There actually are a few applications still out there, that do not do rendering and place all graphics streaming into just a single core(CPU), plus the programs own need for CPU is handled as well. The surprising fact is that, these programs actually do enough calculations that it causes modern day CPU's of laptops to increase fan speed.

Doing some looking around, the best and most popular example for this that I could find is Old School Runescape MMORPG. Basically it does not use integrated GPU rendering even, and use CPU rendering. The developers are not even working towards making it rendering based. [1] <---Game developers website discussing the fact that they do not support rendering. [2] <---Where I got the CPU based info of no renderer or GPU is used.

But then again, this is just only a single, out of the multiple, CPU's in modern day processors.

 

[TheELF]

It still could be. In the strictness sense HT isn't SMT as we know of it from everyone else, Intel has HT handled in windows as kind of a backup the have Windows recognize which core is a real core and which is Logical. They did this a long time ago because their implementation really can't handle 2 threads interweaving that well and almost across the board HT lead to reduced performance. So MS patched windows to use that to fill in the gap, this helped alleviate the windows overhead and small tasks interrupting programs but not all threads are created equal. The legacy of that was still in windows in 2017 when Ryzen hit and MS had to patch the scheduler again to get AMD's SMT implementation working by doing what they did originally, which is treat both cores equally and schedule as needed. It's one of the reason AMD's SMT is a lot more effective (30% boost vs. 10%).

No there has been nothing in windows to recognize which core is a real core and which is Logical you can download and use coreprio that was made for threadrippers on intel CPUs and see that this is not so,corepeio will make windows use the real cores first and only then the HTT ones so you can see the difference,windows doesn't do that.
Also HTT gives you up to 100% improvement ,where did you get the joke numbers of 10 and 30 from?
The fewer instructions the threads use the more instructions are left over for the HTT SMT thread,if they use few enough you get 100% speed out of both threads.

Now add the core race. Extra Logical cores aren't as useful as core count goes up. This is made worse by worse by how HT works. Since the logical threads are only assigned jobs in an emergency when physical cores usage is high enough they get used less and less as core count rises. Then look at the Data center usage. 2 things that they do the most is either VM hosting or heavy scientific processing. The Later is going to peg the physical cores at 100% so little room for extra little slices to fit in and in settings where those really don't exist. The former you take a legitimate chance of some VM's impacting others by one being forced to utilize the logical core of another's physical core and it would run like poo anyways because HT wasn't really made for that. When it comes to web hosting, most don't even have it turned on for that very reason, the cores you purchase are cores.

Windows sees all cores the same way as I already said and hardware VT(virtualization) has stopped VM from using 100% of the CPU if they don't need it, loading up a VM nowadays is just like loading up any other program there is not much overhead.You can also use task manager to confine VMs or any other software to run only on the cores or threads you want them to run,you are not forced to just go with whatever the defaults are,confine each to one core+HTT or two cores+HTT whatever they need and HTT will be plenty useful.

We have to remember why Intel developed it in the first place. Windows back in the day had horrible issues with thread lock. Anything that used your CPU basically locked the system out. Simple IO access from windows would momentarily lockup games. Antivirus programs created a perpetual motion performance crashes scanning files. Any of these could lead to a full system lockup and crash. That's what HT was originally meant for. So I think there is a decent chance that as Intel develops new CPU's specially ones that aren't just tweaked Sandybridge, HT could go away. If anything it's the need of the feature on low core count laptops that will keep it alive.

Yes Windows back in the day had horrible issues with thread lock because this is an OS problem as windows got better in using time slices and and task manager in general getting better and better these problems faded away from windows anyway.
Since this is a software problem you can still see these problems in console ports where you see games stutter or parts of the game not loading in because they take over control from windows since consoles don't have an OS in the same way that windows is so they screw up and this happens on CPUs with or without HTT or SMT.

 

[ehume]

So are we going to see both Intel and AMD abandon SMT/HT because of the recently discovered exploits?

 

[SarahKerrigan]

So are we going to see both Intel and AMD abandon SMT/HT because of the recently discovered exploits?

I seriously doubt it. SMT, like out-of-order, is enough of a win that it will likely continue to be worth it despite the side effects.

 

[ehume]

I seriously doubt it. SMT, like out-of-order, is enough of a win that it will likely continue to be worth it despite the side effects.

But the big cloud providers may not find that acceptable, so for them, SMT may not be worth keeping. At least for the AMD CPUs, they cram a lot of physical cores onboard.

 

[TheELF]

But the big cloud providers may not find that acceptable, so for them, SMT may not be worth keeping. At least for the AMD CPUs, they cram a lot of physical cores onboard.

No matter how many cores you have HTT/SMT doubles that without the need to double the physical space.
Cloud providers can run software security solutions if they have to, no matter how much performance they loose due to the software part they would still gain more then enough performance to make it worth while.

 

[Topweasel]

No there has been nothing in windows to recognize which core is a real core and which is Logical you can download and use coreprio that was made for threadrippers on intel CPUs and see that this is not so,corepeio will make windows use the real cores first and only then the HTT ones so you can see the difference,windows doesn't do that.
Also HTT gives you up to 100% improvement ,where did you get the joke numbers of 10 and 30 from?
The fewer instructions the threads use the more instructions are left over for the HTT SMT thread,if they use few enough you get 100% speed out of both threads.

There is Intel and Windows identifies the physical and logical cores. On top of that in the Affinity app, Intel specifically has Windows give the logical thread assignments core number assignments for after the physical. For example a 9900k the Core 8 (the 9th assigned core) will be Core O's HT logical core.The percentage is based on averages of tests that fill the physical cores then fill the logical. Something like Folding at Home. Take it's performance with SMT off and SMT on. In theory in a perfect world there would be no difference. But SMT helps the pipeline get better used. Historically HT gets you about 10% extra and AMD's SMT ~30%.

Windows sees all cores the same way as I already said and hardware VT(virtualization) has stopped VM from using 100% of the CPU if they don't need it, loading up a VM nowadays is just like loading up any other program there is not much overhead.You can also use task manager to confine VMs or any other software to run only on the cores or threads you want them to run,you are not forced to just go with whatever the defaults are,confine each to one core+HTT or two cores+HTT whatever they need and HTT will be plenty useful.

Like I said for general Business VM's this isn't an issue. Cloud Server services though beyond Intel's security issues, a lot have been disabling HT anyways because managing computing power intrusions into other customers computing efforts isn't a worthwhile management effort. To much micromanaging to be worth it (even if you are using software tools to handle said management, to many CPU cycles wasted, that could be used for other clients).

Yes Windows back in the day had horrible issues with thread lock because this is an OS problem as windows got better in using time slices and and task manager in general getting better and better these problems faded away from windows anyway.
Since this is a software problem you can still see these problems in console ports where you see games stutter or parts of the game not loading in because they take over control from windows since consoles don't have an OS in the same way that windows is so they screw up and this happens on CPUs with or without HTT or SMT.

Yes. That's what I was getting at. I wasn't denying that the Windows Schedular has gotten better. But HT was created as a fix for a software issue. I didn't mention consoles and is kind of pointless. Games are written for bare metal so it just always expects the resources to be available. When its not it goes crazy.

No matter how many cores you have HTT/SMT doubles that without the need to double the physical space.
Cloud providers can run software security solutions if they have to, no matter how much performance they loose due to the software part they would still gain more then enough performance to make it worth while.

HT doesn't double anything. It allows extra threads to be assigned without having to flush the previous thread from memory and wait for the pipeline to clear. The advantage is if certain parts of the core are unutilized it can more work done.

 

[TheELF]

There is Intel and Windows identifies the physical and logical cores. On top of that in the Affinity app, Intel specifically has Windows give the logical thread assignments core number assignments for after the physical. For example a 9900k the Core 8 (the 9th assigned core) will be Core O's HT logical core.

That's AMD,intel goes one real one logical.
I gave you a hands on practical test that will show you that windows does not do what you think it does,if you choose to ignore this that's your problem.

The percentage is based on averages of tests that fill the physical cores then fill the logical. Something like Folding at Home. Take it's performance with SMT off and SMT on. In theory in a perfect world there would be no difference.

HTT/SMT where not made to help with perfect or very close to perfect software,how crazy would a company be to put all of this money into research for only 10% performance...
HTT/SMT are made to use the otherwise unused/idle instructions so obviously it;s made for software that leaves instructions idle and not for software that is nearly perfect and does not leave any instructions idle.

But SMT helps the pipeline get better used. Historically HT gets you about 10% extra and AMD's SMT ~30%.

So what you are really saying here is that AMDs cores are so much worse at keeping that pipeline fed.

Like I said for general Business VM's this isn't an issue. Cloud Server services though beyond Intel's security issues, a lot have been disabling HT anyways because managing computing power intrusions into other customers computing efforts isn't a worthwhile management effort. To much micromanaging to be worth it (even if you are using software tools to handle said management, to many CPU cycles wasted, that could be used for other clients).

Yeah explain to us again how this would actually ever even happen?!
If you rent some cloud computing or cloud VM or whatever you get a set number of threads and these threads are yours alone, there is no intrusion from nowhere.

HT doesn't double anything. It allows extra threads to be assigned without having to flush the previous thread from memory and wait for the pipeline to clear. The advantage is if certain parts of the core are unutilized it can more work done.

Look at the video,if you choose to ignore any evidence then why are you even arguing?
Also even if it just allows extra threads to be assigned that's a gain right there,you get twice the assigned threads at the same time,that's still double.

 

[mxnerd]

Even Atom CPU is still in use and OP think i3 will be out of market? How about Celeron and Pentium?

 

[maddie]

April fool's day is only ONE day and it's past for this year. Why am I reading that HT gives 100% improvement and about 'perfect software', whatever creature that is.

 

[escrow4]

OP doesn't understand economics. Intel, like any company that makes $$$$$$$$$$$$$, has a product stack from top to bottom. Even the Iphone has a "cheap" (well cheaper) model now - the XR. If you don't want or can't afford an i5 you drop lower otherwise you'd piss off to AMD. The i3 is a brand that will not be going anywhere, same with the Pentium. I also cannot give up a desktop simply because its the easiest computing device - you can't have a dozen tabs in Chrome with Filezilla and 4 other apps open switching between on a tablet, and I don't laptop anymore (or game on a PC).

 

[whm1974]

OP doesn't understand economics. Intel, like any company that makes $$$$$$$$$$$$$, has a product stack from top to bottom. Even the Iphone has a "cheap" (well cheaper) model now - the XR. If you don't want or can't afford an i5 you drop lower otherwise you'd piss off to AMD. The i3 is a brand that will not be going anywhere, same with the Pentium. I also cannot give up a desktop simply because its the easiest computing device - you can't have a dozen tabs in Chrome with Filezilla and 4 other apps open switching between on a tablet, and I don't laptop anymore (or game on a PC).

This+ While for 2019 I wouldn't consider anything less then 4 core/4 threads for a new system, the fact is, many people will do fine with a 2c/4t processor, and most folks(that I know) can't afford higher end CPUs anyway.

 

[NostaSeronx]

I expect more SKUs.

iF
iD
iB
i9
i7
i5
i3
i1
Pentium Platinum w/ mixed core assortment.
Pentium Gold w/ mixed core assortment.
Pentium Silver and Bronze, etc.
Celeron maybe add levels like some Iron, Steel, Titanium, Carbon
Atom (64-bit optimized)
Quark (32-bit optimized)
Preon (16-bit optimized)
Then, there is the Quantum lineup;
q7
q5
q3
q1

etc. So many SKUs, it is a maze. Then, there are suffixes for L4, Near HBM, Discrete GPU in/on Package, etc.

 

[ehume]

I expect more SKUs.

iF
iD
iB
i9
i7
i5
i3
i1
Pentium Platinum w/ mixed core assortment.
Pentium Gold w/ mixed core assortment.
Pentium Silver and Bronze, etc.
Celeron maybe add levels like some Iron, Steel, Titanium, Carbon
Atom (64-bit optimized)
Quark (32-bit optimized)
Preon (16-bit optimized)
Then, there is the Quantum lineup;
q7
q5
q3
q1

etc. So many SKUs, it is a maze. Then, there are suffixes for L4, Near HBM, Discrete GPU in/on Package, etc.

/sarcasm

 

[TheELF]

April fool's day is only ONE day and it's past for this year. Why am I reading that HT gives 100% improvement and about 'perfect software', whatever creature that is.

Dude, do you even English bro?

 

[coercitiv]

Dude, do you even English bro?

It's not just the English that matters, clarity and brevity are also paramount. Some structure would help too, but let's not ask for everything at once.

 

[DrMrLordX]

let's not ask for everything at once.

I, for one, would like to start with a partridge in a pear tree. No idea why I want those things, but hey, seems like a good place to start. The partridge can have an i3. It does not need more than four cores.

 

[coercitiv]

I, for one, would like to start with a partridge in a pear tree. No idea why I want those things, but hey, seems like a good place to start. The partridge can have an i3. It does not need more than four cores.

This partridge and i3 silicon mix reminds me of my favorite movie:

Rachael: Do you like our owl?
Deckard: It's artificial?
Rachael: Of course it is.
Deckard: Must be expensive.
Rachael: Very. I'm Rachael.

Let's hope for a somewhat distant future of affordable i3 powered artificial partridges.