Why didn't intel enable "hyperthreading" they would have been pretty competitive with AMD

[redbread123]

Title, they could have been really competitive with AMD still, intel would be better in games and still not lack cores/threads. why did they do this? heat isn't a problem most cpu coolers keep the cpu under 80 under full load.

[Mira Yurizaki]

If you're talking about CPUs that don't have it, it's mostly to create a market segment.

 

Otherwise, HyperThreading, or more generally speaking, simultaneous multithreading (SMT) is a hit and miss feature. There's very few use cases where SMT would cause a dramatic performance improvement and in some cases, may hinder performance.

[redbread123]

2 minutes ago, Mira Yurizaki said:

If you're talking about CPUs that don't have it, it's mostly to create a market segment.

 

Otherwise, HyperThreading, or more generally speaking, simultaneous multithreading (SMT) is a hit and miss feature. There's very few use cases where SMT would cause a dramatic performance improvement and in some cases, may hinder performance.

yeah but all the amd fans keep saying intel doesn't have smt, atleast they would be on the spec sheet, on par with AMD

[Suppersupport]

meltdown and spectre was a big reason for some I know.

[redbread123]

Just now, will4623 said:

meltdown and spectre was a big reason for some I know.

what are those.

[Mira Yurizaki]

5 minutes ago, redbread123 said:

yeah but all the amd fans keep saying intel doesn't have smt, atleast they would be on the spec sheet, on par with AMD

Those AMD fans probably don't understand that HyperThreading is Intel's implementation of SMT. But again, situations where SMT shine are few and far between and having it on or off really doesn't matter most of the time: https://www.techspot.com/review/1882-ryzen-9-smt-on-vs-off/

 

2 minutes ago, redbread123 said:

what are those.

Security issues that were found in processors.

 

Meltdown and Spectre have nothing to do with SMT. They affected speculative execution which is a core feature of modern processors.

[mariushm]

Intel processors have a lot of bugs and issues related to hyperthreading ... meltdown and spectre are the names of a few sets of vulnerabities discovered in Intel processors. For example, a 'bug' could allow an application running on one core to just 'listen' and silently collect bits of information from applications running on other cores or threads, which is something normally not allowed by operating systems.  So an application could sit and listen and catch passwords or database records and send them away, for example.

They made fixes for a lot of these issues which are "uploaded" into the processor through bios updates and through operating system upgrades but these fixes slow down the cpu by some amount.

 

For some of the new processors, rather than have their programmers fix the code stored in the processor or change the actual silicon (add or remove transistors) which takes months of manufacturing and testing silicon, they just disable hyperthreading.

 

 

[redbread123]

1 minute ago, mariushm said:

Intel processors have a lot of bugs and issues related to hyperthreading ... meltdown and spectre are the names of a few sets of vulnerabities discovered in Intel processors. For example, a 'bug' could allow an application running on one core to just 'listen' and silently collect bits of information from applications running on other cores or threads, which is something normally not allowed by operating systems.  So an application could sit and listen and catch passwords or database records and send them away, for example.

They made fixes for a lot of these issues which are "uploaded" into the processor through bios updates and through operating system upgrades but these fixes slow down the cpu by some amount.

 

For some of the new processors, rather than have their programmers fix the code stored in the processor or change the actual silicon (add or remove transistors) which takes months of manufacturing and testing silicon, they just disable hyperthreading.

 

 

oh

[Mira Yurizaki]

HyperThreading flaws are related to issues with SMT, which are separate from speculative execution issues that Meltdown and Spectre attack

 

https://arstechnica.com/information-technology/2018/11/intel-cpus-fall-to-new-hyperthreading-exploit-that-pilfers-crypto-keys/

https://en.wikipedia.org/wiki/Hyper-threading#Security

 

Meltdown and Spectre are also not specific to Intel

https://en.wikipedia.org/wiki/Meltdown_(security_vulnerability)

https://en.wikipedia.org/wiki/Spectre_(security_vulnerability)

[mariushm]

2 minutes ago, Mira Yurizaki said:

Meltdown and Spectre are also not specific to Intel

https://en.wikipedia.org/wiki/Meltdown_(security_vulnerability)

https://en.wikipedia.org/wiki/Spectre_(security_vulnerability)

Yet AMD processors are not affected by them, or it's practically impossible to take advantage of them.

[Mira Yurizaki]

2 minutes ago, mariushm said:

Yet AMD processors are not affected by them, or it's practically impossible to take advantage of them.

AMD is not affected by Meltdown, but Meltodwn does affect some ARM and POWER processors (hence, not specific to Intel). AMD processors are affected by variants of Spectre. From https://meltdownattack.com/  (emphasis added)

 

Quote

Which systems are affected by Spectre?

Almost every system is affected by Spectre: Desktops, Laptops, Cloud Servers, as well as Smartphones. More specifically, all modern processors capable of keeping many instructions in flight are potentially vulnerable. In particular, we have verified Spectre on Intel, AMD, and ARM processors.

 

[.Apex.]

33 minutes ago, Mira Yurizaki said:

Otherwise, HyperThreading, or more generally speaking, simultaneous multithreading (SMT) is a hit and miss feature. There's very few use cases where SMT would cause a dramatic performance improvement and in some cases, may hinder performance.

22 minutes ago, Mira Yurizaki said:

situations where SMT shine are few and far between and having it on or off really doesn't matter most of the time:

In fact the opposite for Intel, there's only very few cases where HT will hinder performance but in most cases it does improve it, maybe not much in gaming but there's always some improvement regardless and the difference is dramatic in programs that can utilize all of the cores or while multitasking, that anomaly with having better performance with SMT off is only present on AMD CPU's, otherwise the 9700K would be the chart topper but it isn't, the 9900K is.

[redbread123]

9 minutes ago, Mira Yurizaki said:

AMD is not affected by Meltdown, but Meltodwn does affect some ARM and POWER processors (hence, not specific to Intel). AMD processors are affected by variants of Spectre. From https://meltdownattack.com/  (emphasis added)

 

 

what are arm and power, can they not work in a pc

[Mira Yurizaki]

3 minutes ago, _Syn_ said:

In fact the opposite for Intel, there's only very few cases where HT will hinder performance but in most cases it does improve it, maybe not much in gaming but there's always some improvement regardless and the difference is dramatic in programs that can utilize all of the cores or while multitasking, that anomaly with having better performance with SMT off is only present on AMD CPU's, otherwise the 9700K would be the chart topper but it isn't, the 9900K is.

I could've found someone doing more generic tests, but eh, PC gaming is the most encountered use-case on the forum

 

3 minutes ago, redbread123 said:

what are arm and power, can they not work in a pc

ARM and POWER are separate CPU architectures. AMD and Intel use the x86-64 architecture in their CPUs

 

I could've been less confusing with the wording there

[redbread123]

33 minutes ago, Mira Yurizaki said:

I could've found someone doing more generic tests, but eh, PC gaming is the most encountered use-case on the forum

 

ARM and POWER are separate CPU architectures. AMD and Intel use the x86-64 architecture in their CPUs

 

I could've been less confusing with the wording there

by architecture, you mean the puzzle the transistors work in to solve problems?

[Mira Yurizaki]

4 minutes ago, redbread123 said:

by architecture, you mean the puzzle the transistors work in to solve problems?

Kind of. Architecture defines how the software can tell the hardware what to do. So those combination of bits and bytes in software mean something on one architecture, but something totally different (mostly gibberish so it won't run) on another.

 

Or you can think of CPU architecture as the "language" computers speak.

[Bombastinator]

1 hour ago, Mira Yurizaki said:

If you're talking about CPUs that don't have it, it's mostly to create a market segment.

 

Otherwise, HyperThreading, or more generally speaking, simultaneous multithreading (SMT) is a hit and miss feature. There's very few use cases where SMT would cause a dramatic performance improvement and in some cases, may hinder performance.

The cases are quite specific but they are getting a lot more frequent, specifically for games.  My suspicion is they will continue to become more and more frequent.  Is there an actual good reason for it? No.  They’re all bad reasons.  They’re very common ones though and getting more common.

[Mira Yurizaki]

7 minutes ago, Bombastinator said:

The cases are quite specific but they are getting a lot more frequent, specifically for games.  My suspicion is they will continue to become more and more frequent.  Is there an actual good reason for it? No.  They’re all bad reasons.  They’re very common ones though and getting more common.

I don't think people will spend much time, if any, optimizing specifically for SMT. The problem SMT is trying to solve is a thread didn't use all of the CPU's execution units, so it goes and find a thread that could make use of the remaining ones. The problem regarding optimizing specifically for SMT is that the number of execution units is often different between microarchitectures. If you optimized for Skylake, performance won't be as good for say Haswell or Zen.

 

If it just so happens your tasks can scale that well, then it's a happy coincidence.

[Bombastinator]

2 minutes ago, Mira Yurizaki said:

I don't think people will spend much time, if any, optimizing specifically for SMT. The problem SMT is trying to solve is a thread didn't use all of the CPU's execution units, so it goes and find a thread that could make use of the remaining ones. The problem regarding optimizing specifically for SMT is that the number of execution units is often different between microarchitectures. If you optimized for Skylake, performance won't be as good for say Haswell or Zen.

 

If it just so happens your tasks can scale that well, then it's a happy coincidence.

There’s another problem though: what if the app was written with a minimum thread number larger than the number of thread spaces available?  If that happens cueing occurs and overhead builds.  Overhead that thread programming is very very bad at dealing with.

[Mira Yurizaki]

2 minutes ago, Bombastinator said:

There’s another problem though: what if the app was written with a minimum thread number larger than the number of thread spaces available?  If that happens cueing occurs and overhead builds.  Overhead that thread programming is very very bad at dealing with.

Many apps that are running are likely using more than a dozen threads. Not every thread is running at the same time. If you have an application that's compute bound, then sure, you should limit the number of worker threads to the number of threads available in the system.

[Underi]

I don't think the modern i5 processors need hyper-threading. Mine already has 6 cores which give enough juice for multi-core intensive workloads, while keeping the single core performance high, which is what matters in games. That's why it can perform on par with AMD's 3600 while being two years older. Intel doesn't need the hyper-threading in lower-end processors because it makes up with it's single core performance. If you want to build a budget workstation (and maybe even an expensive one), the obvious choice is AMD.

[Princess Luna]

4 minutes ago, Underi said:

Intel doesn't need the hyper-threading in lower-end processors because it makes up with it's single core performance

You're justifying the artificial product segmentation? Intel should make HT standard already, since Skylake it no longer depends on silicon purity to be enabled.

 

They are trying to overcharge you for a feature that's present on every other present day CPU, AMD brought it to their entire line up already and so should Intel specially at the current condition where they are getting behind.

 

Qualcomm, Apple and any other CPU out there on multiple devices already counts with some sort of multi-threading capacity per core, Intel is the only one behind in their greed to profit as much as possible.

[Mira Yurizaki]

4 minutes ago, Princess Luna said:

You're justifying the artificial product segmentation? Intel should make HT standard already, since Skylake it no longer depends on silicon purity to be enabled.

And what if i5s are just rejected i7s that couldn't pass muster?

 

If you're going to argue that Intel should do better QA, then go tell AMD to do a better job on their ends too because they have plenty of products that are basically "the same silicon as the higher end variant, but less feature rich"

[Kisai]

2 hours ago, redbread123 said:

Title, they could have been really competitive with AMD still, intel would be better in games and still not lack cores/threads. why did they do this? heat isn't a problem most cpu coolers keep the cpu under 80 under full load.

SMT doesn't improve the single-thread performance of a CPU, which is what matters for games. Games absolutely do not make use of hyperthreading in any practical way that doesn't reduce the performance of the CPU.

 

Hence, if you get a CPU without hyperthreading, it typically performs better at single-threaded tasks clock for clock. Hyperthreading on Intel CPU's splits one physical core into two virtual cores by splitting the ALU's, but they're still start behind the same part of the pipeline.

 

From P4 era:

 

If Intel simply made it two cores with half the pipeline length in the first place (which is ultimately what happened), then you end up with two cores that have no resource contention.

 

So on i7's with Hyperthreading you get this:

 

 

Same 10-15% improvement over the hyperthreading being disabled.

 

Basically if a program would benefit from many less-pipelined cores, then Hyperthreading will always have a performance boost over non-hyperthreaded loads. Yet, as stated, a higher single-threaded performance is always preferable to hyperthreading, and when all the spectre/meltdown/zombieload stuff came out, guess what took the biggest hit?

 

 

 

So ultimately, all Intel did was sow confusion by adding cores to processors but not changing the part number to indicate so.

 

The original i7-970, had 6 cores, 12 threads. The i7-960 had 4 cores, 8 threads.

The i7-2700k had 4 cores, 8 threads. i7-3770k, 4 cores, 8 threads.

i7-4770/k 4 cores, 8 threads

i7-6700/k 4 cores, 8 threads

i7-7700/k 4 cores, 8 threads

 

and then suddenly Intel adds an i9 tier

i7-8700/k 6 cores, 12 threads, vs i9-8950HK also with 6 cores and 12 threads

i7-9700/k 8 cores. No HT, vs i9 -9900/k with 8 cores and 16 threads.

 

 

So basically Intel pulled a "Grocery Shrink ray" 

Yes, ultimately the i9 part with hyperthreading is no faster than the i7 part without hyper threading (higher turbo clock applies to single core 2.6%), it still has a 14.7% gain from the hyper threading. But here's the fun part. 2.6% x 8 cores is 20.53%, so that entire hyperthreading improvement could have come from the cpu speed alone.

 

 

Looks like Hyperthreading on the 9900k actually makes little difference, even under synthetic bechmarks. We might see Intel drop hyperthreading completely at some point, because the security benefit outweighs the 9% performance gain, when a 30% performance drop comes from making it secure.

[Princess Luna]

3 minutes ago, Mira Yurizaki said:

And what if i5s are just rejected i7s that couldn't pass muster?

 

If you're going to argue that Intel should do better QA, then go tell AMD to do a better job on their ends too because they have plenty of products that are basically "the same silicon as the higher end variant, but less feature rich"

Crippling the CPU on purpose from a feature that's universally adopted in order to profit more from it is not the same as having to downgrade the CPU tier because silicon is unable to achieve the needed specs.

 

i5's and i7's both should have HT because to Intel it's literally a switch already to turn it on, so you'd have 6c/12t and 8c/16t parts, the difference is in core count as it should be and not as Intel done for ages with their quadcores where HT was the only difference when all those i5's could have it enable.

 

Cache and cores and out of the box frequency is already a great way to do product segmentation since it makes sense as not all silicon can achieve the same specs however SMT and HT both are no longer depended of silicon, it's like stated a switch to turn on and off at hardware level.

 

AMD is doing a much better job.