Intel CEO Confirms SMT To Return to Future CPUs

[Panther_Seraphin]

So reading the 1st post here is my take

• Datacenter - We have NO way of competing in the short/mid term against AMD and ARM without the use of SMT. The security concerns around SMT are outweighed by the performance deficit we would be left in without it.
• Client - No changes in inital architecture/plans but that leads to
• Longterm plans - New architectures are required and we cant just keep iterating on current ones. ALL new designs have to be reviewed and approved by me before going to actual hardware testing.

My guess is regarding the longterm plans is to try and getaway from the multi platform offerings Intel seems to have a thing for in the datacenter. Also I would suspect there has been a bit too much freedom between the different teams to suit their own needs (Datacenter/Client) with little to no co-ordination causing design decisions that have hurt them.

Hyper-threading is a crutch and is usually there to maximize profits by allowing for simpler and smaller cores

Which is wrong as HT actually adds a small amount to the die area per core due to the more complicated front end required.

What "job" are you talking about? Be specific. If you look on the broad spectrum of normal PC users, how many people could benefit from it and in what extent?

And you can argue the same point about having a core desinged focusing on any single task load. Having something that is purely focused on Database may not be ideal for someone doing scientific work and the one designed for scientific work will be imbalanced for gaming etc. For example. how many games are really taking advantage of AVX512 instructions and even the lower versions. How impacted would scientific work be in a core without any of those instructions.

SMT was that tool that suited the mantra "perfection is the enemy of good enough". By using SMT they were able to utilise an "imperfect" core design for certain tasks but still perform admirably across a wide breadth of tasks

It's odd how Intel hasn't made GAMING processor with just big cores and crap load of cache. Surely their new approach can just glue whatever together with tiles...

Never going to see it as from a business perspective as it makes 0 sense. Your spinning a completely bespoke design for a market around 2% of the PC market that arent even people who spend big on these parts either. I would love to see it as I can imagine a pure P core design with limited graphics etc similar to AMDs IO Die output would be a really interesting proposition.

Gaming PCs & Laptops - Worldwide | Statista Market Forecast

Worldwide: Revenue in the Gaming PCs & Laptops market worldwide is projected to reach US$62.36bn in 2025. Definition: The market for gaming PCs and laptops encompasses a dynamic range of specialized computing devices tailored to deliver optimal gaming performance.

www.statista.com

 

[Konceptz]

I'll leave this here..

RTX 4090 & 53 Games: Core i9-13900K E-Cores Enabled vs Disabled Review

In this week's TPU50 Megabench we're testing whether you can unlock additional gaming performance by disabling the E-Cores on a Raptor Lake Core i9-13900K processor. For our benchmarks we used the mighty GeForce RTX 4090, with 53 games at three resolutions.

www.techpowerup.com

 

[LittleBro]

Well of course a cpu that was designed with smt in mind will do worse when you disable it, lol. The question is, is it worth the 5-10% extra die space that it requires or would you prefer something else for that die space

It can be calculated. Intel said they first implementation of HT cost 5% of die size and brought 15-30% performance improvement. (Pentium IV)
If you're limited by socket, like AM5 (because Intel would introduce new socket as always), you might come a to point where you can't fit another cores (CCD)
but you may further increase performance by SMT, which will compensate for those cores.
Sure, if you're not limited by space, it's always better to max out count of standard P-Cores.

You seem to be strongly confused.

Thing of opinion.

What "job" are you talking about? Be specific. If you look on the broad spectrum of normal PC users, how many people could benefit from it and in what extent?

Any normal user especially working on Linux can leverage perf. uplift of SMT.

Why are you even mentioning AMD, this thread is about Intel. I own/ed Intel CPUs and I carefully observed how thread utilisation works and what is the benefit of HT and it is exactly how I explained. May I ask you what experience do you have with Intel CPUs?

Holy shit, you mentioned Intel first and now what? I just pointed out that Intel's and AMD's SMT implementation might be different. Did not mean to hurt you.
So you seem to observe things and than make your own opinion on how do things work rather than studying schemes and materials, right?

My experience with Intel ... sure. 1997 - Celeron II 266 MHz, then AMD Duron 800 MHz, next Pentium IV, Pentium D, CentrinoDuo, Core2Duo, Sandy Bridge, Ivy Bridge.

 

[JustBenching]

It can be calculated. Intel said they first implementation of HT cost 5% of die size and brought 15-30% performance improvement. (Pentium IV)
If you're limited by socket, like AM5 (because Intel would introduce new socket as always), you might come a to point where you can't fit another cores (CCD)
but you may further increase performance by SMT, which will compensate for those cores.
Sure, if you're not limited by space, it's always better to max out count of standard P-Cores.


Thing of opinion.


Any normal user especially working on Linux can leverage perf. uplift of SMT.

Holy shit, you mentioned Intel first and now what? I just pointed out that Intel's and AMD's SMT implementation might be different. Did not mean to hurt you.
So you seem to observe things and than make your own opinion on how do things work rather than studying schemes and materials, right?

My experience with Intel ... sure. 1997 - Celeron II 266 MHz, then AMD Duron 800 MHz, next Pentium IV, Pentium D, CentrinoDuo, Core2Duo, Sandy Bridge, Ivy Bridge.

Its way above my paygrade and knowledge to determine if SMT is an overall benefit or not, but i know for a fact that it is detrimenal for gaming and ecores are always better. So if you can save a bunch of space by removing SMT and adding a few ecores instead, im sold. Extra points for removing AVX - if the space savings are big enough you can counteract the performance drop of avx by having more cores. The logic is sound, i have no idea if the numbers (die space saved vs performance drop) check out though.

 

[rattlehead99]

So reading the 1st post here is my take

• Datacenter - We have NO way of competing in the short/mid term against AMD and ARM without the use of SMT. The security concerns around SMT are outweighed by the performance deficit we would be left in without it.
• Client - No changes in inital architecture/plans but that leads to
• Longterm plans - New architectures are required and we cant just keep iterating on current ones. ALL new designs have to be reviewed and approved by me before going to actual hardware testing.

My guess is regarding the longterm plans is to try and getaway from the multi platform offerings Intel seems to have a thing for in the datacenter. Also I would suspect there has been a bit too much freedom between the different teams to suit their own needs (Datacenter/Client) with little to no co-ordination causing design decisions that have hurt them.



Which is wrong as HT actually adds a small amount to the die area per core due to the more complicated front end required.

And you can argue the same point about having a core desinged focusing on any single task load. Having something that is purely focused on Database may not be ideal for someone doing scientific work and the one designed for scientific work will be imbalanced for gaming etc. For example. how many games are really taking advantage of AVX512 instructions and even the lower versions. How impacted would scientific work be in a core without any of those instructions.

SMT was that tool that suited the mantra "perfection is the enemy of good enough". By using SMT they were able to utilise an "imperfect" core design for certain tasks but still perform admirably across a wide breadth of tasks


Never going to see it as from a business perspective as it makes 0 sense. Your spinning a completely bespoke design for a market around 2% of the PC market that arent even people who spend big on these parts either. I would love to see it as I can imagine a pure P core design with limited graphics etc similar to AMDs IO Die output would be a really interesting proposition.

Gaming PCs & Laptops - Worldwide | Statista Market Forecast

Worldwide: Revenue in the Gaming PCs & Laptops market worldwide is projected to reach US$62.36bn in 2025. Definition: The market for gaming PCs and laptops encompasses a dynamic range of specialized computing devices tailored to deliver optimal gaming performance.

www.statista.com

You make the entire core smaller by bottlenecking it to begin with and then you solve it with Hyper-Threading, and you save about 10-20% die area.

 

[AVATARAT]

SMT is not HT and works differently, so don't mix examples with it here.
For server purposes, HT is probably a good thing because it provides better parallelism.
For pure gaming, without HT is better (especially with the current Windows Scheduler), yes, some games use it properly.

 

[Visible Noise]

SMT is not HT and works differently, so don't mix examples with it here.
For server purposes, HT is probably a good thing because it provides better parallelism.
For pure gaming, without HT is better (especially with the current Windows Scheduler), yes, some games use it properly.

They are the EXACT same thing. Hyperthreading is Intels branding of SMT.

 

[Geofrancis]

My understanding is whole point of SMT is to give the core something to do when the main thread is stuck like waiting for memory or IO so the core isn't just sitting idle. that's where the performance gains come from, utilizing compute time that would otherwise be wasted idling, thats also why it doesn't need more power as its just using idle silicone.

 

[Macro Device]

People here talking gaming performance as if everyone here is rocking a 5080+ tier GPU and a 1080p display to truly bottleneck that GPU by any CPU imaginable.

I'd trade some gaming performance for professional performance, efficiency at lighter loads, scheduling optimisations and overall stability. AND SQUARE CPU DESIGN. WHAT IS SO DAMN ATTRACTIVE IN THESE RANDOM RECTANGLES, INTEL?

 

[JustBenching]

thats also why it doesn't need more power as its just using idle silicone.

That's not true. It definitely uses more power at same clockspeeds.

 

[Lycanwolfen]

SMT is beneficial, on AMD platform you get 15% more perf. at no increase in power draw.
Intel was stupid to abandon SMT (HT). Yep, there were multiple security issues present but
instead of abandoning whole SMT technology, they should have make a proper redesign.
Which will they do now. Better late than never, right?

IMHO, SMT is better than e-cores/LP cores, as it:
- does not require additional physical space;
- does not differentiate between instruction sets;
- scheduler issues are much lesser than with multiple types of cores.

By the way, WCCFtech is spreading a rumor that Intel will abandon e-cores and LP cores with generation after Nova Lake.
Having 16-24 good old fashioned P-cores with SMT might be a future. Intel needs another Sandy Bridge effectivity miracle.
i7-2600K was such an amazing processor.

I agree with you. I have a Core i9 9900 KFC with pure 8 cores and 16 threads and it still runs my gaming box today. These E cores and L Cores just waste time. Been pushing intel to bring out 16 core 32 thread desktop CPUs for a long time hopefully it comes back.

 

[AVATARAT]

They are the EXACT same thing. Hyperthreading is Intels branding of SMT.

They do the same job, but in different ways.

 

[Tyler-98-W68]

Because the game isn't graphically demanding, it's computational because of physics applied to rigid bodies and there all it maters is moar cores. Where with actual games you need fast cores and just enough of them to drive everything just right. You can't compare that.


It's odd how Intel hasn't made GAMING processor with just big cores and crap load of cache. Surely their new approach can just glue whatever together with tiles...

Not graphically intensive, wrong again, its very graphically intensive and can easily bring even a 5090 to its knees......it can use ALL of the vram in a 5090 quite easily. You put the resolution to 8k and its 20gb plus in every level.

It can also easily.bring a 5070ti down to unplayable frame rates as well. I've tested the game on multiple systems, and x3d cache doesn't do anything either.


It's not odd, they tried out a bunch of cache over 10 years ago with broadwell. Most people forget about that. Maybe they want a different approach than just gluing shit together or releasing "new" cpus on an old platform constantly

People here talking gaming performance as if everyone here is rocking a 5080+ tier GPU and a 1080p display to truly bottleneck that GPU by any CPU imaginable.

I'd trade some gaming performance for professional performance, efficiency at lighter loads, scheduling optimisations and overall stability. AND SQUARE CPU DESIGN. WHAT IS SO DAMN ATTRACTIVE IN THESE RANDOM RECTANGLES, INTEL?

It's almost like you explained exactly what arrow lake is. Less gaming performance much better productivity, and way more efficiency.

 

[Squared]

SMT is beneficial, on AMD platform you get 15% more perf. at no increase in power draw.

For Intel it's around 20% more performance for 10% more power.

I strongly suggest you to read this article: https://www.phoronix.com/review/amd-ryzen-zen5-smt/8

But you can't see that by simply turning SMT off on a processor that has it, as Phoronix did. Even with SMT off AMD CPUs still pay the power draw price, so these results don't reflect what would've happened if AMD had removed SMT.

 

[Sol_Badguy]

detailing his plan to "step in the right direction." While much of this revolved around AI, its foundry business, and job cuts—a 15% cut in overall head count and a 50% cut in management layers—one of the more relevant takeaways for the PC enthusiast community is that Intel will be reintroducing SMT or Hyper-Threading to its processors in the future.

So AI is a step in the right direction? Great!
Same with job cuts? Awesome!
Bring back that feature that Arrow Lake didn't need in order to beat Raptor Lake? Fantastic!

Yo Lip, here's some free advice, again, how about you guys make sure that Nova Lake doesn't have filler tiles on the die (hello Arrow Lake). Just saying, filling as much as possible with actual processing units might help the CPUs beat Zen6.
Oh and until then how about you drop the price on the 285K?

I don't think Intel can get around the fact that games prefer large fast cores with fuck ton of L3 cache to make framerate go up.

Eh the developers still need to do their fair share of optimizing, it's not like all games are properly coded and only the CPUs are to blame.

Intel needs another Sandy Bridge effectivity miracle.
i7-2600K was such an amazing processor.

GOAT!!!

Hyperthreading is Intels branding of SMT.

Or is SMT AMD's branding of HT?

 

[igormp]

My understanding is whole point of SMT is to give the core something to do when the main thread is stuck like waiting for memory or IO so the core isn't just sitting idle. that's where the performance gains come from, utilizing compute time that would otherwise be wasted idling, thats also why it doesn't need more power as its just using idle silicone.

Not really. When your main thread is stuck waiting for memory or IO you just preempt said thread and run another one in its place. This can be done even on a single core, single thread CPU. It's a matter of concurrency.

SMT is a way to make your internal core's pipeline be always fed with a flow of instructions, and to better utilize the underlying execution units of the core.

 

[evernessince]

If SMT is only a 15% area saving for 20% less performance, just add real cores.

Intel's hyper-threading takes <5% die area for a 15-30% performance boost: https://www.codeblueprint.co.uk/201...oost by using SMT for multithreaded workloads.

That's a huge win and AMD's SMT is even better than Intel's HT, so the benefit is even more in favor of having the tech.

When you have big, wide cores HT / SMT is beneficial. I think that e cores are beneficial on desktop but only 1-2 of them, as they are primarily intended to be used for background tasks they don't impact the user experience. The problem is if it's even worth having to deal with the scheduler issues to add these. Microsoft's incompetence hurts Intel here.

SMT is not HT and works differently, so don't mix examples with it here.
For server purposes, HT is probably a good thing because it provides better parallelism.
For pure gaming, without HT is better (especially with the current Windows Scheduler), yes, some games use it properly.

They are similar: https://en.wikichip.org/wiki/amd/microarchitectures/zen#Simultaneous_MultiThreading_.28SMT.29

They do the same job, but in different ways.

You wouldn't call a toaster with a different orientation of heating elements something else, it's still a toaster.

Same applies here. You can argue the advantages / disadvantages of changes but they achieve the same goal.

 

[ZoneDymo]

Did I mention already how Intel seems to have no focus?

 

[Panther_Seraphin]

A very good representation of how SMT and Multicore differ but also compliment each other from Intel

Orange Thread 1
Green Thread 2
Yellow Thread 3
Blue Thread 4
Grey Idle

Single core - every single thread has to wait for the thread to complete/be put to sleep before the next thread can be loaded into the CPU
Dual core and more - More threads being run at the same time

Single Core Hyperthreading - Both threads are being loaded/processed at the same time on the same physical hardware. As you can see the CPU has actually changed the timing of some instructions to reflect available resources and this is where a strong front end and low latency cache levels are key for this work well.

Multi Core Hyperthread - same as single core hyperthreading but the front end of the CPU has to be even more powerful to be able to schedule the available resources most effectively.

Now in regards to power/heat as you can see with the cores with hyperthreading there is far less idle time on the core so as you can imagine the power/temperature of that core will be higher. It was why I remember in the days of 3xxx+ series Intels one of the reason people were recommending to run Hyperthreading off was purely because running the cores so loaded would add sometimes upwards of 30 degrees per core.

 

[BoggledBeagle]

Intel's hyper-threading takes <5% die area for a 15-30% performance boost

No, that is not correct current information. That could be right in the time when CPUs had 4 cores. That is long gone.

For example, 14900K has 8 P cores, 16 E cores and then 8 P cores - second threads. It does not want to use second threads of the P cores, because they perform the worst. For any loads with 24 or less threads HT is not utilised at all, because it would DECREASE the performance of the CPU. HT only helps in specific case of extremely multithreaded applications. Normal consumers do not use such applications and if they do, 24 threads of 24 cores offer very high performance for any normal person.

NORMAL PEOPLE DO NOT NEED HYPERTHREADING ANYMORE, PERIOD.

 

[Geofrancis]

I have always wondered how much different performance would be if you compared single thread results with single core results with 2 threads on AMD vs a single thread on intel.

 

[TheinsanegamerN]

Not graphically intensive, wrong again, its very graphically intensive and can easily bring even a 5090 to its knees......it can use ALL of the vram in a 5090 quite easily. You put the resolution to 8k and its 20gb plus in every level.

It can also easily.bring a 5070ti down to unplayable frame rates as well. I've tested the game on multiple systems, and x3d cache doesn't do anything either.

Yeah the claim that BeamNG isnt graphically demanding is silly. It wont be demanding if your CPU sucks ballz and cant maintain a reasonable framerate.

However, using 8k as a benchmark is also silly. Nearly any game will be demanding at such a high resolution.

It's not odd, they tried out a bunch of cache over 10 years ago with broadwell. Most people forget about that. Maybe they want a different approach than just gluing shit together or releasing "new" cpus on an old platform constantly

Many people forgot about it because they barely made any. The x3d has shown the benefit of cache. BeamNG doesnt benefit as much because it needs MOAR COARS and the 8 core x3ds are a huge bottleneck.

that doesnt mean the big cache wouldnt help them.

People here talking gaming performance as if everyone here is rocking a 5080+ tier GPU and a 1080p display to truly bottleneck that GPU by any CPU imaginable.

I'd trade some gaming performance for professional performance, efficiency at lighter loads, scheduling optimisations and overall stability. AND SQUARE CPU DESIGN. WHAT IS SO DAMN ATTRACTIVE IN THESE RANDOM RECTANGLES, INTEL?

You act like someone is forcing you to buy an x3d and a 5080.

Nobody is stopping you from buying CPUs more aimed at productivity. Go buy a 9950x or a 265/285.

Its way above my paygrade and knowledge to determine if SMT is an overall benefit or not, but i know for a fact that it is detrimental for gaming and ecores are always better. So if you can save a bunch of space by removing SMT and adding a few ecores instead, im sold. Extra points for removing AVX - if the space savings are big enough you can counteract the performance drop of avx by having more cores. The logic is sound, i have no idea if the numbers (die space saved vs performance drop) check out though.

Removing AVX to add in more cores is like saying you want to remove your car's turbo so you can fit a larger intake. You'd be cutting yourself off at the knees, AVX make a huge difference in any workload built for it.

You can even test this yourself, if you have a coffee lake machine, go find a coffee lake pentium. They did not have AVX, then compare it to a core i whatever with its core and clock rate adjusted to match. The performance difference is shocking. Even back in high school it was noticeable, and that was 14+ years ago on ivy bridge.

 

[Geofrancis]

It's not odd, they tried out a bunch of cache over 10 years ago with broadwell. Most people forget about that. Maybe they want a different approach than just gluing shit together or releasing "new" cpus on an old platform constantly

I remember when the 5775C was launched, the reviews were raving about its gaming performance with that 128mb eDRAM cache. intel could have done x3d style gaming chips long ago but failed to read the room as they were still making quadcores for maximum profit and the cache chip was expensive, it was intended for high-end office computers. I wouldn't be surprised if that's where AMD got the idea from.

A Broadwell Retrospective Review in 2020: Is eDRAM Still Worth It?

www.anandtech.com

 

[Squared]

What I don't get is that when Lion Cove came out, Intel said hyperthreading was out for client use but would still be present in servers. So when Intel says that future server CPUs will have hyperthreading, what is the news?

 

[FoulOnWhite]

I'll leave this here..

RTX 4090 & 53 Games: Core i9-13900K E-Cores Enabled vs Disabled Review

In this week's TPU50 Megabench we're testing whether you can unlock additional gaming performance by disabling the E-Cores on a Raptor Lake Core i9-13900K processor. For our benchmarks we used the mighty GeForce RTX 4090, with 53 games at three resolutions.

www.techpowerup.com

Either way less than 1% average on all of them showing E-cores make near zero average difference. Sure they migh make a difference on some games, but overall just buy a 9800X3D