Intel Ditches Hyperthreading For Lunar Lake CPUs (pcworld.com) 56
An anonymous reader shares a report: Intel's fastest processors have included hyperthreading, a technique that lets more than one thread run on a single CPU core, for over 20 years -- and it's used by AMD (which calls it "simultaneous multi-threading") as well. But you won't see a little "HT" on the Intel sticker for any Lunar Lake laptops, because none of them use it. Hyperthreading will be disabled on all Lunar Lake CPU cores, including both performance and efficiency cores. Why? The reason is complicated, but basically it's no longer needed. The performance cores or P-Cores on the new Lunar Lake series are 14 percent faster than the same cores on the previous-gen Meteor Lake CPUs, even with the multi-thread-processing of hyperthreading disabled.
Turning on the feature would come at too high a power cost, and Lunar Lake is all about boosting performance while keeping laptops in this generation thin, light, and long-lasting. That means maximizing single-thread performance -- the most relevant to users who are typically focusing on one task at a time, as is often the case for laptops -- in terms of surface area, to improve overall performance per watt. Getting rid of the physical components necessary for hyperthreading just makes sense in that context.
Turning on the feature would come at too high a power cost, and Lunar Lake is all about boosting performance while keeping laptops in this generation thin, light, and long-lasting. That means maximizing single-thread performance -- the most relevant to users who are typically focusing on one task at a time, as is often the case for laptops -- in terms of surface area, to improve overall performance per watt. Getting rid of the physical components necessary for hyperthreading just makes sense in that context.
In other words ... (Score:5, Funny)
They run too hot!
Re: In other words ... (Score:3)
I have always seen hyperthreading as a fake way to run multiple tasks compared to multicore.
Task switching comes with overhead even if you have hardware that supports it.
Re: In other words ... (Score:5, Informative)
I assume it's heavily task-dependent. The only benchmarking I've done was with CFD, where we found that enabling hyperthreading would provide a wall-clock speedup of about 15-20%. Hyperthreading might also provide well better than a 15-20% speedup for other tasks, or even negative under the worst conditions.
And this in no way compares hyperthreading to simply adding more cores, which has its own advantages and disadvantages.
This ignores the most important... (Score:2)
This fails to account for the most important feature of hyperthreading.
Which is of course the satisfaction of seeing 32 nearly idle cores in task manager when in face you have a mere 24.
Re: (Score:3)
Big ass Java app at work after being well tuned would run 64 cores at 99%. Well, ok 63 because one was reserved for OS and other crap, but it was a thing of beauty to watch it go on 'top' or 'ps'.
It can be done but it took tremendous effort to make that happen. At home, playing games and other light work nonsense I never saw more than 3-4 cores doing anything at all much less multiple running at/near 100.
HT was useful for certain loads when we only had a few cores but those days have passed. It's a good
Re: (Score:2)
"At home, playing games and other light work nonsense I never saw more than 3-4 cores doing anything at all much less multiple running at/near 100"
This applies to SSD benchmarking & usage too. Some of the drives have fantastic numbers at queue depths 8 or higher and 16 threads but that is not even remotely what a typical end user would benefit from.
Joe Sixpack gets the most benefit from the lower latency and dramatically better random read/write performance compared to just about any single HDD but isn'
Re: (Score:2)
Truth! I picked my home SSD based on charts for QD 1-4 and performance curves within the cache. The odds I blow out the cache for any normal daily use are about zero.
Re: (Score:2)
"I picked my home SSD based on charts for QD 1-4"
Which did you choose?
Re: In other words ... (Score:5, Insightful)
Re: In other words ... (Score:5, Insightful)
A modern CPU has multiple execution units of each type. If you've got an ideal instruction stream, it'll keep all of them busy simultaneously. But you'll almost never have that happen in the real world. Out of order execution is one way to keep the units busy. Hyperthreading is another. Using either alone will give big gains, but using both is going to offer a much smaller benefit, as there's just less unused CPU power still remaining.
Re: (Score:2)
(Atom model) CPUs are also nice for not having any of the Meltdown/Spectre type vulnerabilities.)
Came to the comments for the Meltdown/Spectre/etc.. stuff.
IIRC, disabling hyperthreading in the bios prevented (some of) those exploits (I can't remember exactly which ones). I wouldn't be the least surprised if those sort of security concerns helped drive this development (Intel CPUs without hyperthreading). [don my tinfoil hat] ... maybe even at the governments request for a more secure general purpose CPU?
FWIW, I've always turned off hyperthreading on my personal computers. Even testing full rebuilds of
Re: (Score:2)
I'm not sure that a government asking for more secure equipment from a vendor counts as a conspiracy.
Re: (Score:2)
IIRC, disabling hyperthreading in the bios prevented (some of) those exploits (I can't remember exactly which ones). I wouldn't be the least surprised if those sort of security concerns helped drive this development (Intel CPUs without hyperthreading). [don my tinfoil hat] ... maybe even at the governments request for a more secure general purpose CPU?
I'm not sure that a government asking for more secure equipment from a vendor counts as a conspiracy.
(NOTE: I didn't use the word "conspiracy").
The government requesting some feature, or lack thereof, would not, in and of itself, be a conspiracy. However, if both the government and the company were aware of security implications impacting all their products with hyperthreading (that are still being sold, btw), but were not telling us and instead marketing it as a performance/efficiency improvement, that's would clearly be a major conspiracy. It would be like the government knowing of (or even playing a han
Re: In other words ... (Score:4, Interesting)
I have always seen hyperthreading as a fake way to run multiple tasks compared to multicore.
Task switching comes with overhead even if you have hardware that supports it.
I'd say it's complementary rather than fake, as it allows you to get more usage out of each core. Sure, all else being equal I'd rather have two cores than one with hyperthreading - but I'd rather have two cores with hyperthreading than two without.
Re: (Score:2)
Re: (Score:2)
Would you rather have 6 cores with hyperthreading than 8 cores without? It does cost silicon area (and development time and effort).
Probably not. I was responding to a post calling it "fake". It does increase performance in many cases - I've seen 0-50% or 30% being ranges quoted, and 0% is also possible - but with 1/3rd more cores instead that would be better. That said, I'm pretty sure the amount of silicon added per core is much, much lower than the 33% extra added by going from 6 to 8 cores. After all, it does share the execution units and the cache with the core. Do you have any figures? I didn't find any with a quick search. The
Re: (Score:2)
Re: In other words ... (Score:5, Insightful)
That's exactly what it is. A single processor core can execute several instructions simultaneously under ideal circumstances, but you're rarely going to keep the processor fully busy. Hyperthreading is a way to use that otherwise unused computing power.
Hyperthreading is basically a fake extra core. Duplicate the register set and the bits visible to software, but share the rest with the main core. The hyperthread core just uses whatever bits of the CPU core that happen to be idle at the moment. The hyperthread will always run at a fraction of the speed of the main thread, but the hyperthread costs very little to add to the chip, so it's worth it if your priority is performance.
Now that Intel is prioritizing power efficiency and heat output over raw performance, they've decided it's not worth it.
nope (Score:1)
CPU vendors wouldn't support it if it didn't help. It's a cheap way to use surplus CPU hardware to get extra performance. The beefiest CPU cores (the ones in workstation CPUs like Power) go up to eight threads per core.
At this point there are enough cores to handle typical laptop loads and Intel made the decision that cutting out the SMT hardware was worth the power savings. Can't argue with that.
It's not about task switching, as you proposed. If anything there's less task switching with SMT, because more t
Re: (Score:2)
Task switching overhead is one of the things that hyperthreading really helps with. Since it effectively doubles the number of cores you have, tasks can run for longer before needing to switch. When they do switch, the whole CPU core isn't stalled writing and reading state data, it can continue to execute another task thanks to hyperthreading.
It looks like Intel thinks that the overhead for hyperthreading is more than offset by higher boost clocks on CPUs with a lot of cores.
speculative execution (Score:5, Insightful)
Seems like this is a positive thing given all of the issues with speculative execution, should help mitigate CPU vulnerabilities.
Re: (Score:1)
Re: (Score:1)
Hello, have you met the Cortex-A65? [arm.com]
Re: (Score:2)
the transistor budget is better spend on other things. this is why no arm/riscV chips employ SMT/HT.
No. The transistor budget is customised to maximise the design goals of an architecture and product. There are ARM chips out there with hyperthreading (A65, A78 to name two), but the majority target really low power applications and thus don't. On RISCV you have a completely different architecture so you won't see mention of hyperthreading or SMT, but you do see references to RISCV processors with multiple harts per core, which is effectively their way of doing what SMT/HT is doing on x86 CPUs.
On most PC CP
Re: (Score:3)
This was my first thought as well. Security.
Pretty sure they gave up trying to make it secure while still being performant. OpenBSD shut it off completely in a "better safe than sorry" moment, and from what I've heard, linux kernel patches reduced the performance a bit too.
Re: (Score:2)
But this is on laptops, not servers. It makes sense to disable it on servers used in VM clusters as this is a way to get secrets leaked out especially between customers.
Laptops generally are used by a single person at a time, so the need to leak secret keys is much lower. I mean, if you're already able to compromise the laptop, I'm sure there are easier ways than using Spectre or
Re: (Score:2)
Not really. Firstly hyperthreading can already be disabled on any system. If you have security concerns than uncheck it in the BIOS. Secondly only a small subset of speculative execution attacks depend on hyperthreading.
It is objectively better to leave this option in the hands of the user. There are still many workloads that can get a 10-15% increase in performance with hyperthreading on.
Laptops have a heat limit (Score:2)
hyper-thread shaming (Score:3)
Lunar Lake is all about boosting performance while keeping laptops in this generation thin, light, and long-lasting.
So... Hyperthreads are fat, heavy and weak? Daaaamn.... I guess when they sit around the laptop, they sit *around* the laptop.
Ya, I get they're really talking about the extra power/cooling needed, but personally I don't need a laptop to be razor-thin and/or feather-light. Also, while I could understand some vendors selling laptops w/o hyperthreading to help achieve those goals, I can also imagine some people wanting more and be willing to buy a laptop that can support the extra power/cooling needed, so disabling the feature all around seems unwarranted.
Also, since when does an electronics (specifically computer) maker *want* their products to be "long-lasting"? I mean, MS just required HW specs for Windows 11 that required people to buy all new hardware (to be officially supported anyway) and the PC makers were like "Um, okay. More sales for us."
Re: (Score:2)
Do you travel often?
The extra 2-4 lbs on a heavier laptop feels mighty heavy after a few long flights, cab rides, and so on.
My last work laptop I got one of those ultra thin ones instead of my usual, "I can run 60 fps hires shooters" laptop. It really did matter on those trips and didn't make any difference in the office because I hooked it to a big monitor in the office.
Re: (Score:2)
Do you travel often?
The extra 2-4 lbs on a heavier laptop feels mighty heavy after a few long flights, cab rides, and so on.
I get that -- that's why I didn't say all. But weight isn't the only, or major, concern for some (many? most?) people. Some people want the form-factor *and* power. Just saying that it should be an option. Also, why does Intel even care (about the reasons they gave)? They don't sell PCs / laptops. Why not leave it up the the actual vendors?
Re: (Score:2)
I used to always travel with a redundant system, so the weight savings were quite important to me.
I ran windows on boot camp on 13 mac pros for a while since it was a nice small light laptop with an excellent screen (the early retina screen when 1080p was a stretch for a 13 inch windows laptop).
Re: (Score:2)
What were you doing that was so important you carried a backup device?
I've never heard of anyone doing that. Curiosity++
Re: (Score:2)
"important" is stressing it.
But I did graphic presentations for trials. I could pull up any document and expand any paragraph and highlight the sentences.
What I could do is switch the input and if a laptop cried nobody would know.
Also, I had backups for hanky external hard drives if that happened.
Basically my job was to make sure what the lawyer said was on the screen, and then not to know I put effort into it (that was kind of a bummer, the better I did my job the less work it looked like I did. But someho
Re: (Score:2)
Very cool, thank you. And totally makes sense now.
I was trying to imagine why it would matter even for a CEO on vacation if their laptop choked and other standard corporation jobs but couldn't figure out why any of them would need a backup unit.
Re: (Score:2)
I don't need a laptop to be razor-thin and/or feather-light.
Lunar Lake is intended for a buyer who isn't you.
I can also imagine some people wanting more and be willing to buy a laptop that can support the extra power/cooling needed, so disabling the feature all around seems unwarranted.
It's not like this will be the only laptop CPU that exists. Lunar Lake will be useful to a different type of user than these imaginary "some people" -- they can buy something else. Different products serve different market segments, and are rarely one size fits all; this is a perfect example.
BS (Score:2)
Bullcrap. HT/SMT has been shown to improve MT performance on average by 30%. A 14% performance uplift is not enough to cover this deficiency. They indeed might have some other ideas in mind, nad probably HT will be reenabled in the future. At least AMD is not giving up on it either now or any time in the
Re: (Score:2)
Bullcrap. HT/SMT has been shown to improve MT performance on average by 30%. A 14% performance uplift is not enough to cover this deficiency. They indeed might have some other ideas in mind, nad probably HT will be reenabled in the future. At least AMD is not giving up on it either now or any time in the future.
There's a power, colling, performance and silicon ageing. Throw that into a 4D constraint space and solve for the optimum architecture. HT didn't make the cut. HT in part makes up for suboptimal assignment of instructions to execution HW. When you get better at the latter, the benefit of HT goes down.
Re: (Score:3)
They're basically disabling hyperthreading but adding more efficiency cores. You're getting extra full, slower cores instead of the extra threads on the faster CPU. It's about maximizing performance within a limited power budget.
Ideally desktop chips won't have the power restrictions and can the threads back.
Re: (Score:3)
Its about physical silicon space. Hyperthreading is not free, and requires a significant amount of space on the chip.
Intel's E-Cores this generation are looking to have ~50% boost in performance, and they're significantly smaller than their P-Cores.
When doing single-threaded tasks, you want a P-Core not sharing its resources. When doing multi-threaded tasks, you want tons and tons of E-Cores. SMT/HT just doesn't make sense anymore in the current technological landscape when you factor in performance-per-die
Re: (Score:2)
What real work jobs are you running that HT makes a difference?
I don't mean theoretical benchmarks.
If you have such jobs, you may very well be the edge case that gets hurt by this decision but given how many cores sit entirely idle 99% of the time on most computers, it is statistically unlikely you are the outlier.
Re: (Score:2)
The jobs I ran with HT were just the same program run multiple times. The HT squeezed a few percent more out of the box.
Extra cores squeeze a lot more percent out of the box.
In my case it's analyzing nondeterministic data from silicon circuits. It's all so much faster these days. I remember running jobs (entropy analysis tasks) on some new silicon in the lab with many, many cores and I didn't have time to go get some coffee before it finished. That was a first.
Reasonable. (Score:5, Insightful)
One of the biggest things that prevent a superscalar out of order processor from being fully utilized is data dependencies between instructions. For, instance, assume you have a 4 way out of order processor and want to calculate:
A = B*C + D*E
There's three math operations in there, and you can do up to 4 at a time, so you ought to be able to issue all three operations at the same time, right? And the answer of course is "nope", because you can't issue the addition until after the results from the two multiplications have already completed.
So, let's assume that you have a idealized 4 way superscalar processor with out of order execution and every instruction will take just 1 clock cycle. But, the average code that the processor sees will 100% of the time be able to execute at least 1 instruction, 50% of the time at least 2 instructions. 25% of the time at least 3 instructions, 12.5% at least 4 instructions, and so forth, halving the percentage of the time that N+1 more instructions can be executed. So, on average, how many instructions per second will this processor handle?
Doing, the math, you'll see the answer is 1.875 instructions per clock cycle. So, in other words, less than half the theoretical capacity of the chip. Now, there is some specialized code that can fully utilize the capability of saturating all 4 execution units, such as matrix multiplications and such. But the average code has that 100/50/25/12.5/.... pattern.
Now, assume you want a faster processor. Adding the ability to execute 5 instead of just 4 instructions at a time would speed things up. So, instead of an average of 1.875 instructions per clock, we get 1.9375 instructions per clock and waste an even larger percentage of the possible power of the chip. The limit would be 2 instructions per clock.
But, if you add a completely separate set of registers and execute a completely independent thread, so that the two threads don't share any dependences with each other, we get, with the same code base, on average, 3.25 instructions per clock, with 1.625 instructions per clock on each thread. So, by going to hyperthreading, we increase the system performance from 1.875 instructions per clock to 3.25 instructions per clock (a 73% increase), but decrease the per thread performance since the two threads are competing with each other for execution resources.
So, yes, dropping hyper threading support will increase single thread performance. And will lower power requirements since more of the chip is idle since it will usually not be capable of using its full capacity due to data dependencies between instructions. Seems a reasonable compromise since a laptop isn't a server handling multiple high computational tasks simultaneously.
Re: (Score:2)
That was beautiful. Thank you.
Re: (Score:2)
Reading your elegant description I wonder.
Does better power management (as in partial chip management, which I think is a thing now) also mean dropping HT leads to greater power savings and skew the value proposition too?
Like when power is important, is the savings of just letting some of the chip be idle more than it used to be?
HT (Score:1)
Re: (Score:2)
It was a workaround for not having enough cores.
Now we have lots of cores that mostly don't do anything.
Security Flaws (Score:3, Interesting)
Re: (Score:2)
I wonder about the security flaws that were related to Hyperthreading.
Me too. In particular, when they say, "The reason is complicated, but basically it's no longer needed," is that because the fixes they had to put in to secure hyperthreading have nullified its performance benefit (to some extent)?
Thread Director (Score:2)
"The reason is complicated, but basically it's no longer needed."
That sounds like a 15 year old answer.
My understanding is HT is making Thread Director hard to use.
We Don't Want Your Pussy Ass E-Cores. (Score:2)
All this 'hype' for Intel putting a band-aid on a bullet wound.
The jerkiness of Windows (Score:2)
Yes but the OS spends a lot of background cycles checking the hardware, even with I/O interrupts for signaling, and refreshing the GUI. Then there's all that background telemetry and to-the-cloud copy/save operations, then there's actual spyware - and Windows 11.
It seems the jerkiness of Windows 11 will increase, greatly.
Worthless article (Score:2)
If I wanted to read an Intel press release, I would read one, not PC World.
Even their longer "deep dive" article lacks any mention of the numerous security flaws associated with Hyperthreading.
It's journalistic malpractice by PC World to not have any mention of this. I'm sure it played a large role in the decision to drop hyperthreading, regardless of what Intel claims the reasons are. Even Forbes mentions the word "security" in relation to hyperthreading, though the sentence is not particularly understanda
Makes sense (Score:3)
When you had 1 or 2 cores, hyperthreading made a big difference.
Now when you have 16+ cores, for SMT to even come into the equation, you need more than 16 threads all wanting maximum CPU
Translation of "no longer needed" (Score:2)