The article's title says "at 35W", but the article text says "the T-series chip is rated at up to 106 Watts". That appears to refer to the chip's short-term turbo power limit (PL2 in Intel parlance), typically effective for a default of 28 seconds if thermal limits don't kick in—but the time parameter can be adjusted by the end user (and this benchmark run has unknown provenance, so who knows what power management settings were in effect).
Since the benchmark in question (Geekbench 5) only runs for a minute or two, it does most of its work before the chip even attempts to throttle down to anywhere near 35W. An actual power measurement averaged over the entire benchmark run would yield a significantly higher value than 35W.
It is a shame these kind of knobs are rarely exposed on a laptop. Sometimes I throttle my laptop cpu down to like 800MHz-1.2GHz to save power/prevent the fan from going. But this makes Vim less responsive for certain events. What I’d actually like is normal behavior but with the turbo clock limited to like… I dunno, a second.
If you're running on Linux, almost all of this is completely configurable using ryzenadj for AMD laptops or powercap and thermald for Intel chips. Here's a document on how you might go about adjusting this w/ powercap: https://github.com/junaruga/framework-laptop-config/wiki/Imp...
Or you can always fine tune the on-demand governor to gracefully speed up the processor as the load increases and keep it there for a second or so and bring it down.
Modern Linux kernels load the `intel_pstate` frequency driver by default these days which has it's own governors that don't quite act the same way as generic ones, but you can control scaling for them w/ the EPB/EPP hints.
In Linux you want powercap tools. You can tune all three parameters in terms of power and time constant, so you can do exactly what you suggest. You can change it on the fly without rebooting so you could have a ridiculous setup like mine where I have the power limits wired up to two unused buttons on my mouse.
Seems to work pretty well. I was initially confused — it seems the default on my system had the “long term” limit set to some very high value, which made my changes to the short term moot. Seems like an odd default but I’ll blame ASUS for that I guess.
Still slightly confused as to what exactly the “time window” means physically. With mine set to 1000us I still seem to get multiple seconds at full power. But it bumps down before the fan spins up so that seems good enough!
Looking forward to the inevitable realization tomorrow that this is going to make my computer unusable in zoom calls.
If Vim isn’t responsive in a machine with hundreds of MHz something is very wrong with your setup. Vim was fast on machines with dozens of MHz. Either a bad terminal emulator or some plugin making it slow, I’d guess.
So there are two problems I guess, user-based and technical.
User problem: since I use Vim for everything, my expectations as far as responsiveness goes are probably not calibrated correctly. It is almost always very good, so the moments of badness are noticeable. The user (me) is unreasonably picky.
Technical: LaTeX syntax highlighting. It is I’m pretty sure, generally known to be a bit of a pain. Lots of nested environments that can go back dozens of lines and change how commands are interpreted. It is happy in a decent (not huge or anything) sized Fortran code, so I think config is mostly OK. Maybe I should look around and see if there’s a nice LaTeX plugin out there.
Some vim plugins are slow. Fortran highlighting used to be super slow for certain use cases 10 years ago. I think there were some bad regexes. Latex documents are often long (cf other code) and the syntax rules are complex.
There are definitely some pathological situations where vim becomes unresponsive even on a beefy box. Usually there's exceptionally long lines of text involved, esp. with syntax highlighting/:setf active IME.
Furthermore there's the whole intentional delay on ESC exit from modes like an aborted search query... that can make it feel unresponsive but it's just a timeout. Try it, hit '/' in vim, then ESC to abort the search, count a full second of unresponsiveness.
This blog speaks to such ESC delays one encounters in vim:
According to https://vt100.net/heath/z19-om.pdf it can do 110 to 9600 baud. It was connected to some Z8000 based Onyx server. It was generally good but incredibly slow when more than four or five of us were compiling code for the tests at university class. Pascal, Berkeley p system or something with a similar name.
It's not vim, but the OS/graphical system that you are most likely using underneath. While having Firefox/Chromium open for reference to what you are working on.
Specifically worth comparing terminal emulator performance. I noticed that for example (not surprisingly I guess) alacritty has been both the worst and the best depending on the setup.
The benchmarks themselves aren't really deceptive, but wccftech's reporting on them is misleading. These aren't claims being made by Intel, but just random user benchmarks that appeared online.
For all benchmarks, context matters. These days, most in-depth reviewers understand how to work with transient power and thermal limits with respect to time. For example, measuring the MacBook Air M1's peak speed/time vs. its lower sustained speed after it hits thermal throttling around the 5-7 minute mark.
It's kind of hard to compare performance/watt meaningfully when the wattages aren't at all similar.
If you limit power on current generation Intel/AMD processors, the perf/watt is pretty good; depending on the specific load, sometimes better, sometimes worse than M2. When you allow more watts, perf goes up, but perf/watt goes down as you see diminishing returns. (With some loads, you may also have a point where more watts reduces performance, whoops)
Geekbench isn't great as it doesn't even attempt to capture power usage, but that may not be available or accurate on all systems anyway.
> Since the benchmark in question (Geekbench 5) only runs for a minute or two, it does most of its work before the chip even attempts to throttle down to anywhere near 35W.
That doesn't sound right at all. Nothing in a PC system (absent water systems with huge reservoirs I guess) is going to buffer excess heat on minute-long timescales, not even close. There's literally nowhere for that energy to go; 125W for a minute would be melting the solder joints.
CPU throttling works on at most small-integer-second scales.
CPU throttling happens on many different timescales. That's why Intel processors have multiple power limits (PL1, PL2, PL3, PL4), in addition to current and temperature-based throttling mechanisms. The time constant for PL2 is usually either 28 seconds or 56 seconds. At those timescales, the concern is usually not with power delivery or CPU die temperature but rather with exterior case temperatures that the user may be directly touching.
Based on the reported benchmark performance, it seems very unlikely that temperature-based throttling kicked in, and it's clear that the chip was operating well above 35W for at least a large portion of the benchmark run. So the PL1 and PL2 turbo power limits are the relevant controls at play here.
Yeah yeah, I know. I'm just saying that there's zero chance that throttling is affecting this measurement. The idea that an Intel machine is significantly faster for 2-3 minutes after the start of a benchmark is just silly, that's not the way these things work. Go start a benchmark of your own to see.
Again, the thermodynamic argument is fundamental here. You're saying that a "35W" CPU is "actually" drawing power equivalent to a 125W CPU for exactly the time of a benchmark, which is several minutes. That excess would have nowhere to go! There's no reservoir to store it. (Obviously the cooling system could take it away, but part of your argument is that the cooling system is only good for 35W!).
> The idea that an Intel machine is significantly faster for 2-3 minutes after the start of a benchmark is just silly, that's not the way these things work. Go start a benchmark of your own to see.
I've done so, on many occasions, with actual power meters rather than trusting software power estimates. You really do commonly see a laptop's power consumption drop significantly ~28 seconds into a multithreaded CPU benchmark.
> (Obviously the cooling system could take it away, but part of your argument is that the cooling system is only good for 35W!).
I make no such claim that the cooling system is limited to 35W. I only claim that the default platform power management settings from Intel impose a 35W long-term power limit, unless the system builder has adjusted the defaults to account for whatever form factor and cooling choices they've made.
Perhaps you haven't realized that the turbo power limits will still kick in even if the CPU die temperature is not too hot—because they're not actually a temperature-based control mechanism?
It's a socketed CPU intended for low-power small form factor systems and thus will usually be running with Intel-recommended power limits or lower, for all the same reasons that laptop CPUs are usually running with low power limits. The control mechanisms don't actually function any differently between their laptop and desktop CPUs, they just have different default parameters (the various turbo limits).
The only relevance of the socketed nature of this part is that it is easy to put it in a normal desktop form factor where a big heatsink and possibly tweaked turbo limit settings can be used to generate misleading benchmark results. But it's not actually certain that this is what's happening; the Intel-recommended default behavior for this chip can plausibly produce the reported results—just not in any way that could be reasonably described as "35W".
Even if the configuration may be different from motherboard to motherboard and from laptop to laptop, exactly as wtallis said, most Intel CPUs are configured by default to consume during the first 28 seconds a power 2 to 3 times greater than the nominal TDP, e.g. 105 W for a 35 W CPU.
Most, if not all, subtests of GeekBench need less than 28 seconds, so it is quite possible for the entire benchmark to be run at an 105 W power consumption. Whenever a subtest finishes, the power consumption momentarily drops, which resets the 28 second timer.
If the computer has poor cooling, it may happen that when the CPU spends too much time and too frequently at an 105 W power consumption the junction temperature limit is reached, which triggers thermal throttling and the power consumption is reduced. This is a different mechanism, independent of the one that reduces the power consumption down to the nominal TDP after 28 seconds, or after another configured time.
Thermal throttling reduces the power consumption only enough to keep the temperature under the limit, so the power consumption may remain greater than the TDP until the 28 seconds pass.
I was curious about exactly how much burst heat could be absorbed, so I asked WolframAlpha [0]. In a 15" workstation laptop, I think the CPU could quite reasonably pull +100 watts over steady-state TDP for 30 seconds. (100-gram aluminum heatsink absorbing 100W * 30sec -> ∆T = +33ºC)
I always get confused with Intel’s clock speeds and power draw now. Because depending on the p-state or c-state it’s in, it can be severely underclocked. Even trying to set “performance” p-state doesn’t work all the time, at least on my Ice Lake Xeon.
100W is also my observed sweet spot on C++ project build times. At 100W I got the minimum build time. At higher limits the cpu draws more power indeed but the build takes just as long.
It still feels insane that a CPU can draw that much power. I mean that's like over 8 amps at 12 volts for doing... well nothing really. It's 100 watts of pointless resistance losses.
Something that could've powered like half of an electric scooter at top speed. And GPUs are like 3-4 times worse.
With its 8 amps—more like 80A at typical core voltages—it retires potentially hundreds of billions of instructions per second. That is a long way from nothing in my book.
Well sure, but would you count that as actual "work" in terms of physics? I think not, at least not to any degree that should be notable.
There's no practical work being done in a CPU, it's just throwing away charge so it can redirect it around and do some calculations in the process, like a river flowing downhill with some mechanical logic gates in the stream.
The main problem I guess is that x64/x86 aren't designed for power efficiency as much as just pure speed. RISC does a lot better, and biological computers are on a few orders of magnitude more efficient. So there's definitely a lot of room for improvement.
During switching both the n and p transistors in a cmos circuit are open simultaneously for a brief period of time. During this time you are essentially shorting your power supply to ground!
The performance ratio relative to power consumption places this in new territory for Intel. While the T-series isn't new, I don't recall seeing it competing with the likes of K processors and Ryzens.
Also the 1.x GHz base clock which turbo boosts to 5GHz, wow. Do other professors scale across such a wide band?
Stunning performance, especially for a (heavily iterated on) 10nm node. Alas I doubt we'll see many mini-pc/1L PCs configured with such a top tier node but what an epic mini-server that would make! So many cores (8p+16e)!
Thanks, yes, I'm aware of those, but they are quite a bit bulkier than the Cirrus7 cases, at least triple the volume. What I also like about Cirrus7 is that you can get a complete build, even though it isn't cheap.
What’s the actual power consumption? Without those figures the article is just a clickbait. Looking at those multicore results I don’t believe for a second that the CPU was drawing less than 90 watts, probably more.
I'm not sure it's fair to say the 5950X is a good point of reference. The 7950X has been out for quite some time at this point and the 7950X3D is announced and expected to be available next month.
And a ton of people are cross-shopping more expensive newer parts with still-available 5000 series CPUs which are much cheaper. There’s an interesting question about whether absolute performance is more important or if you should weight that against cost and/or power.
The GN benchmarks recently have done a good with explaining both the cost perspective and the performance-per-watt.
Edit: for reference, Newegg currently has the 5950x and 7950x at $500 and $600 respectively. And that’s before the higher platform cost for AM5/DDR5.
If reading a review of the newest vendor's product, it is only fair to give the performance of the competitor's similar generation release. If there are strong availability/price concerns (eg generation N+1 is triple the cost of generation N) those should be noted as to why not performing the most appropriate apples-to-apples comparison.
Calculating performance / dollar is an entirely different exercise (though valuable, and rare is it the newest generation that hits the sweet spot).
That’s one definition of fair, sure. I use a different one; how does this option compare to my other options, sans artificial constraints.
As another example of this, there’s a great mkbhd video comparing new cheap smartphones with old flagship smartphones. He does that every couple years, and it always gets a ton of “I never thought of that” comments.
Living in a place where electricity costs are very high, such as California, the AMDs have much better price to performance when compared to the Intel CPUs over a period of 3-5 years (which is how long I normally wait before upgrading CPUs).
The 7950X3D is not going to be faster in non-game benchmarks. Most likely it will be a bit slower in most synthetic tests and non-game workloads (rendering, encoding, FEA etc.)
It's a fair point of reference for those of us with a 5950X, which was a top performer a couple years ago and is still a damn lot of CPU no matter how you slice it.
The 13th gen i9 in question has 8 performance cores and 16 efficiency cores, and 36MB of cache, whereas the 12th gen has 8 and 8, with 30M of cache, so it's not that surprising it performs better.
My guess is (slightly?) better binning in addition to "forced" PL1 and PL2 limits. Better silicon can run stable at lower voltage, ergo lower power, ergo lower temps, so I bet there's some binning for T SKUs so Dell & co. can ship not-overheating micro PCs.
I see this frequently, but what is “better silicon”? Physically something is off about the manufacturing which impacts performance, but does not kill the chip. What are those defects?
There are 0.1% bad transistors instead of 0.2%? Heat output is more uniform? The bad transistors are clustered in such a way that signal routing is more efficient and leads to a measurable throughout difference?
Very slight changes in the transistor junctions will make them slightly more resistive, more inductive, physical variance will make them narrower, wider, etc. All of these factors, although extremely slight, will add up to different response curves of switching time vs current vs voltage.
Makes sense. T series chips generally aren't as widely available as individual components as their mainstream higher power counterparts. To your point, they are available from OEMs as complete systems, who seem to get priority.
There is advice to just buy the non-T series and limit power. Interesting to see that, at least in this example, they aren't quite equal.
I saw a HN comment that chips that perform well at lower voltage don't scale as well to higher power, and vise versa. By that logic, T-binned chips would perform better at a low TDP compared to a downclocked 13900.
I might want to buy a new powerful 2023 laptop, but there's ton of confusion about CPUs, graphic cards, etc. As usual, there's no simple way for a buyer to understand what's best.
This isn’t due to some “amazing process improvement” for Intel that some wealth managers will shill to their clueless clients. Undervolting and underclocking is a thing. Par for the course for any new generation of chips other than Intel’s which are still way behind the curve. It’s like cheering a city bus because the driver wasn’t drunk for their shift.
Undervolt/downclock it? Afterburner or GPU Tweak, etc, will do this relatively easily. (I've undervolted AMD cards that come with ridiculous core voltages without having to drop clocks at all in past)
Almost everything is stock clocked past the point of diminishing returns right now. This Intel part looks to mostly be a downclocked version of existing 13900.
Look at the recent AMD 7900 vs 7900X. You can get 90-95% of the performance for far less power by just backing off the voltage and clocks a bit. (In their TDP terms going from 115w to 65w TDP, loses less than 5-10%)
Everyone's fighting for chart king/significant generational improvement number they can point to and missing the sweet spot on the efficiency curve, but you can bring it back yourself. I bet the 3080 still runs great at 50-100w less power limit/TDP depending on your use case and I doubt that will result in anywhere near a 1:1 perf/power reduction.
Yeah, exactly. I was able to get a decent ~50W drop and stayed at a higher than base clock rate (1920 MHz @ 900 mV). It's still a space heater, but it's better.
I also frame rate limit myself in a lot of games - I don't need my MMOs running at 160 fps, so many games I'm at 150-200 W. I still wish it was less, but that's much more reasonable than 400 W.
AMD's GPU drivers give you a decent range of control over board power limits, fan curves, GPU and memory clocks and voltages, and on at least some models the ability to fine-tune the shape of the voltage/frequency curve used. Really the only thing missing is an automated tool to explore the V/f parameter space to find the limits of stable operation for your particular chip.
I'm using the software that came with my card to do that. It makes sense that the graphic card manufacturer's would ensure stability with the cost of it running hot - by undervolting I'm playing with potential crashes and that won't do for a card defaults settings.
> (I've undervolted AMD cards that come with ridiculous core voltages without having to drop clocks at all in past)
It's often even worse than that. There are plenty of cases where you can undervolt so far that you now have enough headroom in the power delivery and cooling to allow you to run at substantially higher clock speeds.
Arguably Intel is now just catching up with the times after their long manufacturing quagmire, I suspect large part of the "leaps" are due that. As GPUs have been managing more steady progress, such leaps do not happen. Also there is lot of doubt around these sort of benchmarks, especially for P+E setups
Depends what you mean. It generates more heat, it’s just more effective at getting that heat off the die and into your room.
Edit: since it’s performance per watt is higher, if you’re capping frame rates then you can get less heat out of the 4090. Like I said, depends what you mean.
This has better single threaded performance that M1/M2 by a significant margin, but at a higher total power. When the M1 was released I believe it has the fastest single threaded performance of any CPU. That is no longer the case by quite a bit now.
If power efficiency is a concern, you wouldn’t want to look at the absolute top of the line enthusiast-targeted chips.
I care greatly about power consumption in my laptop, but I don’t really care at all in my desktop. If they can make my compiles finish 10% faster by doubling the power usage, bring it on. My CPU is rarely ever running at these 100% usage levels, so it’s not like it makes a difference in my power bill. Modern coolers are plenty quiet.
anyone who buys a 7950x to run in eco mode is a fool, like buying a race car to go 30 km/h. The target market of the 7950x and the i9 are not people worried about their electricity bills.
There's the hope that this is just a stopgap measure from Intel to deal with AMD and Apple competition, and power demands go back to normal once Intel had time to push a couple of architecture improvements through the pipeline.
In a Laptop its quiete stupid to stay at 100w just to have a slightly better Performance. I mean its only a significant margin as Long as you can get the heat away, which is often the biggest problem in most laptops…
(M1 Max uses half the power for geekbench 5 and ~1650 and being a 2 years older chip. If the m2 max can hold the Power limit and still get to ~2k it would be fastly superior, i.e. the normal m2 is already at ~1850 With Less than half of the Power of the 13900t)
So it's more like putting a lipstick on a long dead design. Sure may look fresh from afar, but it still stinks.
What I am trying to say is that at this level of power consumption Intel should be many times faster than M1 - that is offering substantially better performance per watt.
Sure there are niche applications where power consumption doesn't matter and only the single core performance counts, but could older tech be overclocked and achieved the same result?
Not sure. Point being is that what Intel does is increasing pollution and causing people to bin perfectly working machines just because there is a "new" CPU on the block, where in real life they probably won't see a difference.
It's really irresponsible thing to do for Intel. They should go back to the drawing board and stop releasing meaningless products until they actually get something worth upgrading to.
It seems that AMD is losing the performance crown for consumer PCs for both single threaded and multithreaded workloads at high efficiencies. I love AMD, but it does seem that Intel is back and at worse neck and neck. It isn't clear if it can clearly pull ahead.
This isn't great for AMD's stock, the boom times may be over?
What confuses me is that AMD still has a huge lead for server CPUs. I have not seen massive adoption of AMD in the cloud yet.
Intel can reach equal or better performance than AMD only at higher power consumption.
At equal number of active cores and equal power consumption, an AMD CPU has a much higher clock frequency than an Intel CPU. At equal number of active cores and equal clock frequency, an AMD CPU has a much lower power consumption than an Intel CPU.
For desktop and laptop CPUs, Intel can win with this higher power consumption strategy, especially in single-threaded applications, where the fact that Raptor Lake has a higher IPC than Zen 4 also helps.
On the other hand, for server and workstation CPUs, Intel cannot benefit from this trick, because those CPUs are used in multi-threaded applications where they run constantly in thermally-limited conditions.
The Intel Sapphire Rapids CPUs have from many points of view a superior design in comparison with AMD Genoa, but they are severely handicapped by the inferior manufacturing process used by Intel.
Because of that, the Sapphire Rapids CPUs have low clock frequencies when all cores are active and small cache memories. This ensures that they are no match for AMD Genoa and they barely match the previous generation of AMD server CPUs.
The manufacturing process disadvantage is compounded by the stupid Intel policy of randomly disabling various features in most SKUs, with the exception of those that are ridiculously expensive.
This policy ensures that the benchmarks published for Intel are excessively optimistic, because they show the top SKUs, while the SKUs that most people have are crippled, so they may have a much lower performance.
So in servers the Intel CPUs are not competitive at all, except for various special applications where the CPU performance matters very little, e.g. when the cheapest SKUs are good enough for providing a big memory and many PCIe lanes, or when it is acceptable to pay Intel to enable the included accelerators, because the application can benefit from them.
Each brand has its strengths and weaknesses. AMD isn't dominating anymore, but they are still ahead in areas such as AVX512 (absent in Intel Core), the ability to mount more than 8 performant cores in a chip (Intel's intra-core communication architecture is difficult to scale, hence their new P+E approach, which fortunately seems to work well), and their ability to take advantage of the multi-chip process to mount absurd amounts of cache. The multi-chip approach is also a tech advantage that allows them to cut costs.
(Disclaimer: I work at Intel, but not on server products. Opinions are my own.)
There's a lot more that goes into a server platform than just the cores, for instance: the BIOS code, the BMC support, the maturity of the motherboard designs, etc. These are all areas where Intel seems to still have an edge -- but I'm also very excited about our upcoming server architectures on the core/compute side.
> What confuses me is that AMD still has a huge lead for server CPUs. I have not seen massive adoption of AMD in the cloud yet.
This was years ago so the tide may've shifted but: part of it could still be vendor experience and "it works"-experience?
When EPYC gen 1 & 2 came out, we were shopping for a bare metal megaserver spec at work. I got a budget and free rein (except it had to be Dell), and I really wanted to pick EPYC for the better core clocks (which were significant in our build architecture) and more cores and better cost.
With one EPYC spec and one Xeon Gold spec built (EPYC was I think $13k vs $15k?), work was a bit uneasy about AMD processors just yet. Our workload was MSVC compilation but they were concerned about architecture differences, since all of our workstations & laptops were Intels. They preferred paying more because "we already have Xeon servers and they're proven".
So, we ended up getting the Xeon Gold spec instead.
IMHO, any benchmark should run with Turbo disabled. And come with additional tests about how much turbo brings, and how long it can stay on within a given thermal setup. Otherwise all you have is garbage, not data.
Turbo is such an inherent feature of new CPUs that turning them off would make the benchmark completely unrepresentative for real world usage. For example - a 16 core CPU might run a single core at extremely high clock rates during single threaded workloads. What would be the point of turning it off?
Complex devices need complex benchmarking, unfortunately. You won't get a simple, single number that shows how powerful a cpu is.
Why? It’s not like turbo is not a real performance enhancement. With enough thermal headroom, there’s no reason why turbo can’t be sustained, as heat is the limiting factor.
Good CPU reviews include long render and compile benchmarks that would suffer if the turbo performance couldn’t be sustained. If a CPU can sustain high performance, then it doesn’t matter if it’s using a turbo mode.
It's all about the "With enough thermal headroom". If this isn't specified in the benchmark, then the benchmark is useless because you can't compare with your setup. It's even truer in laptops, but this is true of servers and desktops as well.
In addition, recent CPUs will use this thermal headroom to the limit.
You can have both at the same time by publishing a graph of performance over time where you see the ramp up and the time where the turbo can't be sustained. Quite a few reviewers started doing that maybe 2 years ago?
Since the benchmark in question (Geekbench 5) only runs for a minute or two, it does most of its work before the chip even attempts to throttle down to anywhere near 35W. An actual power measurement averaged over the entire benchmark run would yield a significantly higher value than 35W.