12. High end desktop CPUs are noticeably faster than mobile counterparts (higher freq, more cache)
Pick anything over the 800$ range that isn’t a macbook and you’re way more likely to hit clocks than not.
Note that workstation-class (H) laptop CPUs also make compromises on performance - the Ryzen 9 4900H is 8C16T but only has 4MB L2$, 8MB L3$, and a max TDP of 54W. A desktop Ryzen 9 3950X by comparison is 16C32T and has 8MB L2$, 64MB L3$, and a 105W default TDP (and will go much higher with even basic PBO if your cooling allows). The differences on the Intel side are even starker.
Yes. And as a bonus, I can use my Alienware 17 R4 as a throwing weapon. Or for workout. And the power brick is a perfect cup warmer.
But I like it, nonetheless.
It’s obvious that it’s impossible at the moment to get 3950X performance in a laptop format, but you can get laptops able to keep temps reasonable with 35-50W, and that’s what a lot of laptop SoCs target as total power.
Those SoCs hit (and sustain) their top clocks, whatever those are for that specific SKU.
What I understood from OP is a common complaint for macbooks, that fail consistently to sustain their specified to clocks, because Apple deliberately under specifies their cooling solutions for better ergonomics (and design reasons).
This is actually completely wrong. Almost no laptops sustain their top (boost) clock on heavy workloads. Most usually only sustain max performance for minutes (or seconds!) before throttling. Here's an example chart that shows how various premium Athena/Evo U laptops perform: https://www.notebookcheck.net/Asus-Zenbook-S-UX393JA-Laptop-...
On the workstation side, people complain about Macbooks, but recent MBPs actually throttle their Intel H processors less than a comparable XPS 15 for example: https://www.notebookcheck.net/Apple-MacBook-Pro-15-2019-in-r...
If you are interested in how modern Intel laptop chips throttle and what base and boost clocks mean, you want to do a search for PL1, PL2, and Tau. For AMD chips, you will want to look up STAPM, Fast and Slow PPT.
Note that while a i7-10875H's top "boost" clock is 5.1GHz, the sustained "base" clock is only 2.3GHz. This is so low to be meaningless as a top speed. In practice, unless your laptop's cooling is absolutely terrible, you'll probably end up mostly running in the 3-3.5GHz range under full load. In comparison, on a properly cooled desktop system, a same-gen i9-10900K desktop system should be able to maintain a sustained (all-the-time) clock of about 5GHz (very close to its 5.3GHz boost). AMD chips scale a little bit better due to 7nm having better power efficiency and how PPT works, but the same ratio roughly applies.
This is the problem with all the marketing BS. When I talked about “top clocks”, I wasn’t referring about “boost” clocks. I’m talking of the clocks that the SoC is designed for (that won’t appear on the box), i.e. a lot of laptops are not leaving “performance gaps” due to bad cooling, those SoCs are designed for that level of performance and attaching a fat copper heatsink won’t do much difference.
I had a lot of trouble with a client that complained that our board was not properly designed because the performance they were seeing was not “as advertised”. In the end we had to ship the whole thing to AMD, and have them test the system with a thermal sink. Everything was as expected.
If anyone is interested in this kind of stuff, your explanation is really good, so I won’t add anything because I’d probably do a terrible job :)
Back in the day, most CPUs had had a fixed clock, but these days modern Intel and AMD chips simply don't - they all clock opportunistically, which depends on powers, thermals, but also workload (try running an AVX-512 loads for example). How do you characterize "clock" in this context? Base (minimum) and Boost (hard limit, now split to Max Turbo <2C and All Boost MC) seem to be reasonably sensible numbers.
Now we can argue semantics all day, but to bring it back around if you're just going to say "top clocks" is what the SoC was designed for at a specific workload/power envelope (In AMD's PB, that'd be PPT, TDC, and EDC) then every laptops will "hit their top clocks with no issues," but I'd say that argument (statement?) is a bit circular/pointless. ;P
How often is that a problem, really?
Multithreaded builds (make -j 24) can really hammer the drive. Read and write interleaved, which uses up cache in both directions.
A well proven way to move heat out of a silicon package to prolong its ability to perform at its highest potential. And in a desktop, you've got room for 'em.
I installed an EK spreader sandwich on my Samsung NVMe drive, and it made a massive (20° c) difference over stock. It was previously a bare stick with no surface area/thermal sink to pull heat away.
You seem to assume it's the case that heats up to 80C or more, that's not true and also not necessary for NVMe SSDs to go over their limit. Your CPU/GPU has fans moving the heat away, that's a better position to be in...
You will just have to accept that you've been wrong on this topic. Move on, it happens.
Read and learn! And then accept when an initial assumption turns out to be wrong.
Go to https://www.computerbase.de/2020-09/samsung-980-pro-ssd-test.... look at the graph. You see that a bunch of them go to the 80C line or hover above. All of them throttle (what you said does not happen). In that graph are shown, going above the limit:
1. FireCuda 520 1TB
2. Patriot Viper VP4100 1TB
3. Samsung 970 Evo 1TB
4. WD Black SN750 1TB (+ the same one with a cooler)
This is only a small part of the market of course, but it goes to show that the throttling is a real thing that happens with multiple models.
Then you had a moving goalpost there, that those SSDs do not throttle under realistic workloads. However, this is a sequential read that's only 5 minutes long. Hardly unrealistic. The hour long constant load benchmark is a different graph, however, constant load is also realistic if it's longer than 5 minutes.
If you activate the other chart modes you see the measured performance, which shows the drops linked to the too high temperature, and that they did the same thing for write performance.
You can counteract this with a lot of targeted airflow and/or a heatsink, the heatsink will at least help move the throttling to a later moment. Gamersnexus had a very impressive demonstration of this, one where they did get this wrong: They had an article about a MSI SSD heatsink where they claimed it did not help (so the SSD did throttle! Again something you said does never happen), where it then turned out that it did not work only because their applied temperature sensors (glued them to the heatsink), and IIRC they also missed the higher performance they got regardless. GN often gets it right, stuff like that happens, but it made this one memorable and highlighted the positive effect of these heatsink coolers.
I'm into this topic professionally for years now. I'm not wrong here. If you can't take my word for it, look at professional SSD reviews, they have covered this also for years now.
And sure: There are scenarios where this does not matter. Gaming. Browsing. But: In those workloads there is no significant difference to a SATA SSD anyway. These NVMe SSDs are only interesting if you have large (and thus: long) file transfers. This is what they have to get right (and some do, but not all of them).
By the way, by repeating that I'm wrong and by starting with a straight "No", by always commenting without reasoning and politeness, you made sure that I will correct you - and that I'm not buying into your strange attempts to correct your statements to something that is correctish. They don't work anyway, these SSDs throttle.
You should change your tone around here.
But you don't know what causes the throttling. Its stupid to shove a heatsink on nand. It does not help. Not a single bit. Period. Under any normal daily usage, or even if you had a workstation, you're not going to be reading/writing so constantly frequently that you're ever going to cause the controller to heat up and cause throttling. If you experience any excessive heat. You have bigger issues in your case. Period. Your only proof of throttling is benchmarks running constant read/writes over a period of time. This is not real world usage and doesn't make it necessary to go out and start shoving heatsinks on every single nvme drive. If that was the case then all the laptops which have space between the nvme and the case, or motherboards which lack a 'heatshield' like the gigabyte board you linked to, would have throttling issues. Which they don't.
> You should change your tone around here.
So now you're threatening me?
-----
Anyway I'm done, not gonna sit here an argue anymore.
There are all kinds of vendors / models, who knows what kind of throttling they use?
Having a fast machine that can sustain throughput and I/O is a perfectly rational desire, and for some of us, need.
To test that, one can try it with ramdisk first, before getting an expensive ssd.
I mean, hardware wise they are not much faster, but cooling is a different story.
The highest spec macbook comes with a 9980HK and starts at 2800 $. A 3900X has approximately twice the performance in multi-threaded workloads and you can easily build an entire quiet workstation with it for less than 1000 $. Half the performance, thrice the price. Great deal.
Yes, there are also "laptops" with a 3900X in them. But even those still have lower performance than a desktop with a 3900X because of thermals.
Specs can never tell the true story, but it's clear that the mobile processor is going to be much slower for anything remotely processor intensive, and probably much more than twice as slow for anything making good use of multithreading.
Having had to go back and forth between a laptop and a desktop for a processor intensive application (AutoCAD) the difference was painful.
* https://www.cpubenchmark.net/compare/Intel-i7-1185G7-vs-Inte...
* https://browser.geekbench.com/v5/cpu/compare/4354735?baselin...
These are short tests too, so would be a best case. In real life, laptop performance is probably significantly worse due to thermal throttling.
While we're mentioning other minor gotchas, another one is memory latency and bandwidth. While desktop systems commonly have XMP and 1.35V support, very few laptops do (typically running at JEDEC timings at 1.2V). While there's diminishing returns, the difference between JEDEC 2933 CL21 or 3200 CL22 and say 3800 CL16 can actually be noticeable in certain workloads and is often effectively "free" (one-click in the BIOS) extra performance on the desktops.
Also, laptops generally have CPUs that pull 15 watts. Whereas desktops have CPUs that can pull 95 watts or higher (sometimes up to 150). The difference is astounding.
See: https://evanmccann.net/blog/2020/5/13-inch-macbook-pro-revie...
40, 100, 29 in pmset -g thermlog instantly.
CAD, for example, barring new geometry work in progress now, is single threaded.
People working on high surface count models want these things:
Big, fat, fast cache
Sustained sequential compute performance
Sustained I/O
GPU that focuses on geometry and precision. This is not generally an issue today, but can be on laptops.
Desktop machines with active cooling are where it is at.
Desktops are great.
Though with water cooling you might also have a lot of pump noise, especially if the radiator is mounted incorrectly.
The larger the diameter of a fan, the lower the RPM it can spin at to move the same amount of air (same cooling capacity) as a small fan. Provided you can put the air where it's needed (e.g. a 1 foot fan can't "focus" air onto a 6 inch radiator) a larger fan will just about always be quieter and more efficient.
Oh, and having a water cooled PC with the radiator and fan inside the PC itself is silly. If you run the pipes outside or into your basement, your PC is almost completely silent, plus the cooling capacity will usually be much, much higher, because not having to cram fans and a radiator into a small enclosure lets you make them bigger and more efficient.
And historically, bigger case fans have been problematic, those 200mm fans they tried to introduce some years ago. Bad static pressure if I recall correctly?
Also, I wouldn't call having a radiator and fan inside the PC silly. The case eats a lot of the noise already, it's the easy and the common setup, and good AIOs are quiet and cool well. But maybe you just wanted to share a cool big water cooling setup with everything noisy routed into the basement ;)
The hoses aren't welded steel, you know.
In my case I have a DDC style pump, and its very quiet after I got a car wash sponge and cut a square hole in it and put the pump inside. It looks ghetto but its inside a case so who cares.
At first I was worried about heat, but its lasted 10 years so far.
I heard a story about a DIY PC that did nothing but circulate water through the cpu cooler from a fish tank; the tank was big enough to dissipate heat through evaporation and other means, only occasionally needing a top off.
I actually mounted an automotive transmission cooler to the outside of my PC and made it part of the watercooling loop, because it was inexpensive and large. There are no fans on it, but it still reduces the work the standard pc cooling radiator with the fans on it needs to do.
The most limiting thing is TDP, which in the highest performance laptop processors is still capped at 45W, whereas a maxed out desktop processor can draw 100W or more.
See these tables for i9, for example, compare Coffee-Lake-S (Desktop) with Coffee-Lake-H (Laptop) https://en.wikipedia.org/wiki/List_of_Intel_Core_i9_processo...
Maybe a small desktop. My desktop processor is 180W TDP (Threadripper 1950x), while some others are 250W TDP. You can also get a dual-socket workstation, for 2x CPUs (both pulling 200W each).
Thermals and power are significantly higher on desktops, it ain't even funny. Laptops win in power-efficiency, but absolute performance is always going to be a Desktop.
My Macbook laptop, on the other hand, sounds like a jet everytime I run yarn install.
So yes, desktops have higher thermals. But it handles it so much better than a laptop that it almost becomes irrelevant.
As a result, your 10 year old desktop is probably 50-100% faster than the most expensive Macbook Pro or Thinkpad. It can be quite astonishing swapping to even an old desktop after using a laptop for a long time.
I'm planning to drop in a 5950X upgrade in the workstation in a couple weeks as well. Looking forward to both the huge multi-threaded and IPC gains.
I have mine set to 80% and only set it to 100% if I know I'm going to be away from AC for a while.
Time after time, as someone who's been working on his 10 year old desktop (with a replacement ssd + graphics card when the old one died), I meet devs and analysts using laptops who reason more or less: "well it says i7 and it says x GHz and it says ddr3/4 and it's got a gpu with the same marketing number, so laptops perform the same as desktops cause they have the same hardware in them don't they".
Clearly, what they really mean is "I've never worked with and compared with a desktop". I suppose one of the problems is that SOME of them that have 'used desktops' were actually using neutered VMs in a shared corporate environment that run really poorly and are pretty underspec'd in a shared environment.
But every time, it's actually been the case that not only is the desktop faster and cheaper, but things usually remained faster on an X year old desktop hardware vs more modern laptops for any serious workload.
Edit: and in case it needs to be said, I have both desktops and multiple portable devices in my household because the downside of desktops is clearly portability.
More expensive, less performant, and less serviceable with a suite of proprietary bloatware on top.
I might break the pattern this year though, I got a 3900x last year around launch and moved my 4670k to a home server. I've been pretty impressed with the 3900x while the 4670k has been maxing out all cores in the server for some tasks so was considering buying a second 3900X(T) to replace the 4670k again. But with the rumours about the 5900X, I might end up putting the 5900X in my desktop at the same time as a 3000 series GPU and moving the 3900x to my home server.
Most laptop cooling is awful, but you can certainly find laptops that are well built if you look for it.
Not sure you should weigh bloatware, either. You can trivially install a fresh Windows or Linux and you have to on a custom built PC anyway. If you buy a premade PC, it probably comes with the same crap.
For what it's worth, I switched to a desktop once the core race heated up - now I've got a 12 core 3900X.
I can't believe how much faster it is when it's using all the cores. Night and day. Highly recommended. And 12 cores is barely scratching the surface of the crazy workstations you can build these days.
For tasks 4 threads and less, it would be a bit faster than a current gen Intel laptop chip, but I don't think it would have been worth the portability penalty to me personally if that's all I did with it.
The problem is this base clock speed is given by the CPU manufacturer not the laptop maker. And Intel wouldn't know what kind of laptop it's getting crammed into. So careful definitions don't really help. Desktops are already big clunky things that have to be kept plugged in, I can trust them a lot more to deliver the CPU's promised performance. Whereas laptops are notoriously making compromises because customers tend to be very unrealistic about noise, battery life, etc.
Laptops being power-conscious, they're usually much closer to the point of maximum efficiency on the power curve. In that sense you get better performance per watt. But that's negated by the increased front cost.
At full load, it depends on the heat dissipation; Dell's Precision 7XX0 dissipates heat well enough to keep the CPUs from throttling, but the 5XX0 does not.