The i9 option on the new 15” Macbook Pro?
There’s been quite a bit of chatter regarding the i9 CPU option on the new 15” Macbook Pro and suggestions that Apple’s cooling system isn’t up to par to handle it, with some testing even going so far as to demonstrate that the new i9 15” Macbook Pro is actually slower than the old top end i7 15” Macbook Pro. That’s not something you’d expect for something newer and “better”. I actually did to an extent, but I didn’t expect it to present to the degree that it has in the aforementioned case.
To spoil my own article (if you’d rather not read the whole thing), I agree that thermal throttling is occurring with the i9 and you need to take this into account depending on your use case whether or not it’s (an i9) truly needed, however, I disagree that the new i9 is slower than the older i7 without all the benchmarks even out yet. Indeed, as I was writing this article, Apple has now acknowledged a fix for the bug which is largely, but not entirely to blame (get to that in a bit). I believe there is some copy variance going on here with the factory application of the thermal paste and how lucky of a chip you get too which I’ll get to as well. That said, any of the new Coffee Lakes in the 15” model should be faster than any of the i7’s before it. But, the i9 option is going to yield diminishing returns as it’ll get thermal throttled the most. To truly get the most out of the i9, you’ll have to go the route of redoing your TIM with better paste or liquid metal (which I don’t recommend the latter for mobile applications like the Macbook), as the i7-7920HQ before it also needed to be and had the same TDP. Otherwise you may be wise to just get the 2.6ghz i7 configuration and be happy with it, similar to what I did with my 2014 Macbook Pro.
Some time ago (2013-2015) I supported a large Apple pilot program and we had to establish a software and hardware baseline for our users. One of the things we discovered with suitability testing of various configurations of Macbooks was the highest end CPU offerings in particular, often didn’t deliver on the results and thus wasn’t worth the cost to benefit ratio. The results were certainly there, but, we found they were small, and often largely (but not entirely) diminished after placing the Macbooks under prolonged heavy load. For our users which were heavy on CAD and virtualization, it didn’t make a lot of sense vs if it had been an office environment (single threaded tasks) it may have made more sense. Instead we opted for the 2nd highest end CPUs, the largest memory configuration available and 512GB SSDs as the memory did prove beneficial both for virtualizations, CAD and battery life ironically. 512GB was a sweet spot for storage needs. I too, with my own personal Macbook Pro, choose the higher end “stock” configuration of the 2014 model (512GB model), but didn’t go for any of the higher end CTO options (highest end CPU or GPU) knowing this and still, in some cases I see some minor thermal throttling on occasions even without having the highest end CPU/GPU. It would appear the retina (2012 and newer) Macbooks thermal envelope is geared towards their mainstream processors, and not wholly suitable for the higher end customize to order options available for some time, so this isn’t new news to the i9. You can buy CTO options like the i9 or higher end i7’s of the past, you just run a higher risk of thermals throttling your processing the higher you go and we opt’d not to push the envelope, pun intended.
Jumping to another story which I’m going to connect the dots here, even further back my sister’s iBook was randomly rebooting, running with fans at full power often, and suffering poor performance and battery life. I had a hunch based off the symptoms this was a thermal issue and cleaned her fans with a can of compressed air. Normally with those symptoms on a PC laptop, that would do the trick. No dice here though. Since this was a personal laptop, not someone else’s who was paying me to fix and thus no liability as this was my sister not a stranger, I tried something more risky as my sister was talking about just getting a new laptop since it was out of warranty and the problem wasn’t palatable to her. I took it apart and replaced the thermal pads (which Apple was using at the time) with thermal adhesive (not thermal paste) as I needed something more conductive but would fill the large gap of the pad as that would be too much gap for thermal paste, and probably wouldn’t be wise to have a paste glob that big in something mobile. This was risky though as thermal adhesive I used was electrically conductive. Well after a few hours of teardown and careful execution and re-assembly, it fired up, and ran quiet, without rebooting and excellent battery life. I did note, the heat pipe now got very hot but the fans weren’t kicking into high gear, which at first was alarming that perhaps I had done something wrong, nope. You see, the hotter the heat pipe gets from efficient transfer of heat from the CPU, the more efficient the fans can pull heat away from it thus the CPU runs colder and the fans don’t have to run as fast; you just have a hot pipe or heat sink, but your CPU itself is cooler and your fans don’t have to work as hard as they’re blowing out less, but hotter air out IE they can draw out the same amount of energy at less RPMs, or more energy out with same RPMs.
I also some years back bought a dual core 2.3ghz PowerMac G5 tower that had been back to the shop 5 times from the school district and the tech shop couldn’t figure it out for the life of them what was wrong thus the store manager permitted an out of policy return on the tower and firesold the tower for $1000 “as-is”, seemed like a steal to me as I had a PowerMac G3 at the time and was a former Apple Tech that had been promoted to business sales; I figured I could fix it if anyone could. I bought it and ran the ASD (apple service diagnostic) disc on it when I got it home. Thermals were off the charts (80c) despite fans running full power (4000RPM) even while the machine was idle (the picture on the article is of my former G5 CPU module with heat pipe and sync attached, btw). I decided to replace the stock thermal paste with some quality stuff and cross my fingers. My first application of Artic Silver 5 sustainably improved the results. However, I noted that one of the cores still ran 10c hotter than the other and my fans were still around 1500rpm. I went a second round and redid my own thermal paste; this time I used less paste, making sure I used just enough and no more. Sure enough, I was using too much thermal paste the first round and now both cores were quiet as a mouse on the temps (within 2c of each other), and so were the fans (1000RPM). Hmm, two cases of bad application of thermal management on Apple products right out of the factory. By the way, I realized after my 2nd attempt that the real reason I needed to re-grease the CPU in the first place wasn’t that it needed better thermal paste (although I’m sure it helped), but, the original factory paste had been slathered it on too thick IE bad application if it. Quality control issue obviously, and one that was occurring many moons later IE Apple still hadn’t fixed it. Wonder if they still have QC issues with thermal paste today? Maybe, maybe not… Hold that thought for a bit.
So, I maintain that the i9 is faster than the former top-end i7, in spite of thermal throttling, all things created equal (which holding that prior thought, maybe they’re not all made equal or there is a 1+1=3 as to why some, but not most, benchmarks aren’t demonstrating this). Now why would I maintain that (other than it being newer and “better”)? Let’s backup a bit and explain what the new Coffee Lake CPUs represent to explain how they impact different use cases.
First up, let’s compare the former top-end i7 of the 2017 15” Macbook Pro against the base Coffee Lake of the 2018 15” Macbook Pro…
2017 15” Top end: i7-7920HQ – 4 cores / 3.1ghz base / 4.1ghz turbo / 8MB cache / 45W TDP
2018 15” Base: i7-8750H – 6 cores / 2.2ghz base / 4.1ghz turbo / 9MB cache / 45W TDP
So right out the gate the new base entry Coffee Lake has more cores, same turbo and a bit more cache, but, runs at a slower base clock and supposedly generates up to as much heat as the high end 2017 i7 Kaby Lake CPU, I say supposedly as the i9 which we’re getting to also supposedly generates up to the same 45W’s, yet is far more powerful then either.
What this means is, as a race car (single threaded tasks), the new low end Coffee Lake achieves the same turbo as the old high end Kaby Lake, but has a bit more cache as a cherry on top. Theoretically on paper, the new base 2.2ghz Coffee Lake CPU should be a hair faster at simple things like word processing, browsing the web, etc then the old high end 3.1ghz Kaby Lake. We’re not even talking about the 2.9ghz i9 Coffee Lake yet, or the higher end “stock” 2.6ghz i7 Coffee Lake that comes with the 512GB model either.
Now as a truck that is doing heavy lifting, the base frequency that’s supposed to represent worst case scenario (assuming there isn’t a thermal issue, as both are rated for 45W) is slower on the new low end Coffee Lake, but it has more cores to make up for it. Thus, 2.2ghz x 6 cores = 13.2 vs the 3.1ghz x 4 cores = 12.4. Anything that’s well threaded and optimized code, should run faster here too, once again, not by much (6%), but, it’s doing more at a lower price point. Thus you could in theory buy the entry level 2018 15” model, and it’ll be faster than the older top-end 2017 15” model on both fronts. The low end model isn’t likely to see any throttling, I might add.
I could digress into server CPU offerings and how some are optimized for being a “truck” that is lots of cores operating at lower frequencies which is great for big data chomping or virtualization, or as a race car, that is less cores operating at higher frequencies which is better for databases, office tasks and non-optimized code, or sometimes both (IE a bit of both), think race truck, and these CPUs maintain the same TDP with the variance being how far you swing one way or the other generally speaking. But to keep from stretching this out further, just take my word for it, they exist and you see this tradeoff generally speaking. What you see with this tradeoff is as you drop the frequency, you can generally achieve more load through more cores operating at a lower frequency, precisely what we just saw, a benefit of the having more cores if you can squeeze them in (which they squeezed in for Coffee Lake), but, those type of processors usually loose out of single threaded tasks like databases, office processing, basic performance. In this case though, we picked up 2 more cores, without losing turbo though. The new Coffee Lake can have its cake, and, eat it too. It gets the benefits of having more cores that can run slower that gives more combined carry weight if you will, but, without losing top speed, on just the base model. Thus, it’s safe to say any of the options above the base model, IE the 2.6ghz i7 Coffee Lake or even the 2.9ghz i9 should be even faster than the old Kaby Lake, at everything, all things created equal.
So what could possibly go wrong?
Let’s double back for a bit here and tie it together…
My existing, base stock i7 Haswell on my 2014 model is getting thermal throttled at times. Not badly like my G5 or my sisters iBook that I had to fix, but, it hits the ceiling and it’s not even the top-end offering which obviously faster CPUs are going to run hotter and thus throttle even more. Many folks have noted their higher end 3.1ghz Kaby Lake i7-7920HQ’s get thermal throttled, thus no surprise there. To put 2+2 together here, all things created equal (that is they didn’t screw up the thermal paste application, you got a “bad” CPU or they didn’t accidently forget to associate the machine ID correctly in their auto-firmware updates, which they did, and just fixed), the i7 Coffee Lakes and i9 Coffee Lake are going to get throttled as well, as they have the same TDP as the i7-7920HQ which also was getting throttled before it in the same cooling system. But, once again, Coffee Lakes have more cores that can down clock when they hit a thermal issue, and can more turbo, depending which CPU you choose, so even throttled it’s more optimized to deal with the lack of thermal envelope, if that makes sense. On the other hand it can run a faster turbo clock on that i9, but then it’ll tap out at that faster turbo quicker if you do so. Thus you may not see say a 50% improvement, but you should see something pretty substantial. The trouble becomes you may see diminishing returns on those higher end CPUs like the i9, which I’ve warned about before as they were getting thermally throttled well before they got fully loaded.
Translation to English, as I said much earlier in my article, the mid-range i7 found in the stock 512GB configuration is a much wiser choice as it’ll likely be wasted less if that makes sense. There’s a reason the i9 is a CTO option, and not the default.
So what happened with the i9 that was actually slower than the former i7? I suspect that particular copy actually had two problems, not just the fix Apple put out a bit ago. That one, probably had a bad thermal paste job or didn’t win the CPU lottery, plus needed the VRM fix Apple kicked out (they’re not calling it a VRM fix, but in fact based off what I can tell, it is, as the VRM was defaulting in lieu of being associated to the correct firmware key which the correct key was dictating more voltage, uh huh, VRM fix folks, to models IDs for 2018 15” Macbook Pros...).
Now someone might ask, so why would throwing it in a freezer fix a voltage problem, isn’t that a thermal problem? Not necessarily. Electronics need less volts to maintain a given clock speed when you lower their inherent electrical resistance, which you can do by either throwing it in the freezer, or, as professional over clockers do, apply LN2 cooling which makes it super cold really bringing the electrical resistance down to achieve otherwise unachievable clock speeds regardless of how much heat you can dissipate IE bigger fan cooling or water cooling, both of which are dependent on ambient environment temperatures but can dissipate large amounts of generated heat but can’t lower the CPU to colder then ambient temps. Water cooling / big air cooling is not your freezer. Example, I accidentally overclocked my old G3’s L2 cache too far one time, and I was able to dissipate the heat just fine, but it didn’t have enough volts to comply and got stuck in a broken startup loop so I couldn’t undo it. Putting it in the freezer for a while resolved it as I was able to successfully boot, then lower the L2 clock speed, and remove from the freezer and boot normally. Heat generated actually never entered the equation; I had to lower the electrical resistance enough by freezing the component so it could operate at the higher clock speed with the default volts. In this case, they raised the volts to the correct value to meet the spec, something I couldn't do on my G3, thus I had to freeze it for it to work to spec, and then I lowered my spec to meet the volts I couldn't change because I couldn't boot it up to change it in the L2 cache control software.
Of course, throwing it in the freezer solves any thermal dissipation issues too, which I do suspect the firmware in question that was correctly associated for automated update was the SMC, which includes both voltage regulation, and, thermal/fan regulation; odds are both got messed up if they messed up on the key-machine association on their patch database which it sounds like they did. Good one Apple. I’m sure they won’t repeat that mistake again with configuration management. Suffice to say, Apple probably fixed it-fixed it, not band-aid fixed a hardware issue with software. In this case they restored the originally intended SMC firmware, which didn’t make it by accident, not rolled out a fixed firmware, if that makes sense.
Now, something else I alluded to, the thermal adhesive was better than thermal tape at transferring heat which in turn allowed that iBook to run faster, quitter and made the heat pipe hotter. The same is true of the thermal paste, it can be replaced with either better paste, or liquid metal which is the ultimate solution. The tricky part is liquid metal is not so appropriate for mobile applications as it could potentially shift in transit, but on the other side of the coin, even the best thermal pastes still throttled the former i7-7920HQ on the 2017 Macbook Pro which is also a 45W TDP; it takes liquid metal to “get it there” so it doesn’t throttle. Dollars to donuts if you don’t want to the i9 to throttle, or throttle the least, you’re looking at redoing your TIM with liquid metal which is risky business. Would I do it? No. Good paste like Kryonaut is a much safer solution that’ll get you most of the way there without the risk. That I would do, if you’re up to it and have an i9, even with the Apple fix you’re probably still going to get thermal throttled somewhat as it’s still a 45W TDP just like the 45W Kaby Lakes before it got throttled due to heat.
If you’re not technically inclined, just buy the higher end stock i7-8850H 2.6ghz i7 model and pass on the i9 CTO option and save yourself some pennies but still get a decent upgrade. Odds are it won’t have a bad thermal paste job as most machines are fine and throttling should be minimal without making under the hood changes.
In any event, I’m happy to see Apple fixed their last minute whoops on software update database / firmware that was the big culprit here, but, the original suspected culprit of thermal headroom is still an issue for the i9 to a large degree and either getting the next best thing (the high end 2.6ghz Coffee Lake) or replacing the TIM with better thermal paste is a suggested option depending how technically inclined you are. Or just be okay with diminishing returns, that’s an option too; the i9 should still be a bit faster, but not a lot faster than the next best i7 in any event even without making any changes. I have to agree with John Poole of Geekbench and indicate that thermal throttling under load is probably occurring on the i9 at i7 clock ranges (~2.6ghz) IE an i9 is only really going to benefit you for the limited turbo use scenarios of quick light loads, but any heavy lifting like renders will likely only be faster due to the added 3MB of cache, which represents maybe 1-4%, but probably closer to 1-2%, unless once again, you intend to slap on some Kryonaut, or the more risky liquid metal to see substantial gains.
Tuesday, July 24, 2018