A few weeks ago, a friend asked me what the problem might be when all browsers stop working. More specifically: Whenever they tried to launch either Microsoft Edge or Mozilla Firefox, they only got an error message saying they didn’t have sufficient permissions. It was a strange error – why would permissions suddenly be missing after years of trouble-free operation?

After some back-and-forth, we managed to install a new version of Mozilla Firefox with step-by-step phone instructions and finally installed AnyDesk using a working browser, which allowed me to take a look at the problem myself. That’s when a surprise came to light: The PC has been running Windows 8.1 Home, 64-bit, just as it always has!

Having gone without any updates for quite some time (end of life was on January 10, 2023!), the problem ultimately came down to expired root certificates. Continuing to run 8.1 wasn’t a viable option anyway, so I performed an in-place upgrade to Windows 10 22H2 remotely via AnyDesk. That took hours because the factory-installed 3.5″ WD hard drive with 1 TB capacity* is still the boot drive. The already very underpowered fourth-generation dual-core i3 is further severely hampered by the now very limited 4 GB of DDR3 RAM in the form of a single memory module. All in all, a significantly outdated system, but one with the potential to be upgraded very cost-effectively. That’s exactly what I did today, a few days later.

After a deep clean, everything looks so much brighter:

Upgrades

The PC is used solely for viewing photos and occasional web browsing. So there’s no reason to switch to the very latest and most expensive hardware. For a reasonable price, you can still get a lot out of this 2014 Acer Aspire, here’s a comparison of the current components and the planned upgrade:

The UEFI was already up to date; I checked that before taking anything apart. The reason is that newer processors within the same generation are usually only supported by updated microcode in the form of a newer BIOS or UEFI version. The planned i7-4790 is actually a refresh within Haswell-DT, so it’s even newer than the original i7 release models and incompatibilities cannot be ruled out.

Here are the components; everything except the CPU is from my own stash:

Processor replacement

This is incredibly easy thanks to the FC-LGA 1150 socket:

For thermal paste, I used the tried-and-true MX-4 from Arctic* and 99.9% isopropyl alcohol* to clean the components. I added the two new RAM modules, and lo and behold: POST was successful, and it’s running. Everything is recognized correctly in the UEFI. Now I just need to shrink the partitions, clone the hard drive to the SSD, and put everything back together.

Success?

For comparison purposes, I ran Cinebench R23 and the PassMark Performance Test v11.1 on the original hardware at the beginning and then again after the upgrade. Both tests were run on the same Windows 10 Home 64-bit system. This makes it easy to see the performance gains:

The performance gains are consistently very substantial and extremely noticeable. In particular, the integrated graphics unit – which is slightly more powerful in the i7 than in the i3 – benefits further from dual-channel operation with two memory modules (instead of just one as before). However, due to the patches addressing Spectre and Meltdown as well as other factors, the results do not quite match the reference scores. Here are the Cinebench screenshots:

Next, the PassMark results – though, due to a lack of internet connection at the time of testing, these are only available offline:

Last but not least, here is the HWiNFO overview of the respective configuration:

Der Beitrag Acer Aspire XC-605: A Late Upgrade in Times of Skyrocketing Storage Prices erschien zuerst auf flohs blog.

How could it be otherwise when you handle a lot of hardware? Particularly low-priced devices tend to end up sooner rather than later at their final destination—with me.

The combination of mobile, inexpensive, battery life, and x86 rarely bodes well. Add “convertible” to the mix, and well… Such is the case with Lenovo’s “Miix” series. Two tablets from this series found their way into my hands, one after the other, with a little time in between. One was the Lenovo Ideapad Miix 310-10ICR and the other was its direct successor, the Lenovo Ideapad Miix 320-10ICR. Both are equipped with a 10.1″ display.

⧉ Lenovo

⧉ Lenovo

⧉ Lenovo

⧉ Lenovo

⧉ Lenovo

⧉ Lenovo

The first impression when switched off is really not bad. Glass display, the keyboard is securely docked, the workmanship is high quality. After switching it on, disillusionment sets in – the data sheets didn’t lie. Intel Atom with 2 watts TDP in combination with an embedded multimedia card as mass storage is simply very slow – but cheap. Just like the whole device, bought by different people in various Mediamarkt bargain bins on special offer.

At least both have the maximum RAM expansion, even back then (2016/2017) with 4 GB LPDDR3, which was already very small. With 2 GB of RAM, which was also available, these devices would be completely unusable with Windows 10, regardless of the version. Constant swapping to the slow eMMC simply cannot run well.

Upgrade…?

No way. As a classic ultra-mobile device, there is really nothing that can be upgraded in terms of hardware (without special equipment and soldering tools). So the only option is to try to improve the performance of the software. My somewhat hopeful attempt to improve responsiveness by replacing the factory-installed Windows 10 Home with Windows 8.1 Pro failed miserably: There are simply no working drivers, especially for the processor’s iGPU (however, the 64-bit version of 8.1 could be installed without any problems).

It’s a shame, because 8.1 works really well with a touchscreen and, although it is now less widely supported, it is also a less bloated operating system. Only ~700 MB of RAM were in use on the desktop. All it would need is a reasonably up-to-date browser.

Incidentally, 64-bit Windows 7 Ultimate with integrated SP1 does not boot even with Secure Boot disabled; it freezes with a pink stripe at the edge of the screen when attempting to do so. 32-bit operating systems generally cannot be booted; they are not even displayed in the boot menu. The UEFI only wants to load x64 EFI files.

Even Windows 10 probably runs noticeably better on such limited hardware in the 32-bit version than the preinstalled 64-bit edition. Well, well.

So Windows what?

I have gradually tested the following systems on the slightly newer Miix 320-10ICR, all 64-bit:

• Windows 10 Home (21H2)
• Windows 10 Pro (22H2)
• Windows 10 Enterprise 1607 LTSB (2016)
• Windows 10 Enterprise 1803 LTSC (2019)
• Windows 11 Pro (22H2, 23H2, 24H2)
• Windows 11 Enterprise IoT 2024

This took a really long time due to the very limited performance. With all these systems, everything ran smoothly with the drivers provided by Lenovo for “Windows 10 64-bit”. It is very important to install the driver for the “Intel Power Management IC (PMIC)” first, as some devices, such as the GPU, will only appear after this has been done.

Windows 10 Home was preinstalled and, due to certain preloaded programs, noticeably slower than the subsequent fresh installation of Windows 10 Pro. Windows 11 Pro, which was installed shortly thereafter, ran for a total of four years and was repeatedly updated with in-place upgrades. The speed of everything slowed down significantly each time until it became almost unbearable with 25H2:

In the end, it was the significantly aged Windows 10 Enterprise 1607 LTSB N from 2016. It runs best and comes debloated in the standard configuration. Even Windows Defender can be completely disabled with three clicks via group policy, and Edge is not even present. The .NET Framework in the still fairly current version 4.8 runs on it—this is a prerequisite for some applications.

Intended use

What do you do with such undersized hardware?
In this case the obvious: control, monitor, and regulate.

Both tablets run at different locations with different tasks, with the software stripped down to a minimum with Ungoogled Chromium in full-screen mode. In this configuration, it’s not as bad as you might think. I do have some concerns about “continuous power supply operation,” which can’t be good for the tablets’ relatively large batteries, but so far everything is OK.

Nevertheless, the best thing to do is simply not to buy hardware manufactured as “Day One electronic waste”. These tablets never ran well, even when brand new. Probably even worse, due to a lot of pre-installed software from Lenovo.

Technically good hardware simply has its price, and an x86-based tablet that cost less than €180 new at the time can’t be any good, no matter how you look at it. Even modern Microsoft Office runs rather poorly on it; when Windows 11 was still installed, even Cinebench R23 crashed due to a lack of free RAM. At least it can run under Windows 10 1607 LTSB N. However, there will be no more updates for this version of Windows 10 from mid-2026, so it is not recommended for everyday use. However, the two devices will probably remain in use as “monitoring screens” for a long time to come.

Der Beitrag Windows with that special touch erschien zuerst auf flohs blog.

Spectre and Meltdown are two security vulnerabilities in processors that were discovered some time ago, in 2017-2018. There are two variants of Spectre: CVE-2017-5753 (bounds check bypass, Spectre-V1, Spectre 1.0) and CVE-2017-5715 (branch target injection, Spectre-V2), both of which affect a large number of CPUs – x86 (Intel, AMD) but also ARM. CVE-2017-5754 was assigned to Meltdown and only affects Intel CPUs.

Software fixes have been implemented over time for all vulnerabilities. However, since the vulnerabilities exist in the hardware, they cause varying degrees of performance loss depending on the processor architecture. A very useful little program can be used to disable the “mitigations” or “workarounds” in Windows.

Why would you want to do that?
These mitigations can significantly reduce computing power, and it is highly unlikely that the security vulnerabilities will ever be exploited in a meaningful way in the private sector.

⧉ Toms Hardware

To squeeze a few percent more performance out of aging hardware running modern Windows (10/11), this is definitely a quick and sensible intervention. This is especially true when overclocking older processors, where every point in the benchmarks counts.

Switch off.

Gibson Research Corporation has released a super useful little program that allows you to quickly and easily get an overview of your current system and decide for yourself which mitigations you want to enable or disable.

You can download the tool here. The program must be run manually as an administrator. Modern processors have fixed the gaps in hardware, so there is no need for action. But the tool also reveals this information.

Field trial

The effects can be illustrated using the example of the Intel Core i7-7700K used in the previous “Küche PC” project. This was delidded, coated with liquid metal, and overclocked to the maximum. These comparison results were achieved under Windows 11 Professional 25H2:

As you can see, the results are not earth-shattering, but they are measurable. Kaby Lake suffers less from mitigations than older generations before Skylake. “InSpectre” therefore also attests to “Performance: Good” with mitigations. It should not be forgotten that microcode updates also have an impact and cannot simply be switched off or circumvented.

Der Beitrag Spectre and Meltdown – A Blast from the Past erschien zuerst auf flohs blog.

Due to personal preference and a bit of coincidence, I ended up with a first-generation AMD Threadripper, complete with motherboard and cooler. I would describe the purchase price as “unignorable low”; I bought it second-hand via “Kleinanzeigen”.

A old Threadripper wasn’t my first choice, but it was up there. The power consumption, especially when idle, is said to be quite high, which is probably due to the multi-die design of these CPUs. On the positive side, there are plenty of PCI Express 3.0 lanes. The bundle consisted of the following components:

Processor

⧉ AMD

⧉ TPU

⧉ TPU

CPU Cooling

With an emphasis on appearance, the bundle included a powerful Wraith Ripper processor cooler from CoolerMaster. Only compatible with the TR4 socket, it was released specifically for the launch of the second generation Threadripper. I would have bought something flatter (and without RGB) from Noctua, BeQuiet, or Thermalright, but that’s just how it turned out.

⧉ CoolerMaster

⧉ CoolerMaster

⧉ CoolerMaster

Mainboard

I’m a little unhappy with this one – it’s turned out to be an RGB-overloaded Christmas tree, an AORUS Gaming 7 from Gigabyte. Not really the right choice for use in a rack case. From a technical point of view, it will work fine, but I’m turning the RGB off completely.

However, the VRMs are quite weak, so the 1950X (or its successor, the 2950X) is the best you can hope for. With the new UEFI version, second-generation Threadrippers, some of which have a significantly higher TDP (32 cores and 64 threads with the 2990WX -> 250 watts TDP), should also work in theory. Overclocking is not recommended with this board, with any Threadripper variant for TR4. The power supply and its cooling are too weak.

If I had had the choice, I would have gone with MSI or ASUS, as I usually do. Nevertheless, the Gigabyte board is still well suited for its intended purpose.

⧉ Gigabyte

⧉ Gigabyte

⧉ Gigabyte

Let’s move on to the other components, starting with the Memory.

Der Beitrag <h5>TR-Server #2: </h5><h3><b>The Base</b></h3> erschien zuerst auf flohs blog.

This otherwise really good but now outdated workstation notebook from Lenovo had major thermal problems right from the start. My model has the most powerful non-Extreme mobile processor available, an i7-4910MQ. This quickly becomes very hot with Lenovo’s cooling solution. An Nvidia Quadro K2100M is installed as a 3D accelerator, which is not exactly the most energy-efficient mobile graphics chip either.

Based on my experience with liquid metal in other previous projects, I thought, why shouldn’t this also help with this Thinkpad? No sooner said than done. I had taken the notebook apart many times before, and you really have to remove a lot of screws before you get to the cooler. Then, as usual, I made the preparations, applied thermal paste around the die for insulation because a notebook is mobile, and applied liquid metal*.

Better?

In short: no. The difference is a whole degree Celsius. The cooler is simply undersized. And I’m only talking about full CPU load (even without AVX!), which very quickly causes thermal throttling. If the GPU is also under load at the same time, the ThinkPad can no longer be used effectively. Even old games are unplayable because the system repeatedly throttles heavily and the notebook simply shuts down after a short time due to overload.

I am writing this post so that no one else takes this device apart and puts it back together again multiple times for no reason. New thermal paste* only helps if the existing paste has been in use for several years. But even then, you shouldn’t expect miracles. It’s a shame, because in terms of features, this is a really good 15″ notebook.

Der Beitrag Lenovo ThinkPad W540 – Liquid Metal erschien zuerst auf flohs blog.

As previously determined in the RetroBook project, the maximum Intel processor generation officially supported by Windows XP is the third generation, codenamed Ivy Bridge. This represents a slight improvement on Sandy Bridge, which was a really good step forward (clock speeds, performance per watt, AVX, SATA-III, PCI-Express 3.0, etc.).

So the next step is to decide which direction to go in. That’s easy to figure out: the most powerful Ivy Bridge processor available in the consumer desktop segment is the i7-3770K. It’s a quad-core CPU with hyperthreading, and the K indicates that it can be overclocked. After a quick search, you can find this processor used for a relatively low price. So that’s the one we’ll go with, right?

⧉ Intel

⧉ Intel

⧉ Intel

Great, that was quick…

…no, not really. I wanted to push the processor performance to its limits. There are also the Ivy Bridge-based workstation models, the most powerful of which is the i7-4960X. Intel’s naming convention is confusing; you would expect a Haswell processor because of the “4” in the name. It has 6 cores and 12 threads, which is 50% more of everything. This CPU is also overclockable, and the 15 MB L3 cache is a real improvement compared to the maximum 8 MB in the desktop segment with the i7-3770K.

At the same time, it’s important to note that Intel has only been shipping desktop processors from Ivy Bridge onwards with thermal paste between the heat spreader and the die, which causes numerous problems and almost requires delidding for good operation or meaningful overclocking, especially more than 10 years later. However, the workstation processors are still soldered, which is another plus point. The power consumption is not so good – 130 watts in its original state.

Furthermore, the workstation platform is designed as quad-channel DDR3, i.e., the maximum RAM expansion is 64 GB with 8 memory modules (8 x 8 GB) instead of the usual 32 GB with 4 modules (4 x 8 GB) in the desktop segment. The theoretical read and write throughput is also almost twice as high.

⧉ Intel

⧉ Intel

⧉ Intel

So it’ll be this one, it looks great, doesn’t it?

Theoretically, yes. However, after extensive research, I have chosen another candidate that deliberately flew completely under the radar for many years. This model was never officially available in retail stores, only in the form of pre-configured computers, more specifically: exclusively in Apple Mac Pro “Trashcan” models from 2013, as a configurable CPU. For a long time, this CPU was very expensive on the second-hand market because it was highly sought-after and rare. We are talking about the GOAT of the Ivy Bridge generation:

Intel Xeon E5-1680 v2!

Admittedly, it’s a somewhat unwieldy name. And with a Xeon, aren’t they clocked very low?
Yes, the clock speed is relatively low. More on that in a moment.

The technical data will make every computer enthusiast sit up and take notice, especially with the 2013 release, where four cores with high clock speeds were considered high-end:

• LGA 2011 socket, meaning it is also compatible with workstation motherboards in terms of hardware.
• 8 cores, 16 threads
• 25 MB L3 cache (!)
• For some reason, it can be overclocked!

It’s unbelievable, really. I’d love to know how it came about that the single socket versions of Ivy Bridge EP with the name E5-16XX v2 have an unlocked multiplier; even the smaller variants are unlocked. Either macOS has used some tricks with regard to power consumption or something similar, or someone at Intel forgot to lock the multiplier. Regardless of how or who – thank you.

⧉ Intel

⧉ Intel

⧉ Apple

Here is an overview of the differences between the i7-4960X and Xeon E5-1680 v2:

This processor will be a welcome rarity in this Windows XP retro PC. After keeping an eye out for deals, I was able to purchase one on eBay for €90 after a few weeks.

Der Beitrag <h5>Retro-XP-PC #2: </h5><h3><b>Four Ivy Bridge Cores?</b></h3> erschien zuerst auf flohs blog.

To get an overview of the ordering options that were available at the time and are therefore guaranteed to be compatible, the documents from HP are helpful:

The original service and maintenance manual for both models can be found here:
https://www.levnapc.cz/ProductsFiles/hp-elitebook-8540p-8540w-manual-en.pdf (archive.org)

On page 78, there is a nice list of supported processors:

⧉ HP

A decision must be made…

We see that, with the intended use in mind, a decision must be made as to whether the goal should be to maximize the performance of a single core (in this case, probably a dual-core CPU with two cores) or to have more cores (quad-core CPU with four cores).

The number of cores in a processor does not generally determine the maximum possible performance of a single computing core, but in this particular case, it makes sense to choose between these two options. It will always be a compromise, but Intel’s Turbo Boost mitigates this significantly, as we will see.

The i5 and i7 processors in this list all additionally support “hyper-threading,” meaning they can process four threads with two cores and eight threads with four cores. This increases efficiency and boosts performance.

But first, let’s take a look at what CPUs were actually available for this socket, because just because HP lists these specific models for release doesn’t mean that only these will work in the 8540p.

Does it make sense to look at other processor models from the same generation?

Support is generally determined by Intel’s code module, which is integrated into the BIOS by the manufacturer (in this case, HP). This is often compatible with all or most of a particular CPU generation, regardless of what the manufacturer configures/sells/equips.

In this case, Nehalem or Westmere as the overarching architecture, Clarksfield or Arrandale as the name of the core design. In this first generation of Core i processors, the whole thing is still a bit confusing; later, with a few exceptions, there were always only models within one architecture and one core design.

A good list can be found here, based on the chipset:
https://www.cpu-upgrade.com/mb-Intel_(chipsets)/QM57_Express.html (archive.org)

Now we have an overview of the possible options:

These are the most useful processors with two cores and four threads (dual-core with hyper-threading):

Designation Lithography	Data
i7-640M (SLBTN) Arrandale (32 nm)	Most powerful dual-core, with HP support Base clock: 2.80 GHz Cores / max. Turbo-Boost: 1 – 3.46 GHz 2 – 3.20 GHz Caches: L1 : 128 KB L2: 512 KB L3: 4 MB Possible RAM clocks: 800 MHz or 1066 MHz TDP: 35 Watts Integrated HD Graphics (Ironlake), uninteresting due to dedicated GPU in the 8540p, but potentially important for other notebooks!
i7-620M (SLBPD – C2 | SLBTQ – K0) Arrandale (32 nm)	Base clock: 2.66 GHz Cores / max. Turbo-Boost: 1 – 3.33 GHz 2 – 3.06 GHz Caches: L1 : 128 KB L2: 512 KB L3: 4 MB Possible RAM clocks: 800 MHz or 1066 MHz TDP: 35 Watts Integrated HD Graphics (Ironlake), uninteresting due to dedicated GPU in the 8540p, but potentially important for other notebooks!

If we want a powerful processor with four cores and eight threads (quad-core with hyper-threading) to be more flexible with applications and have more reserves, these models are worth considering:

Designation Lithography	Data
i7-940XM (SLBSC) Clarksfield (45 nm)	Most powerful quad-core Base clock: 2.13 GHz Cores / max. Turbo-Boost: 1 – 3.33 GHz 2 – 3.20 GHz 3 – 2.40 GHz 4 – 2.40 GHz Caches: L1 : 256 KB L2: 1 MB L3: 8 MB Possible RAM clocks: 1066 MHz or 1333 MHz TDP: 55 Watts No integrated graphics!
i7-920XM (SLBLW) Clarksfield (45 nm)	Base clock: 2.00 GHz Cores / max. Turbo-Boost: 1 – 3.20 GHz 2 – 3.06 GHz 3 – 2.26 GHz 4 – 2.26 GHz Caches: L1 : 256 KB L2: 1 MB L3: 8 MB Possible RAM clocks: 1066 MHz or 1333 MHz TDP: 55 Watts No integrated graphics!
i7-840QM (SLBMP) Clarksfield (45 nm)	Most powerful quad-core with HP support Base clock: 1.86 GHz Cores / max. Turbo-Boost: 1 – 3.20 GHz 2 – 2.93 GHz 3 – 2.00 GHz 4 – 2.00 GHz Caches: L1 : 256 KB L2: 1 MB L3: 8 MB Possible RAM clocks: 1066 MHz or 1333 MHz TDP: 45 Watts No integrated graphics!
i7-820QM (SLBLX) Clarksfield (45 nm)	Base clock: 1.73 GHz Cores / max. Turbo-Boost: 1 – 3.06 GHz 2 – 2.80 GHz 3 – 1.86 GHz 4 – 1.86 GHz Caches: L1 : 256 KB L2: 1 MB L3: 8 MB Possible RAM clocks: 1066 MHz or 1333 MHz TDP: 45 Watts No integrated graphics!

Digression: Model selection

A general note on the topic of “always choosing the largest model”:
The mobile Extreme series (identified by “XM”) has a TDP of 55 watts, which is 10 watts more than the maximum specified by HP.
This may or may not work. Furthermore, these models are not included in HP’s support list.

In addition, these models (i7-920XM and i7-940XM) can be overclocked, which is not permitted by HP’s BIOS and therefore offers no added value. At the same time, these models in particular are often many times more expensive when purchased second-hand for various reasons, simply because they are the “largest” models (“Extreme!”). Marketing is “a hell of a drug.”

I always recommend looking at several models close to the top of the performance range and comparing them.
I can think of an example when I upgraded the first-generation dual-core i3 processor in an Acer desktop PC (which now serves as a workshop PC, more on that in another post):

The largest supported processor would have been an i7-880, but it was virtually unavailable anywhere and was almost prohibitively expensive when new. The next smallest model would have been an i7-870 – at the time, this cost well over €40 on the usual platforms (it is now also available at a reasonable price).

I ended up with an i7-860. Why?
€12 including shipping. Performance?

A ridiculous difference. Saved quite a bit of money back then without really losing anything technically.

One should never forget that everything that is upgraded or retrofitted and is of this age is only touched for “reasons,” be it nostalgia, and that it is rarely financially worthwhile. Triple the price for 4% would be an example of irrationality.

Let’s move on to the processor selection:

Returning to the 8540p, these two models were options for me due to their reasonable used price at the time:

These two are actually from two different architectures and were manufactured using two different lithography processes. Both are on HP’s support list, which is good.
What is the difference in performance? It must be significant, right?

I’ll sum it up with “Wow, the dual-core isn’t really any faster on one core, which is unexpected“.
Of course, that’s only half the story, as you can see from the table above—the quad-core model slows down significantly more under increasing load and consumes 10 watts more than the dual-core.
Nevertheless, I’ve made my decision, especially after looking at the meaningful individual performance in Cinebench R10:

In our case, it will be the i7-840QM. For me, double the number of cores is a good trade-off for a minimal loss in single-core performance. Especially since four cores are now the low-cost entry-level standard, and more cores and threads can only be beneficial for modern software. Nevertheless, it is remarkable how well the i7-620M performs in comparison in the multicore test; the improvement or reduction of the lithography process from 45 nm to 32 nm is having an effect.

Apart from that, the more powerful quad-core processors have twice as much L3 cache with 8 MB, which greatly increases performance in some situations. Not to be forgotten is the higher possible RAM clock speed (1066 MHz maximum for the dual-core processors vs. 1333 MHz for the quad-core processors).

For the sake of completeness, here is a comparison of the then-flagship i7-940XM with the i7-840QM:

It’s funny how the TDP increase from 45 watts in the i7-840QM to 55 watts in the i7-940XM can be converted almost directly into multi-core performance, as 10 watts more corresponds to 22.22%, meaning that performance scales almost exclusively through more energy.

The 10 watts “saved” regardless of price will become important later on when the thermal and electrical power budget of the 8540p becomes tighter.

Detailed view and notes:

Until Intel’s fourth generation of mobile processors (Haswell), there were both soldered (BGA – Ball Grid Array) and socketed models (PGA – Pin Grid Array). Since then, only soldered models have been available for mobile devices, which makes upgrading or retrofitting nearly impossible (only with specialized, expensive BGA equipment) and very risky.

Here you can see the two first-generation CPUs from the table above.
The pins and the corresponding socket are also visible. Note the golden arrow on the CPU, which indicates the installation direction (there is also an arrow on the socket at the bottom left, which is barely visible).

⧉ Intel

⧉ Intel

⧉ Intel

General conclusion:

For anyone who is solely interested in single-core performance that should be sufficient under all (realistic) circumstances (i.e., even under higher loads), I highly recommend the i7-640M dual-core processor with 35 watts TDP, provided it is available at a reasonable price. This “saves” another 10 watts compared to the quad-core processors, which also benefits the GPU thermally.

For notebooks without a dedicated GPU, this is the best possible CPU of this generation anyway, thanks to its integrated graphics.

With the quad-core processors, the price is basically the deciding factor, although I would consider the i7-840QM I chose to be the maximum, as 45 watts is already close to or above the thermal limit in combination with a powerful GPU.

Don’t forget:
If the processor is to be replaced, the thermal paste on the CPU and GPU must be replaced.
I recommend Arctic MX-4*, which I have had good experience with over many years:

⧉ Arctic

Let’s move on to the memory (RAM).

Der Beitrag <h5>RetroBook 8540p #3: </h5><h3><b>Upgrade – CPU</b></h3> erschien zuerst auf flohs blog.

The following procedure has proven to work well, regardless of the newly built system:
First, check whether there are any BIOS/UEFI updates available from the motherboard manufacturer, in my case ASUS—there was a newer version (v1501). I installed this using a FAT32-formatted USB stick. Even if there had been no newer version, continue with:

Set everything in UEFI to factory settings and then to minimum settings, i.e.:

• RAM at minimum clock speed, i.e., simply leave the existing setting (presumably 2133 MHz for DDR4) unchanged.
• Processor settings all “original”
• Prepare Windows—disable CSM/Legacy, enable TPM, boot order, enable Secure Boot if desired, etc.

First, install a fresh copy of Windows, if you don’t already have one, to rule out software problems later on and avoid unnecessary background processes and legacy issues that could interfere with overclocking. I installed Windows 11 Pro 24H2, bypassing the “incompatibilities.”

Required software

Once Windows is installed, you can let the system install the drivers itself via a working Internet connection—that’s the easiest way. Then install the updates offered and restart the whole thing. Next, check in Device Manager to see if drivers have been installed for everything relevant—usually they have.

However, I always take the “old-school” approach:
Before I even install Windows, I download the latest drivers for the components I am using from the manufacturer’s websites (in this case, ASUS, Nvidia, and Intel). Then I install Windows offline and, once that is complete, I install the drivers I downloaded earlier, also offline.

This has the advantage of usually resulting in newer driver versions than those offered by Windows Update. Sometimes the manufacturer’s websites also contain additional information, such as notes on current or past problems, etc., which you would never see if you used the automatic method.

I used the following software to overclock the CPU and GPU:

• HWiNFO – For monitoring
• Prime95 – For load generation, optionally with or without AVX / AVX2
• Cinebench R23 – For generating faults and testing differences
• ASUS ROG RealBench – To generate load in order to test stability
• Unigine Heaven – To generate load on the GPU in order to test stability
• MSI Afterburner – For overclocking the GPU
• MemTest86 – To generate faults on the RAM and memory controller in order to test stability

My goal is to achieve stability in ASUS ROG RealBench, as I believe it reflects moderate to heavy everyday use equally well.

Processor overclocking

Actually, the whole thing isn’t complicated, but you need to have a lot of time and patience.

I first overclocked the i7-7700K statically, ensured stability, and repeatedly pushed it to its limits (voltage, temperatures). Since my goal was 5.00 GHz on all four cores with hyperthreading enabled, I started with that—the delidding I had done beforehand should have made this feasible in terms of temperature.

So I simply set the core voltage to 1.30 volts in the BIOS and set the “CPU Core Ratio” for “All Cores” to “50”. Then I saved, booted Windows, and started HWiNFO / Prime95 with AVX and AVX2 and let it run.

First observation: It runs stably for at least 1-2 hours, but gets quite warm (around 80-85 °C).
Then back into the UEFI and down with the voltage, now to 1.25 volts – then tested again.
Temperatures have dropped to just under 80 °C, still everything stable.

I did this until I reached 1.18 volts, at which point I had my first “quick” blue screen.
So I increased it again to 1.22 volts. It then ran for many hours, stable with acceptable temperatures (still ~ 80 °C). Without delidding, even that would have been either barely possible or not possible at all; I’m at the limit.

Next, I ensured stability with ROG RealBench, which is a realistic everyday measure for me; it has to be stable in the end. It was, it runs smoothly and, of course, cooler; I saw a maximum of 75 °C on the hottest core.

Great, I would be satisfied with that in principle.
But now, of course, I want to know what will limit me in the end:
The cooling or the voltage?

How far will it go?

Then I dared to increase the clock speed, setting the multiplier to “51,” i.e., 5.10 GHz. I also adjusted the voltage to 1.30 volts to match. Lo and behold: Prime95 with AVX quickly forces the cooler to its knees thermally, the first thermal throttling kicks in, and later the first blue screen appears (well over 100 °C!).

So I continued with 1.28 volts. Same result, thermal throttling—but later. No blue screen.
But you can already see the limit with AVX, 5.00–5.10 GHz. Then the air cooler, which otherwise does a really good job and is also tolerably quiet, reaches its limit.

After only a few minutes of ROG RealBench, a blue screen appears.
The journey with AVX seems to be over here, from now on I’ll optimize the system for “stable everyday use with a little AVX.”

And I’ve still come a long way:

The final result:

First of all: I made it very easy for myself with the RAM. I simply loaded the stored XMP profile for 2400 MHz 16-16-16-39 at 1.20 volts, slightly increased the DRAM voltage to 1.2540 volts, and set the memory frequency to 3000 MHz, nothing else!

The whole thing ran for days with Memtest86 without any errors. That was a 25% performance boost for free.
Above 3000 MHz, errors occurred that would have required very time-consuming fine-tuning to fix, but that wasn’t my focus.

For the processor, I switched from static overclocking at 5.20 GHz with over 1.35 volts to adaptive overclocking. 5.30 GHz on all cores got too hot – it took a few minutes, but throttling kicked in.

In the end, of course, I spent hours experimenting and days testing to see what worked best and how close I could get to the limits.

The CPU now runs as follows:
Load on 1–2 cores = 5.30 GHz at LLC 6 with adaptive auto voltage, slightly above 1.42 volts in HWiNFO under load.
Load on 3–4 cores = 5.20 GHz at LLC 6, also adaptive, then 1.385 volts under load.

Cache = 4.90 GHz, above that it quickly became very unstable.

Without AVX, I am therefore at both the temperature and voltage limits; it would not be possible to go any further without entering even more dangerous voltage regions. I find 1.42 volts to be borderline; my limit was actually 1.40 volts.

I couldn’t use an AVX offset without the turbo and power-saving mechanisms only working halfway afterwards, or the single-core clock speeds never being reached, as Windows 11 seems to use AVX instructions quite often. I wanted to achieve a dynamic overclock that is also capable of downclocking in idle mode to save heat and power – hence the use of an adaptive “auto” offset in the first place.

Conclusion:

This inexpensive cooler packs a punch – without AVX, I can push the CPU to almost its maximum capacity. At the same time, the board’s power supply is so good that I have never experienced any problems in this regard (I didn’t have any additional airflow over the VRM coolers!).

I would describe the clock speeds as very high, achieved at the expense of equally high voltage. Here is a comparison of the changes:

Cinebench and PassMark results:

First, Cinebench R23 with single-core overview – 1423 points.
Then with multi-core overview – 7112 points.

Finally, Cinebench R24 – 86 points or 419 in multi-core.

When compared to other CPU models, it is easy to see how far the i7-7700K has risen, especially in terms of single-core performance. Temperatures were consistently around 90 °C while Cinebench was running.

Click here for the PassMark PerformanceTest v11.1 Baseline.

What about the graphics output?
We’ll continue with that in the next post.

Der Beitrag <h5>i7-7700K OC #7: </h5><h3><b>Now the overclocking can finally begin</b></h3> erschien zuerst auf flohs blog.

I consider delidding the processor to be one of the most important steps in this project.
EKWB describes the whole process well: EKWB (archive.org)

Firstly, you generally achieve much lower temperatures even with original clock speeds, and secondly, this greatly increases the scope for overclocking. At the same time, power consumption is reduced and the cooling system has to work less hard.

In my situation, it makes sense to do this first—before purchasing other necessary hardware. There is always the risk that the CPU will be irreparably damaged during delidding and the whole project will go down the drain, so it’s better not to buy anything that you won’t be able to reuse later.

What is required?

There are several different delidding tools available from various suppliers to minimize the risk of damage. I printed a tool myself and also ordered a very inexpensive one from China*, and ultimately used the Chinese one (die-cast aluminum, more stable than PLA). To securely but reversibly and heat-resistantly reattach the heat spreader after delidding, I used “Dirko HT*” from Elring, which is very well known in the automotive industry. I used this 3D-printed mold to press the parts together.

Ultimately, the inferior thermal paste used by Intel ex works should be replaced with something much more efficient. Liquid metal* is really the only option if you want to get the most out of it:

⧉ Thermal Grizzly

⧉ Elring

This will also be my first project using liquid metal, which has interested me for a long time—there just hasn’t been a really useful application for it until now, and it’s not exactly cheap. That’s about to change.

Let’s go…

First of all, it should be noted that you should proceed with extreme caution, as there are many small, sensitive SMD components on the underside of the CPU that could potentially be damaged.

With the necessary care, place the processor in the delid tool and screw it in by hand until the movable part slightly clamps the heat spreader. At this point, you should check carefully that the CPU is inserted correctly (the right way round!) and has enough space around it, then insert the Allen key provided into the screw head and continue screwing the screw into the tool. It should be difficult at first and then suddenly easy. When this point is reached, unscrew the screw, remove the slider again, and remove the CPU.

Don’t keep screwing when it got easier!

If you now carefully press between the heat spreader and the circuit board with your fingernails or a plastic spatula, the heat spreader should be easy to remove.

⧉ bit-tech

⧉ TechPowerUp

Cleaning:

Now everything should be cleaned thoroughly. The black adhesive used by Intel, which is probably silicone-based, is stubborn. It is best to scrape it off the heat spreader with a plastic or wooden object. Of course, you should use something that is softer than nickel-plated copper (the material the heat spreader is made of) to avoid scratches or similar damage.

On the circuit board side, you have to do the same thing, but with extra care.
I scraped it off very carefully with my fingernails and “rubbed off” the remaining residue with isopropanol on a paper towel.

The thermal paste can be removed as usual with isopropanol*, nothing special.
If the result looks like this, everything has gone well:

⧉ PCGH

Isolation:

Now the processor must be prepared for use with liquid metal. This means that all potentially conductive surfaces must be insulated, either with (nail) polish* or Kapton tape*.

I found polish to be safer and therefore coated all marked surfaces with nail polish:

In retrospect, I would say that insulating the four contacts (marked in green) near the die with Kapton would have been sufficient. Since this was my first project with liquid metal, I proceeded with excessive caution.

Applying liquid metal:

Now comes the actual application of the liquid metal.
It is supplied in a syringe, together with cloths and a kind of special “cotton swab.”
First, the processor is placed on a soft, flat surface and a small “ball” of liquid metal is applied to the center.

This is then spread over the die with the cotton swab using circular movements and a little pressure until it is evenly distributed.

I then spread a really thin line of Elring Dirko HT sealant over the heat spreader.

⧉ Phiarc

⧉ TechPowerUp

Assembling the CPU:

Now you can reassemble the CPU. I used the 3D-printed tool linked above for this. I used it as intended, but in addition to the clamp, I pressed the whole thing a little more in a small vise* than the clamp was able to do. In principle, however, many things work, even a small vise with rubber protectors should do the trick.

The whole thing has to be pressed together with force over a longer period of time so that the adhesive layer dries as thinly as possible to achieve a minimal gap between the heat spreader and the circuit board.

The crazy folks at TechPowerUp even tested this with superglue*, which hurts (irreversibly!).

The whole thing should now be set aside for a few hours to dry, rather than too little time.

Did the CPU survive?

At that point, I only had Gigabyte’s significantly undersized B250M motherboard with the appropriate 1511 socket at hand. So I had to use it again for testing, and lo and behold:

The processor is still running. Perfect.

I then aligned a 120mm fan* with the VRMs on the motherboard and let it blow at full speed while the CPU was “burned in” for hours with Prime95 Small FFTs with AVX and AVX2.
As usual, I used MX-4 from Arctic* as thermal paste between the heat spreader and the cooler.

Of course, despite the fan, the motherboard throttled the processor down to below the base clock every 10 minutes after about half an hour, but I now knew what I wanted to find out:
The processor runs stably and, even with the reused, similarly undersized Alpenföhn Sella, is easily 20 °C cooler than before. Even under these very unfavorable circumstances, there is no longer any thermal throttling of the CPU!

I read this several times during my research; temperature reductions of 10 to 30 °C (when overclocking) are supposed to be possible. But seeing it live was something else entirely—I had never seen such an extreme improvement before and have never seen it again since, and since more modern processors inherently reach their thermal and performance limits, I probably never will again.

What are you doing, Intel?

The real insight is that Intel (intentionally?) used cheaper thermal paste instead of more expensive indium solder in production to either limit these CPUs or to achieve the effect over time that they become hotter and slower because the thermal paste dries out and will probably even lose contact between the die and heat spreader in some cases. Since “Ivy Bridge,” thermal paste has been used instead of solder, up to and including “Kaby Lake,” the generation from which this i7-7700K also originates.

The FAT and SLIM models of Sony’s Playstation 3 were technically given a best-before date in the same way (thermal paste between the heat spreader and die on the CPU and GPU). The Super Slim is then Direct-Die, presumably for space reasons.

This project was a complete success, and now suitable hardware is being sought to adequately overclock this CPU.

Der Beitrag <h5>i7-7700K OC #3: </h5><h3><b>Dangerous things first</b></h3> erschien zuerst auf flohs blog.