Four Screws Fasten NVIDIA GTX 1070 Dust Cover

I recently took apart three external hard drives to retrieve their bare drives within to use in an internal application. In all three cases, there were no externally accessible fasteners to help with disassembly. I had to pop plastic clips loose, breaking some of them. For laughs, I thought it’d be fun to talk about a time when I had the opposite problem: I was confronted with far too many screws, most of which weren’t relevant to the goal.

I have a NVIDIA GTX 1070 GPU and it had been in service for several years. NVIDIA designed the cooling shroud with a faceted polygonal design. Part of it was a clear window where I could see dust had accumulated through years of use. I thought I should clean that up, but it wasn’t obvious to me which of the many visible screws held the dust cover in place. The visual answer is in this picture:

In case these words are separated from the illustrative image, here is the text description:

Starting from the clear window, we see four fasteners closest to them. These four fasteners hold the clear plastic to the dust cover and not useful right now. We need to look a little bit further. There are two screws further away between the clear window and the fan. Those are two of the four dust cover screws. The remaining two are on the top and bottom of the card, closer to the metal video connector backplate. Once these four screws are removed, the dust cover can slide sideways (towards the backplate) approximately 1cm and then the dust cover can lift free. After that, the four screws holding the clear window can be removed to separate the window from the rest of the dust cover.

In the course of discovering this information, I had mistakenly removed a few other screws that turned out to be irrelevant. Since they’re not useful to the task at hand, I put them back without digging deeper into their actual roles.

Goodbye and Good Riddance To This Generation of Hybrid Drives

I have successfully upgraded a HP Pavilion Split X2 (13-r010dx) to use a modern M.2 SATA SSD with the help of an adapter circuit board. The results were stellar and, even though the adapter does not mount inside the computer properly, I have no plans to put the original drive back in. It was a compromised solution of its era and times have moved on.

A little over ten years ago, flash-based solid state drives started edging towards price levels affordable to normal computer buyers. A few bargain basement devices were released, and they were terrible. I had two JMF602-based drives that lasted through their 90-day warranty periods but not much beyond that. The big breakthrough in affordable and durable performance is credited to the Intel X25-M, but the capacity was quite small. SSD performance is something that I had to experience to be converted to a fan, and it was hard to get non-converts to accept living with 80GB of capacity when we had become accustomed to hundreds of gigabytes.

The compromised solution was the SSD/HDD hybrid drive: there is a magnetic platter hard disk drive, but there is also a small piece of flash memory acting as a small solid-state drive cache. The advertising proclaimed that we would get the capacity of a HDD with all the performance of a SSD. I thought the concept was enticing, but never actually got one to try.

I’m glad I did not, if this computer’s stock WD5000M21K hybrid drive is representative of the breed. Its performance was absolutely terrible. Maybe modern workloads overwhelmed its meager 8GB of flash cache. Maybe years of use has worn out the flash and there was no caching anymore. Whatever the reason, its performance was no better than a HDD, definitely nowhere near the performance of a real solid state drive.

Now solid state drives with plenty of elbow room are quite affordable, giving old computers a new lease on life. The hinderance of oddball connectors like the SFF-8784 become just a speed bump with help of adapters, so we don’t need to put up with the compromises of SSD/HDD hybrid drives anymore.

No Good SSD Adapter Mounting Option

Upgrading to a SATA SSD was a hugely successful transformational upgrade for this old HP Pavilion Split X2 (13-r010dx). Physically, it was only a tiny improvement to the problems of heavy weight and it doesn’t nothing to help the low 1366×768 resolution screen, but the computer is now immediately responsive to user input and no longer miserable to use. That is the good news.

The bad news is that Sintech’s SSD adapter didn’t have mounting holes to line up to original mounting brackets. I originally thought I could 3D-print something to adapt its vertical mounting holes to horizonal, a simple rectangular loop should do the trick. Then I took a closer look at the dimensions and realized it’s not so simple.

The circuit board height where the interface plugs into is fixed and that height is in the middle of the screw holes. In order to clear mounting screws I would have to cut into the circuit board, which I’m not prepared to do. I already had to wait for a replacement to be shipped to me once, I didn’t want to ruin a board and have to wait again.

I considered offsetting vertically to clear the screw holes, moving in one direction or the other. But the overall height of this adapter + M.2 socket is barely any thinner than the original hard drive, leaving very little margin to work with. Pushing towards the screen, I could not move far enough to clear the screws. Pushing towards the back, clearing the screws would bump against the back cover.

So no 3D-printed adapter bracket for me. I ended up using double-side tape sold for attaching Raspberry Pi heat sinks, and used that to attach the SSD to the chassis. This solution is not ideal, but I’m not willing to revert to the stock hard drive because it was something better left to the past where it belongs.

HP Split X2 (13-r010dx) Transformed with SSD Upgrade

With a SSD adapter circuit board temporarily held in place with painter’s tape, I powered up this old HP Split X2 (13-r010dx) to see if it’s willing to run on a modern M.2 SSD. The answer is yes — and quite well!

I have a USB flash drive prepared by the Windows Media Creation Tool for installing Windows 10 2004, and I could select it as the boot device by holding down F9 upon power-up. From there it was an uneventful Windows 10 installation, automatically activated by a Windows 8 license embedded in hardware.

Here is the “Before” picture, taken a few minutes after the computer has booted up on the original WD5000M21K hard drive. The computer is completely saturated with Windows startup tasks and it takes a few minutes to even get Task Manager up and running and a screenshot tool. From here we can see we’re doing well on memory, with only half used. The CPU is largely idle. They are all waiting for the disk.

Here is the “After” picture, taken shortly after the computer started up for the first time running Windows 10 freshly installed on the M.2 SSD. Disk overhead has stopped being a constraining factor. The memory is about the same, and now the humble low power Core i3 CPU is the constraining factor as this computer is chewing through information to decide what to download from Windows Update.

Once Windows was up to date, all drivers were automatically installed, and this computer was ready to go. A reboot verified that startup time has gone from several minutes down to less than 30 seconds. Launching and switching between applications are nearly instantaneous. When the meager 4GB RAM starts running low, virtual memory on the SSD is perfectly usable and not the molasses slow torture session it used to be.

The Core i3 might be the low end of the Core line, but it is far faster than an Atom chip of similar vintage. Freed from the shackles of its molasses slow stock HDD, this computer is now perfectly comfortable running up-to-date Windows 10 and modern applications. This SSD upgrade has proven to be a hugely beneficial transformation for this old computer. Which was fantastic! Except for one problem… how do I make sure it doesn’t rattle around inside the machine?

HP Split X2 (13-r010dx) SSD Upgrade: Round 2

I now have a M.2 SATA SSD available for experimentation, mounted on a Sintech ST-NG8784 adapter circuit board that lets me plug a M.2 SSD into a SSF-8784 connector. This unusual slim connector is used by a HP Pavilion Split X2 (13-r010dx) tablet/laptop convertible computer, which foiled an earlier attempt at SSD upgrade. This time I am better prepared.

Here’s the “Before” picture, with the stock SFF-8784 hard drive in the center. From the factory the interior of this device had a lot more tape and foil, including foil completely wrapping the hard drive. They’ve all been pulled off in earlier adventures.

In addition to disconnecting the AC adapter, the connector at the center of the image (with black wires, not white ribbon cable) is for the battery and should be disconnected before working on this machine.

Four screws held the drive in place using some metal brackets, which were in turn mounted to the drive via some screws on the side. Removing them were trivial, but that exposed the next problem: the PCB could match bottom mounting holes, but we need horizontal ones, so there’s no good way to fasten the drive.

I decided not to worry about proper fastening for the moment, because I had no idea if the computer would even accept this drive with adapter. Some temporary painter’s tape is enough to make sure the board doesn’t flop around while I experiment.

Examining Sintech M.2 to SFF-8784 SATA Adapter (ST-NG8784)

It took a second try before I received an adapter card that looked good enough to proceed. The objective of the exercise is to put a common M.2 2280 SATA SSD into an old HP Pavilion Split X2 (13-r010dx) convertible laptop which came with an unusually thin (5mm) spinning platter hard drive in the super rare SFF-8784 form factor. The form factor foiled my first attempt at an SSD upgrade for this computer, but now I have a M.2 SATA SSD available for experimentation and now I have the adapter card I bought (*) for the project.

This card is made and sold by Sintech, which has a product page for this item where I learned it is designated model ST-NG8784. I was fascinated by how simple the adapter is. There are only a few surface mount components and very few traces on the circuit board. C1 and C2 are obviously capacitors, but I’m not sure what U1 is. Searching on “84-33 2012DC” didn’t result in anything enlightening, but by its general shape and arrangement of nearby capacitors I guess it is a voltage regulator.

The M.2 connector has many, many pins but the SFF-8784 plug has significantly fewer, resulting in a superficially simple layout. I guess that makes sense, after all the S in SATA stands for Serial, so it wouldn’t need many pins to do its thing. I count just two differential pairs on top for data. Most of the other connections are either power or ground. But it does highlight the fact there is no active signal conversion on this adapter: this would only work for SATA M.2 SSDs and I would not expect it to work with NVMe M.2 SSDS.

Mechanically, this adapter card has provisions for several of the popular M.2 card lengths. A threaded standoff has been press-fit into the spot corresponding to the longest supported size M.2 2280. If the user has a SATA SSD in one of the shorter form factors, there is a small Ziploc bag with a screw-on standoff to be installed in the appropriate slot. Since my M.2 SATA SSD is in the 2280 format, I did not need the Ziploc bag. I installed my SSD into this adapter and turned attention to the laptop.


(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

SSD Upgrade Project Delayed By Shipping Damage

I’ve been aware that the performance of an old HP Pavilion Split X2 (13-r010dx) is constrained by its hard drive to some degree. But when I tried to remove that constraint by upgrading it to a commodity SATA SSD, I found that it did not use the connector type I was familiar with. Rather, it used a much thinner and rarer variant called SFF-8784. Native SFF-8784 drives are expensive due to their low volume, but I found an Amazon vendor selling SFF-8784 adapter circuit boards to use mSATA or M.2 SATA SSDs. I resolved to come back to this project later, when I have a spare M.2 SATA SSD to try.

It is now later. Thanks to some end-of-year sales, my computers received upgrades and the cascade of hand-me-down freed up a M.2 SATA SSD for this experiment. I proceeded to order the M.2 to SFF-8784 adapter board I found earlier (*) eager to see how it might improve the old HP’s responsiveness.

Unfortunately, the first adapter arrived damaged. It was shipped in an anti-static bag and enclosed in a padded envelope. The padding was apparently not enough, because the M.2 connector was crushed out of shape. I doubted it would accept a M.2 SATA SSD and I didn’t want to risk a perfectly functioning SSD to try.

I contacted Sintech and they sent a replacement. When the replacement arrived, I noticed a modification. It was still in an anti-static bag in a padded envelope, but there was the additional padding of a block of pink foam to protect the M.2 connector.

With the help of this pink foam block, the onboard M.2 connector survived shipping and looked good enough for this SSD upgrade project to begin.


(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Seeed Studio Odyssey X86J4105 Has Good ROS2 Potential

If I were to experiment with upgrading my Sawppy to ROS2 right now, with what I have on hand, I would start by putting Ubuntu ARM64 on a Raspberry Pi 3 for a quick “Hello World”. However, I would also expect to quickly run into limitations of a Pi 3. If I wanted to buy myself a little more headroom, what would I do?

The Pi 4 is an obvious step up from the 3, but if I’m going to spend money, the Seeed Studio Odyssey X86J4105 is a very promising candidate. Unlike the Pi, it has an Intel Celeron processor on board so I can build x86_64 binaries on my desktop machine and copy them straight over. Something I hope to eventually be a painless option for ROS2 cross compilation to ARM, but we’re not there yet.

This board is larger than a Raspberry Pi, but still well within Sawppy’s carrying capacity. It’s also very interesting that they copied the GPIO pin layout from Raspberry Pi, the idea some HATs can just plug right in is very enticing. Although that’s not a capability that would be immediately useful for Sawppy specifically.

The onboard Arduino co-processor is only useful for this application if it can fit within a ROS2 ecosystem, and the good news is that it is based on the SAMD21. Which makes it powerful enough to run micro-ROS, an option not available to the old school ATmega32U4 on the LattePanda boards.

And finally, the electrical power supply requirements are very robot friendly. The spec sheet lists DC input voltage requirement at 12V-19V, implying we can just put 4S LiPo power straight into the barrel jack and onboard voltage regulators will do the rest.

The combination of computing power, I/O, and power flexibility makes this board even more interesting than an Up Board. Definitely something to keep in mind for Sawppy contemplation and maybe I’ll click “Add to Cart” on this nifty little board (*) sometime in the near future.


(*) Disclosure: As an Amazon Associate I earn from qualifying purchases.

Samsung 500T Disappointments

I pulled out my old Samsung 500T to see if it could run ESA’s ISS tracker. It could, and quite well, so I think the role might be the best use of this machine. Because it has proven to be a huge disappointment in so many other ways.

I knew of the machine when it launched alongside Windows 8. This was when Microsoft also launched the original ARM-powered Surface tablet to demonstrate what’s possible by using ARM chips: look at how thin and light they are! Samsung & friends launched the 500T as counterpoint: just as thin and light as the Windows RT tablets, but with an Intel CPU for full x86 compatibility. Judging by the spec sheet, it was a spectacular device to undercut and humiliate the “Windows on ARM made thin and light possible” story.

But that’s only on the spec sheet and not on the price tag. The 500T was expensive and Surface tablets sold for far less. Due to that fact, I didn’t get my 500T until much later when it showed up as a secondhand refurbished unit on Woot. When it arrived, I was sure there was something wrong with the machine. Maybe it somehow slipped past testing in refurbishment? It was completely unusable, taking several minutes to boot and user commands took many seconds (5-10) to respond or were ignored entirely. I went online and found a firmware update, which took all night to apply and upgraded performance from “disastrous” to “merely horrible”.

The screen was another cool feature that didn’t panned out. Not just a touchscreen for fingers, it was also a pen digitizer. Compatible with passive Wacom stylus used by the much thicker Surface Pro tablet, the 500T also had a tiny stylus holder built in. It held the promise to be a digital sketchpad with pressure sensitivity making it superior to a contemporary iPad. But slow system response killed that dream. Who wants to sketch on a pad when strokes don’t show up until a few seconds after we draw it?

Judging by Windows Task Manager, this device’s major implementation flaw was its eMMC flash storage, constantly showing 100% activity. The Atom CPU was not exactly a stellar performer, but it wasn’t the reason for delay as the 500T was constantly waiting to read from or write to storage. Generally ruining user experience across the board.

Not to let Intel entirely off the hook, though, as its Atom Z2760 CPU turned out to be a long term liability as well. This CPU was part of Intel’s Clover Trail family, and they had problems running Windows 10 features newly introduced in 2016. Intel had discontinued that line and declined to do anything about it, so Microsoft blocked Clover Trail devices from advancing beyond Windows 10 build 1607. They will still receive security fixes until January 2023, but features are stuck at July 2016 levels forever.

All of the above are things that I might be able to overlook as unfortunate result of things outside Samsung’s control. The eMMC storage might have performed well when new but degraded with time as solid state storage sometimes do. (The TRIM command could help, but they had to make use of it.) And Samsung had no way of knowing Intel would just abandon Clover Trail.

But let’s talk about what Samsung have chosen to install on the machine. As is typical of machines around that age, there was the usual useless bucket of trials and discount offers. There are also Samsung features that duplicate existing Windows functionality, others thinly veiled advertisement for Samsung products, and more. The worst part? I could not get rid of them. I thought they would be gone once I wiped and installed Windows 10, but they were bundled with critical device drivers so I had no choice but to reinstall them as well. Holding device drivers hostage to force users to accept unrelated software is consistent with Samsung’s anti-user behavior I saw across the board.

The image at the top of this post is just one example. SWMAgent.exe appears to be some sort of Samsung software update engine (What’s wrong with Windows Update?) and it asks for elevation. If the user declines to grant elevated privileges, the software is supposed to respect that choice and go away. But not Samsung! We see a black border visible around that dialog box, which might look strange at first glance. Windows 10 adds a subtle dark shadow to dialog boxes, why this ugly black thing? It is because we’re not looking at a single SWMAgent.exe dialog box, but a huge stack of of them. Each popping on top of the last, and another one added every minute or so. The thick black border is a result of that subtle dark shading stacked deep and combining, because Samsung would not take no for an answer.

I don’t need that in my life. The upside of this machine being disappointing was that I had no motivation to put up with it. Into the unused bin it went, and I haven’t bought a Samsung computer since.

HP Stream 7 Power Problems

I wanted to see if I can employ my unused HP Stream 7 as an International Space Station tracker at home, displaying ESA’s HTML application. The software side looks promising, but I ran into problems on the hardware side. Specifically, power management on this little tablet currently seems to be broken.

The first hint something was awry is the battery runtime remaining estimate. It is unrealistically optimistic as shown in the screen image above: 46% battery may run this little tablet for several hours, but there’s no way it would last 4 days 4 hours. At first I didn’t think it was a big deal. Battery-powered devices that I’ve dusted off would frequently give wildly inaccurate initial readings on battery. It is common for power management module to require a few charge-discharge cycles to re-calibrate.

In the case of my tablet, a few battery cycles did not help. Battery estimates remained wildly inaccurate after multiple cycles. But I was willing to ignore that estimate, since battery life is not a concern in a project that intends to run tethered to power around the clock. The bigger problem was the tablet’s behavior when plugged in.

HP Stream 7 plugged in not charging

Once power is plugged in, the battery life estimate disappears and (plugged in) was added to the description. This is fine, but I had expected to see something about charging the battery and there was nothing. Not “charging”, not “2 hours until full”, not even the occasionally infuriating “not charging”. There is a complete lack of information about charging in any form.

Still I wasn’t worried: if the tablet wants to run off plug-in power and not charge the battery, that’s fine by me. In fact I am happy to leave the battery at around 50% charge, as that is the healthiest level for long term storage of a lithium chemistry battery. But that’s not the case, either: the tablet will run mostly on plug-in power, but still slowly drain the battery until it was near empty, at which time the tablet would power down.

Only after shutting down did this tablet begin to charge its battery. Now I am worried. If I can’t run this tablet on plug-in power alone, requiring a battery that can’t be charged while it is turned on, that combination would make it impossible to build an around-the-clock ISS tracker display.

What I wanted to do next was to poke around with the hardware of this tablet and see if I can run it without the battery. Fortunately, unlikely most modern compact electronics, the HP Stream 7 can be opened up for a look.

ESA ISS Tracker on HP Stream 7

After I found that Amazon Fire HD 7 tablet was unsuitable for an always-on screen to display ESA’s HTML live tracker for the International Space Station, I moved on to the next piece of hardware in my inactive pile: a HP Stream 7. This tablet was an effort by Microsoft to prove that they would not cede the entry-level tablet market to Android. In hindsight we now know that effort did not pan out.

But at the time, it was an intriguing product as it ran Windows 10 on an Intel Atom processor. This overcame the lack of x86 application compatibility of the previous entry level Windows tablet, which ran Windows RT on an ARM processor. It was difficult to see how an expensive device with a from-scratch application ecosystem could compete with Android tablets, and indeed Windows RT was eventually withdrawn.

Back to this x86-based tablet: small and compact, with a screen measuring 7″ diagonally that gave it its name, it launched at $120 which was unheard of for Windows machines. Discounts down to $80 (when I bought it) made it cheaper than a standalone license of Windows software. Buying it meant I got a Windows license and basic hardware to run it.

But while nobody expected it to be a speed demon, its performance was nevertheless disappointing. At best, it was merely on par with similarly priced Android tablets. Sure we could run standard x86 Windows applications… but would we want to? Trying to run Windows apps not designed with a tablet in mind was a pretty miserable experience, worse than an entry level PC. Though to be fair, it is impossible to buy an entry level PC for $120 never mind $80.

The best I can say about this tablet was that it performed better than the far more expensive Samsung 500T (more on that later.) And with a Windows license embedded in hardware, I was able to erase its original Windows 8 operating system (locked with a password I no longer recall) and clean install Windows 10. It had no problems updating itself to the current version (1909) of Windows 10. The built-in Edge browser easily rendered ESA ISS tracker, and unlike the Kindle I could set screen timeout to “never”.

That’s great news, but then I ran into some problems with power management components that would interfere with around-the-clock operation.

Inviting My FreeNAS Box To The Folding Party

Once my Luggable PC Mark I was up and running, I have one more functional desktop-class CPU in my household that has not yet been drafted into my Folding@Home efforts: it was recently put in charge of running FreeNAS. As a network attached storage device, FreeNAS is focused on its main job of maintaining files and serving them on demand. There are FreeNAS plug-ins to add certain features, such as a home Plex server, but there’s no provision for running arbitrary programs on the FreeBSD-based task-specific appliance.

What FreeNAS does have is the ability to act as a host for separate virtual environments that run independently of core FreeNAS capability. This extension capability is a part of why I upgraded my FreeNAS box to more capable hardware. The lighter-weight mechanism is a “jail”, similar in concept to the Linux container (from which Docker was built) but for applications that can run under the FreeBSD operating system. However, Folding@Home has no native FreeBSD clients, so we can’t run it in a jail and have to fall back to plan B: full virtual machine under bhyve. This incurs more overhead as a virtual machine will need its own operating system instead of sharing the underlying FreeBSD infrastructure, consuming hard disk storage and locking away a portion of RAM unusable by FreeNAS.

But the overhead wasn’t too bad in this particular application. I installed the lightweight Ubuntu 18 server edition in my VM, and Folding@Home protein folding simulation is not a memory-intensive task. The VM consumed less than 10GB of hard drive space, and only 512MB of memory. In the interest of always reserving some processing power for FreeNAS, I only allocated 2 virtual CPUs to the folding VM. The Intel Core i3-4150 processor has four logical CPUs which are actually 2 physical cores with hyperthreading. Giving the folding simulation VM 2 virtual CPUs should allow it to run at full speed on the two physical CPUs and still leave some margin to keep FreeNAS responsive.

Once the VM was up and running, FreeNAS CPU usage report does show occasional workload pushing it above 50% (2 out of 4 logical CPU) load. CPU temperature also jumped up well above ambient temperature, to 60 degrees C. Since this Core i3 is far less powerful than the Core i5 in Luggable PC Mark I and II, it doesn’t generate as much heat to dissipate. I can hear the fan increased speed to keep temperature at 60 degrees, but the difference is minor relative to the other two.

Old AMD GPU for Folding@Home: Ubuntu Struggles, Windows Win

The ex-Luggable Mark II is up and running Folding@Home, chewing through work units quickly mostly thanks to its RTX 2070 GPU. An old Windows 8 convertible tablet/laptop is also up and running as fast as it can, though its best speed is far slower than the ex-Luggable. The next recruit for my folding army is Luggable PC Mark I, pulled out of the closet where it had been gathering dust.

My old AMD Radeon HD 7950 GPU was installed in Luggable PC Mark I. It is quite old now and AMD stopped releasing Ubuntu drivers after Ubuntu 14. Given its age I’m not sure if it even works for GPU folding workloads. It was designed and released near the dawn of the age when GPUs started finding work beyond rendering game screens, and its GCN1 architecture probably had problems typical of first versions of any technology.

Fortunately I also have an AMD Radeon R9 380 available. It was formerly in Luggable PC Mark II but during the luggable chassis decommissioning I retired it in favor of a NVIDIA RTX 2070. The R9 380 is a few years younger than the HD 7950, I know it supports OpenCL, and AMD has drivers for Ubuntu 18.

A few minutes of wrenching removed the HD 7950 from Luggable Mark I, putting the R9 380 in its place, and I started working out how to install those AMD Ubuntu drivers. According to this page, the “All-Open stack” is recommended for consumer products, which I mean to include my consumer-level R9 380 card. So the first pass started by running amdgpu-install. To verify OpenCL is up and running, I installed clinfo to verify GPU is visible as OpenCL device.

Number of platforms 0

Hmm. That didn’t work. On advice of this page on Folding@Home forums, I also ran sudo apt install ocl-icd-opencl-dev That had no effect, so I went back to reread the instructions. This time I noticed the feature breakdown chart between “All-Open” and “Pro” and OpenCL is listed as a “Pro” only feature.

So I uninstalled “All-Open” and installed “Pro” stack. Once installed and rebooted, clinfo still showed zero platforms. Returning to the manual, on a different page I found the fine print saying OpenCL is an optional component of the Pro stack. So I reinstalled yet again, this time with --opencl=pal,legacy flag.

Running clinfo now returns:

Number of platforms 1
Platform Name AMD Accelerated Parallel Processing
Platform Vendor Advanced Micro Devices, Inc.
Platform Version OpenCL 2.1 AMD-APP (3004.6)
Platform Profile FULL_PROFILE
Platform Extensions cl_khr_icd cl_amd_event_callback cl_amd_offline_devices
Platform Host timer resolution 1ns
Platform Extensions function suffix AMD

Platform Name AMD Accelerated Parallel Processing
Number of devices 0

NULL platform behavior
clGetPlatformInfo(NULL, CL_PLATFORM_NAME, ...) No platform
clGetDeviceIDs(NULL, CL_DEVICE_TYPE_ALL, ...) No platform
clCreateContext(NULL, ...) [default] No platform
clCreateContext(NULL, ...) [other] <error: no devices in non-default plaforms>
clCreateContextFromType(NULL, CL_DEVICE_TYPE_DEFAULT) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CPU) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_GPU) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ACCELERATOR) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_CUSTOM) No devices found in platform
clCreateContextFromType(NULL, CL_DEVICE_TYPE_ALL) No devices found in platform

Finally, some progress. This is better than before, but zero devices is not good. Back to the overview page which says their PAL OpenCL stack supported their Vega 10 and later GPUs. My R9 380 is from their Tonga GCN 3 line, which is quite a bit older than Vega, which is GCN 5. So I’ll reinstall with --opencl=legacy to see if it makes a difference.

It did not. clinfo still reports zero OpenCL devices. AMD’s GPU compute initiative is called ROCm or RadeonOpenCompute but it is restricted to hardware newer than what I have on hand. Getting OpenCL up and running, on Ubuntu, on hardware this old, is out of scope for attention from AMD.

This was the point where I decided I was tired of this Ubuntu driver dance. I wiped the system drive to replace Ubuntu with Windows 10 along with AMD Windows drivers. Folding@Home saw the R9 380 as a GPU compute slot, and I was up and running simulating protein folding. The Windows driver also claimed to support my older 7950, so one potential future project would be to put both of these AMD GPUs in a single system. See if the driver support extends to GPU compute for multi GPU folding.

For today I’m content to have just my R9 380 running on Windows. Ubuntu may have struck out on this particular GPU compute project, but it works well for CPU compute, especially virtual machines.

Naked HP Split X2 (13-r010dx) Sitting In A Breeze Runs Faster

Mobile computer processors must operate within tighter constraints than their desktop counterparts. They sip power to prolong battery life, and that power also eventually ends up as heat that must be dissipated. Unfortunately both heat management mechanisms and batteries are heavy and take up space, so finding the proper balance is always a difficult challenge. It is typical for laptop computers to give up its ability to run sustained workloads at full speed. But if we’re not worried about voiding warranties or otherwise rendering a mobile computer immobile, we can lift some of those constraints limiting full performance: run on an AC adapter to provide power, and get creative on ways to enhance heat dissipation.

For this experiment I pulled out the most powerful computer from my NUCC trio of research project machines, the HP Split X2 (13-r010dx). The goal is to see if I can add it to my Folding@Home pool. Looking over the technical specifications published by Intel for Core i3-4012Y CPU, one detail caught my eye: it lists two separate power consumption numbers where most processors only have one. The typically quoted “Thermal Design Power” figure is at 11.5W, but this chip has an additional “Scenario Design Power” of 4.5W. This tells us the processor is designed for computers that only expect to run in short bursts. So even if TDP is 11.5W, it valid to design a system with only 4.5W of heat dissipation.

Which is likely the case here, as I found no active cooling on this HP Split X2. The outer case is entirely plastic meaning it doesn’t even have great thermal conduction to the environment. If I put a sustained workload on this computer, I expect it to run for a while and then start slowing itself down to keep the heat manageable. Which is indeed what happened: after a few minutes of Folding@Home, the CPU clock cycle pulled back to roughly half, and utilization was pulled back half again meaning the processor is chugging along at only about a quarter of its maximum capability.

HP Split X2 13-r010dx thermal throttling

For more performance, let’s help that heat escape. Just as I did earlier, I pulled the core out of its white plastic case. This time for better ventilation rather than just curiosity.

HP Split X2 13-r010dx tablet internals removed from case

Removing it from its plastic enclosure helped only a tiny bit. Most of the generated heat are still trapped inside, so I pulled the metal shield off its main processor board. This exposed the slab of copper acting as CPU heat sink.

HP Split X2 13-r010dx CPU heat sink under shield

Exposing that heat sink to ambient air helped a lot more, but passive convection cooling is still not quite enough. The final push was to introduce some active airflow. I was contemplating several different ideas on how to jury-rig an active cooling fan, but this low power processor didn’t actually need very much. All I had to do is to set the computer down in the exhaust fan airflow from a PC tower case. That was enough for it to quickly climb back up to full 1.5 GHz clock speed with 100% utilization, and sustain running at that rate.HP Split X2 13-r010dx receiving cooling

It’s not much, but it is contributing. I can leave it simulating folding proteins and move on to another computer: my Luggable PC Mark I.

Desktop PC Component Advantage: Sustained Performance

A few weeks ago I decommissioned Luggable PC Mark II and the components were installed into a standard desktop tower case. Heeding Hackaday’s call for donating computing power to Folding@Home, I enlisted my machines into the effort and set up my own little folding farm. This activity highlighted a big difference between desktop and laptop components: their ability to sustain peak performance.

My direct comparison is between my ex-Luggable PC Mark II and the Dell laptop that replaced it for my mobile computing needs. Working all out folding proteins, both of those computers heated up. Cooling fans of my ex-Luggable sped up to a mild whir, the volume and pitch of the sound roughly analogous to my microwave oven. The laptop fans, however, spun up to a piercing screech whose volume and pitch is roughly analogous to a handheld vacuum cleaner. The resemblance is probably not a coincidence, as both move a lot of air through a small space.

The reasoning is quite obvious when we compare the cooling solution of a desktop Intel processor against one for a mobile Intel processor. (Since my active-duty machines are busy working, I pulled out some old dead parts for the comparison picture above.) Laptop engineers are very clever with their use of heat pipes and other tricks of heat management, but at the end of the day we’re dealing with the laws of physics. We need surface area to transfer heat to air, and a desktop processor HSF (heat sink + fan) has tremendously more of it. When workload is light, laptops keep their fans off for silent operation whereas desktop fans tend to run even when lightly loaded. However, when the going gets rough, the smaller physical volume and surface area of laptop cooling solutions struggle.

This is also the reason why different laptop computers with nearly identical technical specifications can perform wildly differently. When I bought my Inspiron 7577, I noticed that there was a close relative in Dell’s Alienware line that has the same CPU and GPU. I decided against it as it cost a lot more money. Some of that is branding, I’m sure, but I expect part of it goes to more effective heat removal designs.

Since I didn’t buy the Alienware, I will never know if it would have been quieter running Folding@Home. To the credit of this Inspiron, that noisy cooling did keep its i5-7300HQ CPU at a consistent 3.08GHz with all four cores running full tilt. I had expected thermal throttling to force the CPU to drop to a lower speed, as is typical of laptops, so the fact this machine can sustain such performance was a pleasant surprise. I appreciate the capability but that noise got to be too much… when I’m working on my computer I need to be able to hear myself think! So while the ex-Luggable continued to crunch through protein simulations, the 7577 had to drop out. I switched my laptop to the “Finish” option where it completed the given work unit overnight (when I’m not sitting next to it) and fetched no more work.

This experience taught me one point in favor of a potential future Luggable PC Mark III: the ability to run high performance workloads on a sustained basis without punishing hearing of everyone nearby. But this doesn’t mean mobile oriented processors are hopeless. They are actually a lot of fun to hack, especially if an old retired laptop doesn’t need to be mobile anymore.

Preparing Retired Laptops For Computing Beginners

I’ve just finished looking over several old laptop computers with an eye for using them as robot brains running ROS, a research project made possible by NUCC. Independent of my primary focus, the exercise also gave me some ideas about old laptops used for another common purpose: as cheap “starter” computers for people getting their feet wet in the world of computers. This intent involves a different set of requirements and constraints than robot building.

In this arena, ease of use becomes paramount which means most distributions of Linux are excluded. Even Raspbian, the distribution intended for people to learn in a simplified environment on a Raspberry Pi, can get intimidating for complete beginners. If someone who receives a hand-me-down computer knows and prefers Linux, it’s a fair assumption they know how to install it themselves as I had done with my refurbished Dell Latitude E6230.

Next, a hand-me-down laptop usually includes personal data from its previous owner. Ideally it is inaccessible and hidden behind password protection, but even if not, the safest way to protect against disclosure is to completely erase the storage device and reinstall the operating system from scratch.

Historically for Windows laptops such cleaning also meant the loss of the Windows license since the license key has almost certainly been separated from the computer in its lifespan. Fortunately, starting from Windows 8 there is a Windows license key embedded in the hardware, so a clean install will still activate and function properly. For these Windows laptops and MacOS machines, it is best to preserve that value and run its original operating system. This was the case for the HP Split X2 I examined.

If a Windows or MacOS license is not available, the most beginner-friendly operating system is Chrome OS. It is available for individuals to install as Neverware CloudReady: Home Edition. Putting this on a system before giving it to a beginner will allow them to explore much of modern computing while also sparing them much of the headaches. And if they dig themselves into a hole, it is easy to restart from scratch with a “Powerwash”. This was what I had done with the Toshiba Chromebook 2 I examined.

But modern computing has left 32-bit CPUs behind, limiting options for older computers lacking support for 64-bit x86_64 instruction set. It meant Neverware CloudReady is not an option for them either. It is possible the user can be served by a machine that is a stateless web kiosk machine, in which case we can install Ubuntu Core with the basics of web kiosk.

And if we have exhausted all of those options, as was the case for the HP Mini netbook I examined, I believe that means the machine is not suitable as a hand-me-down starter computer for a beginner. Computers unable to meet minimum requirements for all of the above would only be suitable for basic command-line based usage. And whether computing veterans like it or not, current convention wisdom says a command line is not the recommended starting point for today’s computing beginners.

So in order of preference, the options for a beginner-friendly laptop after wiping a disk of old data are:

  1. Windows (if license is in hardware) or MacOS (for Apple products)
  2. Either original Chromebook/Chromebox or Chrome OS via Neverware CloudReady.
  3. Ubuntu Snappy Core in Web Kiosk mode.
  4. Sorry, it is not beginner friendly.

[UPDATE: Since the time I wrote this up, I have discovered a lightweight Debian distribution suitable for old x86 computers made by the Raspberry Pi foundation. I put it on the HP Mini as well as an even older Dell Latitude X1 and it appears to be a valid option between steps 3 and 4 above.]

A Tale of Three Laptops

This is a summary of my research project enabled by the National Upcycling Computing Collective (NUCC). Who allowed me to examine three retired laptop computers of unknown condition, evaluating them as potential robot brain for running Robot Operating System (ROS).

For all three machines, I found a way to charge their flat batteries and get them up and running to evaluate their condition. I also took them apart to understand how I might mechanically integrate them into a robot chassis. Each of them got a short first-pass evaluation report, and all three are likely to feature in future projects both robotic and otherwise.

In the order they were examined, the machines were:

  1. HP Split X2 (13-r010dx): This was a tablet/laptop convertible designed for running Windows 8, an operating system that was also designed for such a dual-use life. Out of the three machines, this one had the longest feature list including the most modern and powerful Intel Core i3 CPU. But as a tradeoff, it was also the bulkiest of the bunch. Thus while the machine will have no problem running ROS, the mechanical integration will be a challenge. Its first pass evaluation report is here. For full details query tag of 13-r010dx for all posts relating to this machine, including future projects.
  2. Toshiba Chromebook 2 (CB35-B3340): This machine was roughly the same age as the HP, but as a Chromebook it had a far less ambitious feature list but that also gave it a correspondingly lighter and slimmer profile. It is possible to run a form of Ubuntu (and therefore ROS) inside a Chromebook, but there are various limitations of doing so. Its suitability as a robot brain is still unknown. In the meantime, the first pass evaluation report is here, and all related posts (past and future) tagged with CB35-B3340.
  3. HP Mini (110-1134CL): This was a ~10 year old netbook, making it the oldest and least capable machine of the bunch. A netbook was a simple modest machine when new, and the age meant this hardware lacks enough processing power to handle modern software. While technically capable of running ROS Kinetic, the low power processor could only run the simplest of robots and unable to take advantage of the more powerful aspects of ROS. The first pass evaluation report is here, and all related posts tagged with 110-1134CL.

While not the focus of my research project, looking over four old laptops in rapid succession (these three from NUCC plus the refurbished Dell Latitude E6230 I bought) also gave me a perspective on preparing old laptops for computing beginners.

HP Mini (110-1134CL): First Pass Evaluation

This HP Mini netbook was the oldest of three laptops in this NUCC-sponsored research project. As a netbook, it was a very limited and basic machine even when new, and that was around ten years ago. A lot has changed in the computing world since then.

Today, its 32-bit only CPU limits robot brain applications, as only the older ROS Kinetic LTS released prebuilt 32-bit binaries. Outside of robot brain applications, any modern graphical user interface is sluggish on this machine. From Chrome OS up through Windows 10 and everything in between. When running Ubuntu Mate, it actually felt worse than a Raspberry Pi running the same operating system, which came as a surprise. Both had ~1GHz CPUs and 1GB of RAM. And even though a 10-year old Atom could outperform a modern ARM CPU, the 10-year old Intel integrated graphics processor has fallen well behind a modern ARM’s graphics core.

So it appears the best position for this machine is in running command line computing or data processing tasks that work well on old low-end Intel 32-bit chips. It would be a decent contender for the type of projects that today we would think of running on a Raspberry Pi. With the caveat of weaker graphics effects, it offers the following advantages over a Raspberry Pi:

  • Intel x86 (32-bit) instruction set.
  • Higher resolution screen than the standard Raspberry Pi touchscreen.
  • Keyboard (minus the N key in this particular example)
  • Touchpad
  • Battery for portable use
  • Actual data storage device in the form of a SATA drive, not a microSD card.

It is also the only one of the three NUCC machines to have a hard wire Ethernet port. As someone who’s been burned by wireless communication issues more than once, this is a pretty significant advantage over the rest of the machines in my book.

 

HP Mini (110-1134CL): Command Line Adept

So far I’ve determined a ~10 year old netbooks lack the computing power for a modern desktop graphical user interface, even those considered lightweight by today’s standards. Was it always sluggish even in its prime? It’s a little hard to tell from here, because even though computers have undoubtedly gotten faster, our expectations have risen as well.

But there’s more to a computer’s capability than pushing pixels around, so we fall back to the next round of experiments with command line interface systems. And since we’ve already established that a solid state drive was not a great performance booster on this platform, I put the original spinning platter hard drive back in for the next round.

This time instead of Ubuntu Desktop, I installed Ubuntu server edition instead. This minimalist distribution lacks the user friendliness of a graphical user interface, but it also lacks the graphics processing workload of displaying one as well. As a result this machine is quite snappy and responsive. I found it quite usable, especially now that I’ve learned about virtual consoles and use the Alt key plus F1 through F6 to switch between up to six different sessions. Simple tasks like running Python scripts and running a basic server were done easily and quickly.

I started experimenting with Ubuntu 16, because Ubuntu did not release prebuilt installation binaries for 32-bit Ubuntu 18. However, once Ubuntu 16 server was and and running, I was able to rundo-release-upgrade to move up to Ubuntu 18. From minor tinkering I didn’t notice any significant difference between them.

Then I remembered I had played with an even more minimalist Ubuntu earlier, on an even older machine. Ubuntu 18 Snappy Core is available for 32-bit i386 processors, and it installed successfully on this laptop. Now I have one more incentive to learn how to build my own snaps to install on such a system. I just have to remember to that I can only connect to an Ubuntu Snap machine via SSH, and the list of valid keys associated with an account do not auto-update. I typically generate a SSH key every time I reset a machine, and I no longer have the keys to access my previous snappy core experiment. I ended up reinstalling snappy core to pick up the current set of SSH keys.

HP Mini (110-1134CL): Ubuntu Mate and Chrome OS Slow Even With SSD

After a ~10 year old netbook was upgraded with a solid state drive, we can now confirm the hard drive is not the only thing holding back performance. The following experiments indicate the old Atom CPU at the heart of this machine lacks the power to run any modern operating system graphical user interface.

First up at bat was Ubuntu Desktop 16.04 i386. It ran sluggishly when loading from this machine’s original spinning platter hard drive, and it was not significantly better when loading from the upgraded solid state drive. Watching the HDD activity light earlier, I thought this might be the case, but wanted to verify firsthand, which I have.

Next candidate was Ubuntu Mate, which has a 32-bit installer for 18.04. (Mainline Ubuntu stopped supporting 32-bit in 18.) Even though Ubuntu Mate advertised itself as a lighter-weight alternative to mainline Ubuntu, it was unfortunately still far from pleasant to use. But if needed, one reason to run Ubuntu Mate 18.04 is for the longer supported timeframe of Ubuntu 18. According to Ubuntu releases list, 18 is supported until April 2023.

I then tried an even more constrained operating system: install Chrome OS and make a faux Chromebook out of this thing. I had known Neverware CloudReady as a build of Chrome OS that anyone can install on an old laptop to turn it into a Chromebook. I had trouble making it work before on an old machine before, and wanted to try again.

I noticed the minimum recommended amount of RAM has increased as I remembered it was 1GB, now it is up to 2GB. But that was just a recommendation and I was able to load CloudReady on this netbook with just 1GB RAM. Once launched and running, CloudReady proved to be about as sluggish as Ubuntu Mate 18.

But that’s not the biggest problem:

CloudReady 32-bit EOL

CloudReady, originally advertised to help give old machines new life, has been forced to leave 32-bit CPUs behind. After seeing this notification I went online to find their announcement, as well as confirmation that Chrome OS v76 was about the right vintage for this to happen.

For interactive graphical desktop use, it really doesn’t get any more lightweight than Chrome OS and this machine still struggles. It looks like we need to fall back to a text-based server edition of operating system software.

[UPDATE: I found an even lighter weight distribution of Linux for old 32-bit x86 machines that would be familiar to users of the Raspberry Pi.]