Projects
The Website
Over the weekend, I have been rushing to finish writing the blog posts for the last three weeks.
Polyvinyl Chloride
With expected parameter counts for my AI workloads now expected to be well over 128 billion FP16, I decided I am going to need much more RAM to accommodate for this. What I have planned now may be well out of budget, pressuring me more to earn and save more funds for this build.
Updated list as of August 10, 2025:
- Motherboard: Supermicro H11DSi
- CPUs: 2x AMD EPYC 7371
- RAM: 512GB via 16x 32GB DDR4-3200 (at DDR4-2666 on 7371) 1Rx4 Registered ECC Samsung M393A4G40AB3-CWE
- GPU: upcoming Nvidia RTX 5090 SUPER, either 48GB or 64GB
- Storage: TBD still
- Power Supply: Super Flower Leadex Titanium 1600W 80+ Titanium
- Case: Phanteks Enthoo Pro Tempered Glass
Notes on each component:
- Motherboard + CPUs: As I mentioned in previous posts, the use of a Supermicro H11-series board with Naples CPUs will severely limit the bandwidth between the GPU and CPUs as it only supports up to PCIe 3.0 while the RTX 5000 series has PCIe 5.0. With PCIe 3.0 x16, the bandwidth between the CPU and GPU would be up to 16 GB/s instead of the 64 GB/s possible with PCIe 5.0. This may pose a bottleneck when transferring model weights from system RAM to the GPU's VRAM.
- RAM: Samsung consistently seems to be the lowest cost option for RAM on second-hand channels like eBay. Samsung does fit the aesthetics I plan to have for this build. While memory is limited to DDR4-2666 with use of AMD EPYC Naples, DDR4-3200 memory is used to offer a possible upgrade path if I upgrade to an H12DSi-NT6 with Zen 2/3 CPUs.
- GPU: The RTX 5090 SUPER is currently speculated, it is unknown at this time how much VRAM is going to have available
- Power Supply: While being a US$400 unit as of writing, the power supply is considered one of the most important parts of the build if not the most important, especially with the immense amount of power expected to be used in this configuration. A bad power supply could mean the end of all of the other parts in the system. I even plan to test the power supply out of this build to ensure it works out of box and does not immediately fry everything.
Dedicated Task/Calendar system
Also over the weekend, I switched out the Ryzen 5 3600 desktop for the much lower power Intel Celeron J4105-based system. It felt like a waste using what was essentially a gaming PC to do something that a Raspberry Pi could do (did not feel like configuring Raspberry Pi OS). Before, I was not able to use this Celeron-based desktop as one of the outputs would start to flash minutes after booting which I thought was due to VRAM running out. After reinstalling Fedora, upgrading to Fedora 42 as a result, I was able to use the iGPU (onboard graphics) to run both monitors. It was likely an Intel graphics driver issue than allocated VRAM; 256-512MB should be more than enough for a 1280x1024 and 1920x1080 monitor.
2nd Life Inc.
Day 13 - August 6, 2025
On this day, I continued on networking equipment, coming across Cisco routers. As these are Cisco systems, the IOS console is just like what I learned to use during the 2020-2021 school year at the Cisco Tech Center program. It brought back many memories of that time.
When testing monitors, I came across a small device with an HDMI connector and USB connector that had the Microsoft logo on it. I found out it was a Microsoft Wireless Display Adapter from 2017. After connecting it to a monitor, the connection interface showed. I was so impressed by the fact that something this small is able to drive an HDMI output and decode 1080p video in real time. FCC teardown pictures showed that it consisted of a Cavium PureVu SoC and an eMMC flash on the other side.
On this day, I worked on a 2021 iMac, being the first Apple M1 (or any Apple Silicon) system I worked on here. The front glass was chipped and cracked but the display itself was intact.
Day 14 - August 7, 2025
I finished up the 2021 iMac this day. As with every Thursday, scrap sorting started at noon and lasted throughout the rest of the day.
Day 15 - August 8, 2025
On this day, I was tasked with some unique systems.
I worked on some tower-format servers, a Dell PowerEdge T430 and T540. Later on, I was tasked on getting the specifications of a Dell blade, including getting the information on the specific RAM modules used, which is the first time I was actually told to do this.
After getting the specifications of the Dell blade, I worked on inspecting a Supermicro tower server, the first Supermicro-branded system I have worked on (not a rebranded unit with a Supermicro motherboard). While interesting, specifications I found somewhat disappointing, using a C2SBC-Q which does not have ECC memory support unlike every other Supermicro board I have worked with.
I came across two Dell Precision 690 desktops. Upon booting, I noticed that both systems used Intel Xeon X5355 CPUs that is part of "Clovertown". While anticipating to upgrade my 2006 Mac Pro to X5365s since 2020, I have not worked with a system with Clovertown CPUs until now. As part of the "Clowntown" inside joke relating to an online friend misreading Clovertown as Clowntown, I made sure to take a picture of the system with a can Faygo.
I was sought out to fix a recurring kernel panic that consistently occurred across several 2019 MacBook Pros after macOS Sequoia 15.6 was installed. The kernel panic on every system I worked on seemed to have been fixed by merely resetting the NVRAM.
Near the end of the shift, I found some bins of hard drives to wipe. I brought over the Supermicro tower which had a front 4x HDD bay which was helpful to clear these hard drives.