PC Build

Jumping ahead a bit: After the build was complete, I was talking to a co-worker whose son recently built a computer. That approach turned out to be much different than the one outlined below. His son purchased a few flagship items and pieced together the rest with budget buys and items on hand. Who knew that a top-of-the-line processor (Ryzen 9 7950X3D) and graphics card (Radeon RX7900X3D) wouldn’t overcome the limitations of a 10-year-old 7200RPM HDD? Needless to say, the son was disappointed with his build. 

AMD Ryzen uses a homogeneous core topology (all performance cores (P-cores)) whereas Intel 13th Gen uses a hybrid topology (P-cores and efficiency cores (E-cores)). My thought process was: Many interactive creative workloads are lightly threaded or scale poorly past a handful of performance cores; the hybrid architecture would allow the scheduler to preferentially allocate high-performance threads to P-cores, and background tasks to E-cores. Offloading background tasks to E-cores prevents them from stealing frequency and thermal headroom from the P-cores. Though this does depend on how well the scheduler allocates processor resources. 

Price-wise, between comparable Intel and AMD processors, the Intel 13th-gen processor undercut comparable Ryzen 9 pricing. I was strongly leaning Team Blue, and then I found benchmark data specific to one of the use-cases, which showed Intel's hybrid architecture outperforming similarly priced Ryzen processors, reinforcing the architectural reasoning. Thanks to the data, I settled on the i7-13700K as a balance of performance and cost.

Eventually, I decided on four logical drives: One for the OS and documents (C), one for program files and libraries (D), a cheaper but faster large drive for large project files (photos, videos, etc.; E), and a large, slow drive for local backup and redundancy (F). 

C: 1x 1TB PCIe 4.0 SSD. Boot, temp file, and document drive (OS and document files (Word, Excel, etc.))
D: 1x 2TB PCIe 4.0 SSD. Program file data, project libraries, program caches, active scratch work, etc. 

E: 2x 2TB SATA III SSD (RAID0, RST). Large data files (pictures, videos, etc.). 

F: 2x 12TB SATA III HDD (RAID1, RST). Boot/program drive backups, document and library backups, and a mirror of the RAID0 drive.  

C/D Drives: While there isn’t much of a speed advantage having the OS on one drive and program files on another anymore, I chose to keep this architecture for organizational and restore granularity, not performance reasons.
E Drive: Ideally, this would have been combined with the D drive using a 6-8TB PCIe SSD; however, this would be extremely cost-prohibitive. The cost-performance compromise is two SATA SSD drives in a RAID0 configuration, almost doubling the read/write speed of one single larger drive for large, sequential workloads (for random I/O, the speed remains SATA-limited). This is the highest-risk element of the storage design, and is mitigated by daily mirroring to mitigate significant data loss (see ‘Backup’ below). 

F Drive: This is two 7200RPM SATA HDD drives in a RAID1 configuration that serve as space for backing up the Boot and Program drives, in addition to mirroring the RAID0 drive. 

RAM was nearly the last choice, and came down to ‘what do I have left in the budget?’ Although four DIMMs would have provided the same capacity, I opted for one DIMM per channel to avoid the signal-integrity penalties associated with 2DPC configurations. On modern DDR5 systems, fewer DIMMs per channel generally translate to higher stable speeds and better overall memory performance. While 16GB met the minimum requirement, it seemed reasonable to add more given Lightroom catalogs, video editing, and AI experimentation. The budget allowed an increase to 32GB 5600MHz. 

DIMM A1: Empty

DIMM A2: 1x 16GB 5600MHz DDR5

DIMM B1: Empty

DIMM B2: 1x 16GB 5600MHz DDR5

The first step is adjusting the processor core clock rate and operating voltages. To do this, I set all fans to run at 100% in the BIOS, enabled the Intel default PL1/PL2 power limits, then booted and started the stress-test software while monitoring each core's performance. With Intel's default PL1/PL2 limits enforced, the CPU still reached its maximum operating temperature within a minute, triggering CPU throttling of individual P-cores and E-cores to prevent exceeding temperature limits. The first adjustment was to slightly underclock and undervolt. Eventually, I settled on underclocking the P-cores turbo frequency by 300MHz (4.9GHz) and E-cores turbo by 200MHz (4.0GHz), and a small -100mV offset. This allowed for a 30-minute CPU stress test without internal CPU thermal throttling. Remember, we’re tuning for a marathon, not the 60-meter dash, and even in a marathon, the CPUs will have some duty cycle, which is why I decided 30 minutes was sufficient. Sidenote: By reducing the working and boost voltages, I may have missed the Raptor Lake voltage issue that was traced to voltage overshoot and aggressive board defaults.

The next step was to check that the RAM was stable under the given XMP profile. Stress testing the memory indicated stable operation for the enabled XMP profile, and that’s pretty much all I was concerned about at this point, so I moved on with tuning the fan speed curve. 

To minimize fan noise at a light load and moderate load. To do this, I dug out two thermocouples and my benchtop multimeters, embedding one thermocouple in the CPU heat sink and another in the GPU heat sink. To give a light load, I recoded a browser use-case in Selenium and ran it in a loop while trying to find the minimum fan levels that brought the system temperature to equilibrium. For a moderate use case, I put a few MP4s back-to-back in OpenShot, limited it to 2 P-cores, and encoded a 3-hour 4K video while adjusting the fan speed to find equilibrium. With these two settings, I finalized the fan curve. The nice part is that at the light load, only the CPU fan and one rear exhaust fan were on at a low level, and the case was nearly silent.

There are a lot of backup options out there, which can be a rabbit hole in and of itself. Ultimately, I decided to go with a recommendation from a trusted colleague who has been building PCs and designing embedded architecture for longer than I've been alive. 

The backup software is automatable, which makes the following backup tasks fairly easy: 

• The boot drive is automatically imaged on the first Monday of every month, then incrementally imaged twice a week for the duration of the month. The software will automatically clean up old images; right now I have it save the last two months of full+incremental images, and four months of full images before that. 

• There’s a shortcut sync documents from the Boot drive and libraries from the Data drive to the RAID0 drive, then mirrors the RAID0 drive to a folder on the RAID1 drive and shuts down or hibernates the computer, which I use at the end of the day. 

• Once a week, there’s a scheduled boot in the middle of the night which images the Data drive to the RAID1 drive, then mirror-syncs the RAID1 drive (boot drive images, data drive images, documents/library backup, and RAID0 drive backup) with my NAS (if connected). 

What this means: 

• In the event of a Boot drive failure (or broken update) or Data drive failure, I’m out a maximum of one day’s work of documents or edits stored in the libraries, and a maximum of four days' tweaks to the OS or one week of drive programs that were installed/uninstalled/updated. 

• In the event of a failure in the RAID0 drive(s), I’m out a maximum of one day’s work. 

• In the event both of the RAID1 drives crash, well, probably minimal other than about a day of rebuild time, as what’s stored on them should be recoverable from the other drives or NAS.