Video summary

I’m glad I didn’t invest, so I can talk the Framework Desktop

Main summary

Key takeaways

Product Review

Product reviewed

Framework Desktop — Framework’s first fixed/desktop PC.

The video focuses on a mostly/fully built unit packed into a tiny ~4.5 L chassis, using a highly specified configuration intended to showcase both gaming and AI performance.

Key features mentioned

  • Repairable / serviceable / upgradeable design philosophy
    • Described as the Framework “DNA,” similar to their laptops.
  • Barebones purchase model
    • Sold without core components; the user installs/selects parts.
  • Customizable I/O via modular “IO modules/tiles”
    • Presented as similar to the Framework laptop approach.
  • In-the-box assembly
    • “Not as much as I thought,” but assembly is still required.
    • Includes a screwdriver described as “two screwdrivers in one.”
    • Not fully assembled at the time of the video.
  • Storage configuration in this build
    • Two 4TB WD SN850X Gen4 NVMe SSDs (strongly recommended by the speaker).

High-end compute configuration (this build)

  • CPU: 16-core AMD Ryzen Zen 5
  • GPU: 40 RDNA 3.5 GPU cores
  • RAM: 128 GB, noted as soldered (with controversy mentioned below)
  • Form factor: very compact ~4.5 L, with cooling/PSU integrated for that footprint

Power & networking

  • PSU: 400W Flex ATX 80+ Gold
  • Power cable availability/pricing discussed (see cons)
  • Networking: Wi‑Fi 7 chip with an internal antenna

Port list (as listed)

  • HDMI 2.1
  • Dual USB4 Type‑C (40Gbps)
  • Dual DisplayPort 2.1
  • 5 GB Ethernet
  • Pair of 10 GB USB‑A
  • 3.5mm audio
  • C13 power inlet (back)

Expandable / accessory concept

  • Top/front aesthetics with a magnetic panel/tile system
  • Demonstration of custom 3D‑printed adapters/tiles
  • Conceptual mention of adding an SD-card ingest station

User experience / setup notes

  • Build difficulty: surprisingly easy for their first self-build on their channel.
  • Operating system options:
    • Windows 11 Home or Pro (download provided), or no OS (user supplies)
    • Their system runs Windows 11 Pro 24H2
  • Driver workflow:
    • After OS install, download the Framework driver bundle (chipset/graphics/Wi‑Fi/Ethernet).
    • They also tried latest AMD drivers, but caution that Framework’s bundle is more reliable, since the hardware behaves more like a mobile SOC than a typical desktop platform.

Windows performance mode is critical

On Windows, performance depends heavily on power settings:

  • Switch power plan: Balanced → Best performance
  • If not set correctly, reviews reportedly show performance around ~100W
  • With correct settings:
    • can boost to ~160W briefly
    • then ~140W for longer
    • eventually leveling near ~120W
  • The video emphasizes this is an easy-to-miss step that significantly changes results.

Performance results cited

PassMark (CPU + GPU)

  • PassMark CPU/overall score: ~61,000
  • Comparison:
    • Versus Ryzen 9950X (16 cores): ~61,300 vs ~66,000 average
    • Despite a lower power budget and “mobile SOC” nature, performance is presented as very close.

GPU scores

  • ~23,000 (3D test)
  • ~1,200 (2D test)
  • PassMark leaderboard mapping suggested roughly:
    • RTX 4060 Ti / RTX 5060 Ti class, or
    • Radeon 7000 XT class (with the caveat that laptop variants weren’t the focus)

Real games (examples at 1440p)

  • Arc Raiders (1440p medium)
    • ~100–120 FPS
    • described as “way more than playable”
    • noise estimated around ~40–42 dB in performance mode
  • Overwatch 2 (1440p)
    • ~150–180 FPS (ultra preset mentioned)
  • Counter-Strike 2 (low preset)
    • ~300 FPS average
    • ~130 FPS 1% lows

Overall claim: strong across esports and demanding games, but the system isn’t positioned as “only” one thing—it’s framed as mixed-use (gaming + AI + productivity).

AI / productivity performance notes

  • AI workflow using LM Studio on Linux (Cache OS / Arch-like) and Vulkan mode
  • AMD ROCm guidance referenced:
    • best performance expected on AMD stack
    • “not great in Windows” per the speaker
  • Soldered memory limitation
    • The speaker states socketed RAM isn’t possible because the Strix Halo-class hardware requires soldered RAM
    • Desoldering is mentioned as technically possible but effectively not a supported path

Example LLM outcomes (tokens/sec)

  • “OpenAI 120B model” in reasoning mode low: ~50 tokens/sec
    • (btop confusion mentioned: it showed ~32GB system RAM, because the remainder is GPU VRAM)
  • Compared to MacBook M5 for the same/workload-like quant:
    • ~40 tokens/sec vs ~25–26 tokens/sec on M5
  • Compared to their M4 Max:
    • ~70 tokens/sec on M4 Max
    • so the new system isn’t always faster, but still strong
  • Larger 70B model (~71GB VRAM used):
    • runs, but “not going super fast”

Video conclusion: the system is “seriously competent” for gaming, AI, and productivity.

Main pros mentioned

  • Strong repairable/serviceable/upgradeable approach consistent with Framework
  • Small ~4.5 L footprint with high performance
  • Excellent modular/customizable I/O via tiles/modules
  • High 1440p gaming performance (high FPS)
  • Good thermals/noise in practice (~40–42 dB cited)
  • Strong AI inference performance for its class
  • Driver + power-mode guidance improves real-world out-of-box results

Main cons / criticisms mentioned

1) Soldered RAM (128 GB variant)

  • Presented as the most notable controversy.
  • The speaker argues it’s not Framework’s fault, since it’s driven by AMD chip requirements.
  • Advice implied: avoid soldered-RAM variants if that constraint is unacceptable to you.

2) PCIe expansion limitations

  • A PCIe Gen4x4 connector is present “in theory,” but:
    • closed-back slot makes typical GPU usage difficult
    • a blocked-off PCIe slot is described as a strange design choice (would require the chassis to be slightly taller)

3) Front cooling / extra fan

  • Optional mounting for an 80mm fan is mentioned, but it doesn’t come included.

4) Power cable pricing confusion

  • Some users complain about paying extra for a power cable.
  • The video’s stance: it’s not unfair extra pricing, reflecting supply/cost logic.

Comparisons made

  • Versus Ryzen 9950X (PassMark baseline comparison): close despite power differences
  • GPU class mapping via PassMark leaderboard:
    • roughly mid/high 4060 Ti / 5060 Ti / 7000 XT desktop-class range
  • AI tokens/sec comparisons:
    • Versus MacBook M5: system faster (~40 vs ~25–26 tokens/sec)
    • Versus M4 Max: M4 Max faster in that specific test (~70 tokens/sec)
  • GPU laptop variants are noted as not the core focus in the GPU-class comparisons.

Pricing / model lineup mentioned

  • Storage can be expensive if purchased through Framework, which is part of why barebones is offered.
  • For this configuration, the video describes totals roughly as:
    • ~$2,000 base for the build configuration (plus other costs)
    • ~$700 storage
    • plus Windows license and other factors → nearly ~$3,000 total
  • Other SKUs mentioned:
    • Max 385: starting at $1,100
    • 64GB variant with 16-core Max + 395: starting at $1,600
  • Implied guidance:
    • gaming-only users may prefer 64GB
    • AI/video editing/LLMs may justify 128GB

Overall verdict (concise recommendation)

The video concludes the Framework Desktop is a highly capable, compact system that performs very well for gaming and AI, while maintaining Framework’s modular/repairable approach.

The biggest drawback—soldered RAM—is treated as largely inherent to the AMD chip rather than Framework’s design choice. The implied recommendation is to buy the right RAM tier for your use case (64GB for typical gaming; 128GB for AI/LMM/productivity) and to ensure the correct Windows power mode for best performance.

Unique points mentioned (deduplicated)

  1. Framework’s repairable/serviceable/upgradeable goal since 2021.
  2. Desktop is controversial but intentionally tiny (~4.5 L).
  3. Barebones purchasing with user installs/OS selection.
  4. Strong SSD pick: WD SN850X 4TB Gen4.
  5. Modular/customizable I/O via modules/tiles.
  6. Assembly required; not fully assembled at the time referenced.
  7. Screwdrivers included.
  8. Example configuration: Zen 5 16-core CPU, RDNA 3.5 40-core GPU, 128GB soldered RAM.
  9. Port set includes HDMI 2.1, dual USB4-C, dual DP 2.1, 5GbE, dual USB‑A, audio, and C13.
  10. Top panel access/removable panels and internal fan header access; holes for optional handle (handle not included for the user).
  11. Magnetic/tile front panel customization; “cute” customization and color story.
  12. Power cable pricing complaint addressed.
  13. Build includes a sponsor mention unrelated to the PC (Boot.dev).
  14. Customization demo using 3D-printed adapters/tiles; maker-side quests mentioned.
  15. Side panel options mentioned (black vented vs translucent).
  16. PSU described as 400W Flex ATX 80+ Gold, including airflow notes.
  17. PCIe expansion limitations (closed-back PCIe slot; blocked-off half-height slot).
  18. AMD Wi‑Fi 7 with internal antenna.
  19. Optional extra 80mm fan mounting (not included).
  20. CMOS reset button included.
  21. OS options: Windows 11 Home/Pro download or no OS.
  22. Emphasis on Framework driver bundle reliability vs generic AMD drivers.
  23. Critical performance setting: Balanced → Best performance.
  24. Power targets described: ~160W brief, ~140W longer, then ~120W.
  25. Pricing rationale: barebones to reduce expensive storage bundling; near ~$3,000 total for this build.
  26. PassMark CPU: ~61k, close to 9950X expectations due to power budget.
  27. PassMark GPU: ~23k 3D class mapping to desktop 4060 Ti / 5060 Ti / 7000 XT ranges.
  28. Gaming benchmarks at 1440p:
    • Arc Raiders ~100–120 FPS, noise ~40–42 dB
    • Overwatch 2 ~150–180 FPS
    • CS2 ~300 avg / ~130 1% lows
  29. Memory tier guidance: 64GB for gaming, 128GB for AI/LMM/productivity.
  30. Soldered RAM explained as required by AMD Strix Halo; socketed memory not compatible.
  31. AI setup on Linux: Cache OS/Arch-like, LM Studio, Vulkan, ROCm mention.
  32. Token/sec examples: ~50 for 120B low reasoning; ~40 vs M5 ~25–26; M4 Max ~70.
  33. 70B model runs but slower; ~71GB VRAM used.
  34. Final hope: a future board with swappable memory and/or more chip options.

Speaker views / contributors (at end)

  • Primary speaker (implied “Jack,” the video author):
    • Owns the walkthrough of ports/specs, setup steps, driver/power-mode guidance, benchmark interpretation, and the soldered-RAM controversy framing.
  • No other meaningful technical contributors appear in the provided subtitles beyond generic mentions of reviewers/commenters and a non-PC moment (“Arlo” demo).

Original video