Free RAM, not a SCAM

Summary of technological concepts & product features

• Background on Nvidia mining GPU (CMP 170HX)

  • The video argues the CMP 170HX is essentially an A100-derived compute card intentionally crippled for crypto mining.
  • Key “crippling” mechanisms mentioned:
    • No graphics hardware (it is not a real graphics-capable GPU).
    • PCI Express bus restrictions (later discussed as unlocking to PCIe Gen2; further performance via capacitor mod).
    • HBM memory reduced from 64GB (HBM2E) down to 8GB.
    • Floating point performance locked down (with modding/unlocking discussed below).

• Core idea: modders unlock the disabled capabilities

  • The video claims enterprising modders have found ways to restore:
    • Floating point performance
    • Hidden memory
    • PCIe interface / bandwidth limitations (at least partially)
  • Overall result: turning the card from “e-waste” into something usable for AI workloads.

---

Guides / tutorials / procedures highlighted

1) Baseline testing (stock configuration)

• Purpose: establish what can run before unlocking, using a controlled comparison.
• Benchmark method: Llama Bench (not comprehensive, but used for like-for-like comparisons).
• Test setup details:
  • Multiple GPUs included to contextualize results:
    • CMP 170HX mining cards (2 units)
    • GeForce RTX 3090 Ti (compute comparison; described as faster compute but less RAM)
    • GTX 1650 (used for video output)
  • Pricing context noted: CMP prices increased substantially compared to when previously reviewed.

2) Unlock procedure (software/driver-based)

• Unlock difficulty: rated around 6/10, with more complexity during re-locking.
• Key steps mentioned:
  • Disable Secure Boot
  • Purge/remove existing Nvidia drivers
  • Blacklist “nouveau” drivers
  • Rebuild initramfs / RAM filesystem (Linux-specific steps)
  • Use a specific cloned GitHub repo and a modified driver approach
• Important behavior described:
  • The unlock works via a modified driver that unlocks cards in that system.
  • Moving the card to another machine will re-lock it until that new system is unlocked again.

3) Verification & reliability testing (memory test)

• Memtest procedure: run a memory read/write scan to catch bad HBM sectors.
• Meaning of “bad sectors”:
  • The card may later crash (described as “burp out”) during math/compute workloads.
• Result:
  • One tested card reported memory errors, supporting the idea that unlocked hardware may still be inconsistent due to manufacturing/binning.

4) Optional hardware mod (capacitor / PCIe bandwidth unlock)

• If the software unlock is not enough, the video describes a capacitor mod using leaked schematics.
• A person described as a “soldering wizard” adds back tiny capacitors above the PCIe slot to unlock full 16-lane functionality (contrasted with a constrained mode).
• Tradeoff / implications:
  • Higher bandwidth is claimed, but the mod increases complexity and risk.

---

Performance & analysis results (AI workloads)

• Stock vs unlocked behavior

  • With the mining configuration, the card struggles—especially with larger model sizes—due to:
    • Limited FP throughput
    • Severely reduced usable VRAM (8GB)
    • PCIe limitations
  • After unlocking:
    • Model loading and usability improve dramatically
    • Still described as slower than RTX 3090 Ti in compute-heavy sections, but “worlds faster” relative to stock.

• Model size experiments

  • Smaller model: loads and generates well once unlocked (though still not equivalent to higher-end compute GPUs).
  • Larger “Qwen” model (~28GB class):
    • Mentioned fallback to system memory (system RAM used when VRAM isn’t sufficient).
    • Smaller model speeds up strongly, while larger-model loading/generation is still impacted by PCIe bandwidth and compute limits.

• PCIe bottleneck insight

  • Even after unlocking, the card is said to run in a constrained PCIe mode:
    • unlocked to PCIe Gen2
    • but still limited to x4 bandwidth
  • Further bandwidth improvements are tied to the capacitor mod.

• Token processing / throughput discussion

  • The video contrasts:
    • Time to first token (affected by load/bus)
    • Prompt processing speed
    • Token generation rate
  • It claims the 3090 Ti can beat the CMP unlocked card for generation speed due to faster compute, but the unlocked CMP remains usable.

---

Reliability, pricing, and “not guaranteed” warnings

• Pricing increased significantly since earlier discovery/review.
• No guarantee the card will work perfectly after unlock:
  • Cards may be binned due to failing power thresholds, quality standards, or HBM damage/failures
  • Manufacturing may fuse off failing portions, producing “crippled” mining variants.
• Heat and inconsistency:
  • A borrowed-card owner claims inconsistencies may worsen/improve with temperature (e.g., above 85°C),
  • but the video warns not to rely on temperature behavior for production stability.
• Production guidance implied:
  • Memtest is recommended
  • Expect variability across units.

---

Takeaway / recommendation framing

• The unlocked CMP 170HX is portrayed as a budget/functional alternative for mid-to-large model sizes (roughly 20–40GB class):
  • Enough RAM for context and multiple workloads
  • Usable for interactive and agentic/background tasks
• However:
  • Bottlenecked by PCIe bandwidth/architecture limits
  • Reliability varies depending on card quality (e.g., bad HBM sectors)
  • Unlocking effort and possible hardware mods add friction
  • At current pricing, it’s positioned as “pretty good” only if you can obtain a working unit and/or accept limitations.

---

Main speakers / sources

• Lucas (from “the lab”) — demonstrates/introduces unlocking and mods.
• Dan (“resident soldering wizard”) — performs the capacitor/hardware mod.
• Amog Mooney Cot — referenced as the creator linked in the repo/steps for the unlock process.
• The video also references:
  • earlier research/papers on GPU unlocking
  • the “170th Street repo” / LT Labs companion article.