Cmpunlocker is a GitHub-hosted Linux patch for NVIDIA’s CMP 170HX (GA100) mining card, described by its author as removing imposed throttling and restrictions to restore A100-class compute throughput and memory capacity. The supplied repository snippet targets nvidia-open 580.x, while the evidence title names driver 610.43.03; claims also vary between 64GB and 80GB, and the snippets do not independently verify reproducibility, BF16 performance, or PCIe Gen2 functionality. Treat the full unlock as a project claim pending benchmarks and replication.
2026-08-06T23:37:51Z
Independent field reports establish that cmpunlocker can expose roughly 64GB and run practical inference on some CMP 170HX cards, but observed unlock failure, instability, and constrained PCIe behavior disprove the full reproducible 64GB/BF16/PCIe hypothesis. The useful residual finding is a card-lottery unlock, not a dependable local-inference platform.
2026-08-06T22:25:59Z
The nominal update adds no controlled replication or direct BF16 and PCIe validation beyond the field reports already incorporated. Evidence still supports a useful unlock on some cards, but failures and instability make the full reproducibility claim a hardware lottery rather than an established result.
2026-08-06T21:29:41Z
No new substantive result extends the field reports already incorporated: the patch works on multiple cards and supports practical inference, but unlock failures, instability, and limited PCIe performance prevent treating it as reproducible hardware. Further engagement churn should not move the case without controlled card-to-card success rates or direct BF16 and PCIe validation.
2026-08-06T20:28:41Z
Fresh independent field reports and a hosted-card benchmark now show multiple CMP 170HX cards exposing roughly 64GB and running practical local inference, including multi-card deployments. However, a direct two-card report found one unlock failure and instability in the other, so the patch is demonstrably usable on some hardware but the full reproducible 64GB/BF16/PCIe claim is undermined by a real card lottery.
2026-08-06T20:21:42Z
evidence attached: reddit.post.1vhfio2 — Directly extends the open CMP 170HX unlock episode, though it provides no independent technical verification.
2026-07-31T13:25:26Z
The update is only minor engagement churn on an existing buyer thread, with no new physical-card replication or validation of BF16, PCIe, or multi-GPU behavior. The case remains a single-card quantized-model proof rather than a reproducible full unlock; wait for concrete field results.
2026-07-31T05:22:27Z
The latest trigger adds no substantive field report or second physical-card reproduction. The case remains a single-card quantized-model proof, with reproducibility, BF16, PCIe, and multi-card behavior still unresolved; ignore further engagement churn until concrete results appear.
2026-07-31T04:22:39Z
The new field-report request produced no actual report, leaving the case at a single-card quantized-model proof rather than a reproducible full unlock. Further discussion should not advance it without another physical-card replication or direct BF16, PCIe, and multi-GPU results.
2026-07-31T04:20:57Z
evidence attached: reddit.post.1vbf1s0 — This directly seeks field evidence relevant to whether CMP 170HX unlocking works in practice.
2026-07-29T09:33:58Z
The nominal attachment adds no second physical-card reproduction or direct validation of BF16, PCIe, or multi-GPU behavior. The case remains a single-card proof surrounded by repetitive engagement; revisit only when concrete deployment results arrive.
2026-07-29T01:22:23Z
The nominal evidence trigger adds no new physical replication beyond the lone single-card benchmark. Treat the full unlock as unproven until another card validates memory capacity, BF16 performance, and PCIe behavior; engagement churn no longer merits frequent review.
2026-07-28T22:25:22Z
The latest attachment adds no substantive evidence beyond the lone single-card quantized-model benchmark, so the full reproducibility claim remains unsupported. Repetitive engagement should no longer trigger frequent review; wait for another physical-card replication or direct BF16, PCIe, and multi-GPU results.
2026-07-28T21:26:32Z
The nominal evidence update adds no second physical reproduction or direct validation of BF16, PCIe, or multi-card behavior. The case remains a single-card proof surrounded by repetitive buyer interest, so wait for concrete deployment results rather than reviewing engagement churn.
2026-07-28T19:27:22Z
The attachment does not add a second physical reproduction or validate BF16, PCIe, or multi-card behavior, so the case remains a single-card proof rather than a reproducible unlock. Repetitive buyer interest no longer warrants frequent review; wait for concrete deployment results.
2026-07-28T18:25:01Z
The nominal evidence trigger adds no substantive result beyond the existing single-card quantized-model benchmark. The case remains relevant but unproven until additional physical cards establish reproducible memory unlocks and directly validate BF16, PCIe, and multi-GPU behavior.
2026-07-28T17:27:08Z
The nominal evidence attachment adds no new physical reproduction beyond the lone single-card quantized-model benchmark. Card-to-card reliability and the combined BF16, PCIe, and multi-GPU claims remain unverified, so repetitive interest no longer merits frequent review.
2026-07-28T16:24:41Z
The latest trigger adds no substantive evidence beyond the lone single-card quantized-model benchmark; repeated interest and buyer questions still do not establish reproducibility. BF16 performance, PCIe functionality, multi-card scaling, and card-to-card unlock success remain unresolved.
2026-07-28T13:31:33Z
Continued repetitive engagement (buyer questions, price gripes) with no new reproduction; still resting on a single independent single-card benchmark. Cooling further as the discussion has become steady-state Q&A rather than advancing evidence.
2026-07-28T12:27:12Z
The trigger adds no substantive evidence beyond the existing single-card quantized-model benchmark. Reproducibility across cards and direct validation of BF16, PCIe, and multi-GPU behavior remain unresolved, so further repetitive discussion does not advance the case.
2026-07-28T11:23:24Z
The added current-state thread chiefly reveals continued demand for real-world experience and benchmarks, not a second independent reproduction. The single-card result still does not establish card-to-card reliability, BF16 performance, PCIe behavior, or multi-GPU scaling.
2026-07-28T11:20:59Z
evidence attached: reddit.post.1v8uad7 — The current-state discussion bears directly on whether CMP 170HX unlocking is practical, though the limited engagement provides weak corroboration.
2026-07-28T08:27:20Z
The nominal evidence trigger adds no substantive reproduction or deployment result beyond the existing single-card quantized benchmark. The combined claim remains a hardware-lottery proposition pending card-to-card replication and direct BF16, PCIe, and multi-GPU validation.
2026-07-28T07:27:38Z
The refreshed discussion adds only modest amplification and no new physical reproduction, benchmark, or deployment result. The full claim remains supported by one single-card quantized-model benchmark, while card-to-card reproducibility, BF16 performance, PCIe behavior, and multi-GPU scaling remain unresolved.
2026-07-28T06:23:45Z
The nominal evidence update adds no substantive result beyond the existing single-card benchmark and planned deployment. Reproducibility across cards, BF16 performance, PCIe behavior, and multi-GPU scaling remain unverified, so the case stays a consequential hardware-lottery claim awaiting physical replications.
2026-07-28T05:22:22Z
No new reproduction or deployment result advances the case beyond the existing single-card benchmark. The full claim remains a hardware-lottery proposition pending card-to-card replication and direct validation of BF16, PCIe, and multi-card behavior.
2026-07-28T02:23:16Z
The buyer’s planned deployment creates a plausible path to replication but supplies no result yet. The full reproducibility claim remains supported by only one independent single-card benchmark, with BF16, PCIe, multi-card scaling, and card-to-card success still unresolved.
2026-07-28T02:21:08Z
evidence attached: reddit.post.1v8k1fl — A real buyer's planned multi-card deployment makes the unlock case's PCIe, scaling, and practical local-inference questions concrete, though it is not yet independent benchmark confirmation.
2026-07-25T18:26:51Z
No fresh reproduction or benchmark has extended the single-card result into a repeatable unlock, and BF16 plus PCIe behavior remain unvalidated. With only stale amplification, the case cools while awaiting results from additional physical cards.
2026-07-22T17:29:45Z
An independently authored benchmark repository now supports that at least one unlocked 64GB CMP 170HX can run a large quantized model at practical throughput, moving the case beyond discussion alone. It still does not establish repeatable success across cards or validate BF16 performance and PCIe functionality, so the full combined claim remains unsettled.
2026-07-22T17:21:43Z
evidence attached: reddit.post.1v3mf24 — Independent benchmark results provide useful corroboration that the unlocked CMP 170HX can run large quantized models at practical throughput.
2026-07-22T09:27:51Z
The newly attached material still provides no independent reproduction or benchmark validating the combined 64GB HBM, BF16, and PCIe unlock. The case remains a consequential but lottery-constrained project claim, with repetitive discussion adding no substantive confidence.
2026-07-22T08:24:27Z
Fresh comments remain anticipation, skepticism, and repetition rather than independent hardware reproduction. The claim still describes a potentially useful but lottery-constrained unlock, with no combined validation of usable HBM capacity, BF16 performance, and PCIe functionality.
2026-07-20T07:25:27Z
The new activity is minor amplification, not independent reproduction; no benchmarks validate the combined 64GB HBM, BF16, and PCIe claim. Keep this as a hardware-lottery project claim until arriving cards produce repeatable results.
2026-07-20T05:21:19Z
The latest reobservation adds only marginal engagement and no independent hardware reproduction, benchmarks, or validation of the combined 64GB HBM, BF16, and PCIe claim. The case remains a potentially consequential but lottery-constrained project claim rather than a demonstrated local-inference option.
2026-07-20T04:32:48Z
grounded: converges/medium — The claimed unlock converges with Scott’s usable-mass-over-unusable-power position and could materially expand his gamepc local-inference substrate through unus
2026-07-20T02:21:46Z
The reobservation adds no independent reproduction or benchmarks; the full 64GB HBM, BF16, and PCIe claim remains unverified and constrained by hardware-lottery concerns. Discussion is repetitive rather than meaningfully advancing the case.
2026-07-20T01:36:39Z
The discussion now frames the unlock as a hardware lottery with PCIe, bad-bank, and driver limitations rather than a reproducible 64GB inference solution. No independent benchmarks or replications verify the full HBM, BF16, and PCIe claim; current activity is mostly amplification and skepticism.
2026-07-20T01:35:44Z
evidence attached: reddit.post.1v15s4y — The reported reverse-engineering results bear directly on the claimed 64GB unlock, though the evidence is explicitly unverified and AI-assisted.
2026-07-19T11:24:19Z
case created — The unusually large hardware unlock is consequential for local inference but requires independent reproduction.