Shopping for Unlocked Mining Silicon

The NVIDIA CMP 170HX is a cryptomining card built on the same GA100 die as the A100: passive slab, no display outputs, and 1.49 TB/s of HBM2e that shipped artificially limited. A community unlock — writing feature-override registers after opening the protection layer via a GSP Falcon exploit — restores it to 64 GB. That makes it, on paper, the cheapest high-bandwidth VRAM money can buy. I wanted two for a local inference box.

“On paper” was carrying a lot of weight, so I did the market survey properly: every live listing across Facebook Marketplace (six metro areas), eBay, Alibaba, and AliExpress, with seller-fraud research on each. Findings, in ascending order of surprise:

The card barely trades socially. Three live Facebook listings existed across all six markets combined. This is not a liquid asset.

AliExpress inventory was 100% one scam operation. Every listing traced to a single storefront lineage — a repurposed jewelry shop, renamed, zero percent positive feedback — plus roughly fifty letter-prefixed clones. My favorite bait: a C$2,150 card with C$5,199 shipping. If you’ve seen a too-good 170HX price there, you’ve seen this cluster.

Facebook beat eBay in every cross-listed case. Same sellers, same physical cards, consistently cheaper off-platform — spreads from a few hundred dollars to nearly half price. Marketplace fees are real and sellers price them in.

AI-generated product photos are not a scam tell. The best-vetted seller in the whole survey used obviously generated hero shots; a confirmed fraud used stolen photographs of real inventory (0% positive feedback — 67 ratings across 350 sales — reusing a legitimate listing’s exact title). The photo’s provenance told me nothing; the feedback ledger did.

The market prices variants irrationally. 8 GB and 10 GB stock versions sell for the same money despite different claimed unlock ceilings — nobody bidding on these cards seems to price the thing that actually matters.

What it actually sold for

eBay exposes about ninety days of completed listings, so the whole available history is 81 completed sales between 21 June and 2 September 2026. That search returns 132 completed listings; the other 51 are fan shrouds, brackets, blower adapters and two entries that “sold” at C$1,710,039.59, which is placeholder pricing rather than a sale.

Fig. 1 — Every completed sale, coloured by the memory the seller advertised. Hollow rings are the 31 live asks on 3 Sep, beeswarmed at that date. The bracketed bars are the medians of the two well-sampled windows either side of the step. The grey dashed line is the least-squares fit through the sales — arithmetic, not a forecast. The green band is a judgement: a saturating curve from the current clearing price toward a ceiling set by the cheapest substitute that has actually sold.

The rise is a step, not a trend. Median C$1,776 across the 49 sales in the second half of July; C$3,343 across the 20 between 17 August and 2 September. That is +88% in six weeks, and it happens almost entirely in one jump around 9 August. Before it, a flat ~C$1,800 market with a floor cluster near C$820. After it, almost nothing clears under C$2,600.

The least-squares fit through all 81 sales gives +C$231 a week, and I quote that as arithmetic only. Its R² is 0.27 — it explains about a quarter of the variation, because prices stepped rather than drifted. Extended to year end the same line reaches C$7,188, which is what a line does if you keep drawing it, not something the market has promised.

Asks ran ahead of clearing prices. Median live ask across 31 buyable listings on 3 September was C$3,727, against C$3,224 where the fit had the market that day. Sellers had priced in more of the run-up than completed sales supported, which is the entire reason there was room to negotiate.

The variant premium is imaginary. 8 GB cards median C$1,914 (n=59), 10 GB cards C$1,798 (n=18) — and the 10 GB bin is the one advertised as unlocking to only 40 GB. The gap is a timing artefact: most 10 GB sales landed in the cheap July window and the August run-up is mostly 8 GB. Nobody is pricing the thing that actually matters, which is the unlock ceiling. Confirm the variant before you buy.

Where it landed

Two cards, negotiated in person at C$3,000 each — about 10% under the market median at the time, and the reason the negotiation was worth doing. Post-unlock that’s C$47 per GB of driver-supported, >1 TB/s VRAM, against C$52/GB for the same card at its market median.

Plotted against every comparable I could price from completed sales:

Fig. 2 — Sold medians: eBay.ca completed listings, delivered, from a 115-sale sweep on Sep 3–5, 2026. Hollow markers are thin samples (n<10). The dashed square is the one row that is neither a single card nor a sold median — a four-card 3090 build, quoted with no n, drawn differently because it is a different grade of evidence. † marks two samples whose size overstates their quality: the RTX PRO 6000 median sits above four fraudulent sub-C$4.1k "Brand New" listings, and the RTX 5090 sample mixes Founders Edition with premium AIB. Both are tail problems, which is why medians are used throughout. Re-checked against live asks on 2026-09-22: six of eight land within about 10% of this sweep; the A40 and the L40S are the two that don't, and both are n≤7.

Among single cards with current driver support, nothing is close: the cheapest is the RTX A6000 at C$117/GB, 2.5× the negotiated price. And capacity alone doesn’t set price — the A100 80 GB and the RTX PRO 6000 carry the most VRAM in the field and the worst cost per gigabyte in it.

But the honest comparison isn’t card-to-card, and it’s the one that nearly didn’t make this post. Four used RTX 3090s are 96 GB for about C$4,800, or ~C$50/GB — which undercuts the 170HX at its market median of C$52 and loses only to the C$47 I actually paid. On cost per gigabyte the mining card’s advantage over a stack of gaming cards is a rounding error, and any version of this chart that omits that row is flattering itself. It’s marked as a different grade of evidence above because it is one: a four-card figure with no sample size behind it, not a sold median.

What separates them isn’t price per gigabyte, it’s what the gigabytes are attached to. The 170HX gives you 64 GB of HBM2e at 1.49 TB/s behind one slot and one 8-pin; four 3090s give you 96 GB of GDDR6X at roughly 936 GB/s each, across four slots, four risers and about 1,400 W. The 170HX’s own catch is the link — PCIe 4.0 x4, which some sellers list as 1.0/2.0 x4. That throttles model loading and makes tensor-parallel across two cards a non-starter; in practice it’s layer-split only.

One card is missing from the chart on purpose. The MI100 is 32 GB at roughly C$44/GB — cheaper per gigabyte than anything plotted, including mine. It’s excluded on deprecation cadence rather than support: ROCm 7 dropped gfx906, which makes gfx908 the oldest surviving Instinct, and for a card you intend to hold two or three years that is a real risk in a way it isn’t for CUDA Ampere. That’s a judgement, not a spec, which is why it’s stated here rather than silently left out.

Background reading that earned its place: the VideoCardz spec sheet, and niconiconi’s teardown and watercooling notes — the single best independent documentation of this card. The register-override mechanism is documented for the sibling card in the cmp90hx-unlock repository and the surrounding NVIDIA-patcher discussion.

Update (Sep 8): both cards arrived, validated as genuine boards from the same production batch, matching power dongles included. The unlock, the external chassis, and what 1.49 TB/s actually does for MoE inference are future posts.