Sidenote: Nvidia’s Branching Product Paths of Geforce and CMP, and The Bad Precedent It Sets

Earlier this week, Nvidia, clearly tired of weeks of headlines about how the current supply shortages and cryptocurrency mining craze had made their Ampere-architecture GPUs near-impossible to get for gamers, decided to counter-attack. Their plan? A fork in their GPU sales, taking chips that “don’t qualify to be Geforce GPUs” and shifting them to a new product line called CMP. CMP cards are intended for mining above all else, sold without display outputs and with less-desirable coolers intended to be used in a standard cryptomining operation where racks are spaced out with riser cables and often cooled through lower ambient temperatures via HVAC systems or other room-wide planning, or extreme measures like using a cave.

There was one other shift coming with this though – starting with the Geforce RTX 3060, which launches this week, Nvidia is using a mix of driver and BIOS-level detection of specific operating mechanics used by the Ethereum miner to reduce hash-rates by 50% on the gaming-focused cards, in an attempt to force miners to buy CMP cards instead.

Now, I want to start with my general impressions of cryptocurrency, mining, and the market it has spawned. The whole thing is kind of stupid in my opinion. Mining has such an absurdly large carbon footprint for no real value to society, and while crypto-enthusiasts will crow about the “democratization of money” or what have you, no cryptocurrency has any value of its own, as it is all tied to fiat currency. What makes crypto valuable is how much actual money someone will trade for it, and it has awful boom/bust cycles (just like the real economy under capitalism!). There doesn’t seem to be any real plans from people who tout crypto as the future for how it will even successfully integrate into a future society or who that leaves behind (cashless policies already have negative consequences for poor and unhoused individuals), and outside of all the real reasons, my personal distaste for it is that most people who love talking about crypto are huge douches about it and it has made getting my last two graphics cards a real pain in the ass.

But I do think Nvidia is doing the wrong thing here, and I want to contextualize why I believe that is.

The most immediate concern I had with this news is that chips aren’t exactly falling out of the sky, and introducing what is, effectively, a new product line based on the same silicon means that some amount of Geforce chips are still being taken out of gamer supply and fed into the CMP products. Nvidia claims that this isn’t the case and that there is no “split” of chips going to CMP versus Geforce, but the question then arises – where are the CMP chips coming from, exactly, especially since they are the same silicon? The obvious first answer to an analyst would be already-produced chips with silicon-level defects in the display controller hardware block, or the video encoder/decoder, or things of that nature where you can safely sell the chip as CMP since those blocks are not required there. That then raises the question of how many such chips even exist, brings up prior enthusiast questions about how good the yield is at Samsung on their 8nm process is that were previously dismissed by the company, and also raises a specter of the next steps. Yields improve over time, and with process revisions and product revisions, Nvidia and Samsung can collaboratively make it such that fewer GA10x silicon dies come back with defects at all, which then means that continuing manufacture of new CMP cards means either disabling a usable Geforce product for CMP or tapering the supply of CMP cards. The latter makes sense with the cyclical nature of the crypto-mining market, but depending on selling prices and expected profits, Nvidia could just as easily do the former, while insisting that the dies going to CMP cards just somehow aren’t making the cut. And, to be fair, that is possible, given that non-defect characteristics can affect silicon, like voltage/frequency curve, power response, leakage, and the like.

The actual largest concern I have though is the driver/BIOS-level limits and the precedent they set. With this move, Nvidia is basically placing an artificial limitation on what can or cannot be done with a card you buy. I wouldn’t be surprised if they pushed this limitation to existing Ampere cards either, given that next month, Nvidia is supposed to be offering existing RTX 3000-series owners an updated BIOS that will enable Resizable BAR for those cards. It would be a simple matter to also slip in that restriction – hopefully disclosing it prior to doing so. The reason I bristle at this change is that even if I think cryptomining is a wasteful exercise in futility, it is also a slippery slope from limiting it to limiting other applications or game performance over time. Nvidia already sort of has this segmentation, in that the Geforce driver stack already has reduced performance in many professional applications, and the Geforce hardware often uses the same GPU silicon as the pro-grade Quadro cards, but with reduced floating point computational capability. These segmentations are artificial and enforced by Nvidia alone with the interest of pushing certain audiences to higher-cost cards. It is why naming the RTX 3090 a “Titan-class” card is misleading bullshit at best, since the card does not have the prosumer optimizations that actual Titan cards have had and instead retains the segmentation.

There is another, more seedy, less surface-level concern here though. For this, cast your minds back to 2018. During the last crypto-craze, Nvidia’s Geforce GTX 1000 series cards were snapped up all the way until the early summer of that year for cryptomining. Nvidia announced their Turing-series RTX 2000 cards that August, and then something peculiar happened – no one really bought RTX 2000 cards. Not that they didn’t sell at all, but they didn’t sell anywhere near the volume that Nvidia had obviously hoped, which they themselves disclosed in investor calls. There were two reasons for this – the first being that Turing was a completely underwhelming performance improvement over those prior-gen cards, and the second is that because of that and the cryptocurrency crash, people were, instead, buying up used 10-series parts to put into gaming rigs, as cryptominers were selling off their used cards at prices that made Turing-based new cards look silly, because it was silly. You could buy a new RTX 2080 for $800, or you could buy an equivalent-or-better 1080 Ti for $500-$600. Sure, you missed out on RTX features, but in 2018, RTX was a punchline at best. Most of the RTX library launched that fall without RTX and were patched later with those features, so it was not really a thing at launch that was worth it, which also coincided with the peak of crypto resales. On top of all of that, RTX just generally wasn’t working all that well at launch, with even the 2080 Ti barely able to sustain a playable framerate with RTX on in most games, which led to the further advances on DLSS to the point where RTX became legitimately usable, at least in games with DLSS 2.0 and RTX.

So with CMP, Nvidia is now selling a card to miners that will not offer anything to gamers, at least in theory. Gaming cards bought for mining can be resold later and used for gaming. Depending on how the CMP cards are enabled (or disabled), there may be ways to make them work for gaming. Nvidia did offer display-less GPUs during the first mining craze, like the almost hidden P106 card, which were just gaming GPUs with no display outputs on the card and that refused to work with the Geforce driver stack. But a couple of simple software mods and the use of a motherboard-integrated display output could get them working for gaming – certainly not an ideal situation, but workable for an enthusiast. If CMP cards come just enabled enough to allow that, they could end up being steals on the secondhand market – but my suspicion here, and the broader point I’m getting at, is this – Nvidia is creating an artificial segmentation designed to shut off one of the biggest suppliers of second-hand GPU sales, such that gamers will be forced to buy new product from Nvidia or its partners, or compete for a reduced number of secondhand GPU sales.

So to me, this feels like a move designed to do two things primarily – it is intended to be a feel-good PR stunt (“we care so much about gamers guys, and we just want what is best for you!”) while creating a market condition where they can sell marked-up and costed-down GPUs to miners, which then have no resale value to gamers, forcing gamers to buy whatever Nvidia will push out in the coming years, regardless of how bad of a value proposition it actually ends up being, and doing it through the use of a software and hardware level lockout that artificially limits the GPU’s performance only under detected use of Nvidia-unapproved applications, which is itself a door that opens the way to all sorts of potential trickery and less justifiable performance modifications in the future.

So hey, nice try, but also…not great!

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

This site uses Akismet to reduce spam. Learn how your comment data is processed.