Instead of locking cards down, Nvidia could manage the supply. They could sell their cards in cooperation with Steam: Only people with an active steam account are eligible to a new card. If their performance stats don't improve after the purchase, which means they have resold their hardware, they won't be allowed to buy another card.

And then you have people who lose their cards, or it gets damaged, or who don't want Steam, etc. It's a big can of worms.

To balance the downsides, Nvidia doesn't have to go all in on 'Steam sales', like Sony, who sells some of their PS5 to registered gamers but not all of them.

On the other hand, I don't see the problem with damaged or lost cards. Damaged cards can be treated like guarantees: send in your damaged card and buy a new one. Lost cards on their own are a problem but how do you lose your card? You have to lose your entire gaming rig which most likely means that it is stolen. So send in a copy of your police report. That way, miners will be punished for wrong legal proceedings if they make a business out of 'losing cards'.

I have a baby. I do anything on Steam once a month at best. Crypto miners could pretty trivially make their use cases look like they spend more time gaming than a good 30% of Steam's active user base. This is a non-starter.

Highly unlikely. There might be reasonably reliable heuristics to identify gaming operations, but any signal for mining is likely to trigger for other long-process number-crunching too.

The idea puts me in mind of the other side: the police once raided a home because surveillance via helicopter identified, from the heat pattern, the property as like to have a loft-based cannabis farm. Turned out that the loft was full of equipment running coin miners. Simple checks get significant false positives.

But perhaps miners would just buy both mining and non-mining cards, since it's not a zero sum purchase, every card added to the stack is worthwhile.

Normal games and applications won't just max the whole memory bus at 100% for hours and hours. Even in a high-load situation it will still bounce around 90-100% or whatever. Mining absolutely slams the memory bus 100%, because current implementations are Proof Of Bandwidth so that's the stat you need to maximize.

Normal games or applications will go out of their way to align threads in a warp to access contiguous memory blocks ("coalesced memory access"), because that can be efficiently handled as a single request and broadcast to every thread. Mining can't do that because every thread is working in a totally different place in memory determined by the DAG. So normal applications will have lots of coalescing happening, mining will have zero.

Normal games or applications will have cache hits, because they're re-using at least some data frequently. Mining will have an (effectively) 0% cache hit rate because the DAG is sending threads to random places in memory and memory is much larger than cache.

These are simple things that can be metered with O(1) performance counters on the hardware. If the card sees a load like that, run in cripple mode, nuke the memory bus to 10% performance. Then you can also add the capability for a whitelist - the driver can attest that a certain binary has been signed by windows from the appropriate vendor, and TPM can attest that the driver and hardware hasn't been tampered with (Remote Attestation). So no need to worry about legacy code or DIY cuda code unless it happens to look like a mining application, and then you can whitelist any edge-cases.

TPM is going to be mandatory for OEM Windows 11 systems - home builders technically don't need it, but then if they happen to run some code that performs 100% random unaligned memory access you'd have to get it signed. But most code doesn't behave anything like that, so that is a real edge-case unless you are trying to write a miner.

You may be morally opposed to the above, but it's not impossible at a technical level, mining code actually looks nothing like normal code at a perfcounter level.