the demand for machines that turn a quarter into a dollar is effectively infinite. It is always rational to turn $0.99 into $1, only bounded by the risk that you can't do that in the future and the machine loses its value. But as there are more machines in the market, that $0.25 starts to approach the $0.99 return, because mining rewards are distributed across more GPUs so each has less expected return.
(and this is ignoring some weird effects like some miners being willing to run at a loss due to stolen electricity/using it to move money past currency controls/etc)
So in effect cryptomining has this super weird demand curve where the demand is infinite, then as you approach the break-even point it slows and stops, unless the price spikes again. And it can also flip and suddenly you're left with miners selling a bunch of used GPUs that undercut all the inventory you geared up to produce at 2x your normal sellthrough.
Literally the only way to cope with that is an instantaneous supply chain - you need to be able to produce enough GPUs to immediately saturate demand and push to the "nobody makes any profit anymore" stage of the curve. But that's not possible in silicon fabrication even in a perfectly ideal scenario - you are doing really well if (like NVIDIA) you can gear up to produce 2x your usual production. There just isn't enough VRAM production on the planet to produce 100x your normal quantity for 2 months, let alone the GPUs themselves. And then you'd have to ramp it all immediately back down too.
And even then, once miners are done they flip all that production back into the secondhand market (often with higher defect rates, the current batch are suffering from GPU cores that explode due to moisture from when miners powerwashed them, as well as memory failing due to prolonged high-temperature operation, and miners are repainting markings on dies to obscure this).
You can't build a supply chain to support that. It's not physically possible. Again, 2014 and 2017 showed this. And the inventory overhangs in 2015 and 2018 showed this too.
The only answer is to get miners onto their own line of products so that it doesn't starve the rest of the market for 2 years and then snap back and flood the market with excess inventory (both unsold and secondhand). People really don't like the idea of being told what they can do with their cards but you can't balance a supply chain around crypto boom/bust cycles, it is not physically possible.
Mining during your spare time is a consolation prize for GPUs being unaffordable, it just gets you back to where you started. And we are now seeing the downside of the inventory overhang resulting from mining - both AMD and NVIDIA are slow-walking the new generation to let existing inventory (mostly NVIDIA) burn through. And that happened in 2018 too.
Ethereum is done for now, but I am in the camp that there will likely be more PoW coins that evolve in the 5+ year timeframe and put us back into the same situation. NVIDIA needs to put perfcounters on their cores that look at alignment and memory coalescing. If ~0% of your memory accesses are aligned and ~0% are coalesced that's highly likely to be a mining workload, since everybody else at least tries to align/coalesce to benefit from higher effective bandwidth, but this is inherently impossible in proof-of-work algorithms since the algorithm is supposed to be making inherently random, unpredictable access patterns. So you can detect kernels that are doing inherently unaligned work, and back that up with a whitelist from the driver of things that you know are good (validated via TPM that the driver hasn't been tampered), and slow down everything not on that list that seems like a mining workload.
And I mean slow down significantly, not just slow down 50%. NVIDIA wasn't interested in stopping mining, they only slowed themselves down to the same level as AMD so they wouldn't be preferentially sucked off the shelves.
And it turns out the secondhand dumps from miners aren't particularly cheap either, and didn't magically give us 3080s too cheap to meter. I said before that people often didn't do the math in the 2017 boom either, prices really only dropped about 30% from pre-boom prices (which were already significantly below launch MSRP - 480 4GB/8GB were $150/175 ish in early 2017, 1080 was down to low $400s, etc). It's tough to get a read on what a "true" pre-boom price is for Ampere/RDNA2 since they launched during the pandemic, supply issues, and really into the start of the mining boom ramp-up, there isn't a good reference. But I think that "roughly 30% below the 'true' baseline" figure has held pretty true.
(and again, the miners did dump, that's why there's GPUs with the dies exploding because they powerwashed them! they're also even making their way back into the supply chain as new/refurbished at times, because a powerwashed and remarked GPU looks pretty dang clean!)
What was a legitimate problem was being preferentially sucked off shelves by miners. When NVIDIA was 2x as good as AMD, and it led to NVIDIA cards being priced 2x as high as AMD, that was a problem, they were taking heat in tech media and potentially losing users to AMD. And that was happening - prices lined up perfectly according to hashrate, and AMD cards had lower hashrates because they had narrower memory buses and infinity cache didn't help mining.
Solution: drop the hash rate by half. Now you are no longer preferentially being sucked off shelves and LHR cards should approximate the price of their AMD counterparts. And if there is a bust, you are both equally affected.
So it wasn't a fullhearted "we want miners OUT of the gaming market". If that was the case you'd have slowed it down by like 90-95%. They were happy to have the sales from miners, they just didn't want to suffer a competitive disadvantage relative to AMD due to miners gobbling up inventory. And if miners dumped a ton of inventory and pushed prices down, that would affect AMD equally too.
--
On the other hand that strategy kind of missed in that AMD simply opted not to participate in the GPU market, when consoles were eating up 80% of their wafers off the top and the CPU market could shift their remaining wafers at 10x the margin of GPUs. See my other comment about the impossible boom/bust cycle of GPUs and AMD's answer is: simply don't participate in the GPU market during this cycle.
It took almost a full year for any of the 6000 series cards to show up in Steam Hardware Survey at a mere 0.15%. NVIDIA had 7-8% of the market being 30-series around the same point - and I doubt 30-series are being used in LAN cafes in china, that is a 1060/1650 phenomenon not one that hits all cards equally. And actually 6700XT rapidly overtook all three of the higher cards combined. AMD just barely made any of the 6800/6800XT/6900 series at all, for the first year.
You can't blame them when one $649-MSRP 6800XT with 520mm^2 of silicon and $200-300 of other BOM cost, could be sold as six 5800X processors (81mm^2 CCD, GloFo 12nm IO die) each with an MSRP of $449, or as 3/4 of an Epyc 64C that sells for $10k MSRP, with essentially no BOM cost. You are talking about selling the same piece of silicon for $200, $2700, or $7500, it's not a hard choice to make. And they could absolutely sell the CPUs at or above MSRP, and that's after cranking MSRP and dropping value SKUs already. So they did a token launch (as late as possible) to satisfy their legal obligation to investors and went and made some fucking money, and simply chose not to participate in the GPU bubble this time.
Increasingly I think the discrete GPU market is just a backstop for AMD. It's a place to soak up excess wafers if they order too much for CPUs and Epyc, and it's a slush fund of wafers if they have unexpected demand from CPUs and Epyc and consoles. It's not something they put an R&D focus on (a major chunk of RNDA1 and RDNA2's early-stage R&D were largely funded by Sony and MS) and it's not something they put first in their manufacturing priority. It's also why they're lagging on the features side of the equation - consoles are paying for the R&D and consoles don't care about tensor cores or video encoder quality or optical flow engines, and won't pay for it in their chips or even fund the R&D.
dGPUs are just a sponge to soak up any excess in the rest of the business, or squeeze a little extra out of. Not a focus anymore. Maybe that changes if they're the first to multi-GCD and have a real shot to win, but, in the current scenario they get 80% of the revenue for 20% of the effort and that's what they're doing.
--
Now the other side of the "AMD simply chose not to participate in the GPU market" coin: it's all still a single market and AMD has to wait for the inventory to burn off too, even if it's NVIDIA inventory. OEMs like Gigabyte, Asus, MSI, etc aren't buying RDNA3 GPUs for their beige boxes or laptops if they have a bunch of Ampere sitting around, even if AMD inventory isn't high itself. Same for the secondary market, it doesn't matter if you are selling new RDNA3 cards if there are a bunch of used miner cards at great prices and people are buying used instead of new. It's all one market.
Some of the drop in marketshare despite deep price cuts is due to exactly this: OEMs aren't buying right now. Not buying anyone, but what's sitting in their stockpile is NVIDIA cards. And AMD can't make the big deals with OEMs that make up the majority of the market - the DIY market is a fraction of the overall market and AMD is selling great there (see: mindfactory numbers) due to aggressive pricing, but OEMs had piles of Ampere they need to get rid of. JPR did have some errors in the last set of numbers but still, the numbers for AMD are not great even then and that's the reason IMO, it's not DIY sales it's OEM sales.
I think the fact that they launched laptop Navi 33 in January but delayed the desktop until June is due to this: it's an inventory problem, and there is a lot of dGPU inventory still piled up. Not theirs, but that doesn't really matter.
It seems difficult to reconcile “we’re actively trying to stop cryptominers from buying our product” actions with “we’re against cryptomining only because they’re not buying our product” views.