Nvidia's 3000 Series RTX GPU [3090s with different memory capacity]

[CarstenS]

I guess that depends on what kind of ROI plan you are. Even with the current, relative low of crypto/fiat rates, you still make money mining ETH even here in germany, where electricity costs are pretty much top 5 in the world, I guess (30 euro-cents per kWh, so aroun 36 us-cents is very common). I can only imagine how it must be with cheap or free electricity and no need for AC b/c of cold climate.

edit: Regarding the calculation of Geforces sold in the last quarter... There are a lot of notebooks, that don't end all in mining, but there are a lot of slower cards, GTX and Turing are still produced. Also, there are all the AMD cards, which mine very efficiently as well if tickled in the right places and there are a couple of other coins, where mining power is also being invested.

 

[Albuquerque]

Let's put some numbers together, shall we? Let's completly strawman the shit ouf of this.

Coinbase ETH Global Hashrate chart

Set the chart start date to September 17th 2020; the release date of the RTX 3080. You should find the global hashrate for ETH is 266.878 TH/s. As of the date I'm writing this post (June 11th 2021) the global hash rate I see as 637.818 TH/s. Let's make it even a more worst case for my argument, let's instead use the absolute highest value on the chart: 676.820 TH/s on May 21st of 2021.

Beginning value, before any 30xx series GPUs were in the public's hands: 266.878 TH/s
Peak value, after almost precisely eight months of 30xx series GPU sales: 676.820 TH/s
A total of 409.942 TH/s of compute was added to mining ETH in a time span of near-exactly eight months.

Let's also assume the absolutely worst case, that literally EVERY GPU added to mining ETH was an NVIDIA Ampere and not AMD. Let's then contrive the argument even moreso, we'll use the slowest mining Ampere: the RTX 3060 (even though it wasn't actually available the whole time.) Why are we doing this? Because it will result in the most video cards needing to be sold in order to make the argument most favorable to a statement of "miners are buying all the cards!!" According to Mistake Enables Full GeForce RTX 3060 ETH Mining - Over 50 MH/s - Legit Reviews Getting Over 50 MH/s on the GeForce RTX 3060 we could see around 42MH/s without any tuning. To be the most gracious to my opponent's argument, we will completely ignore the tuning potential for both power efficiency and mining rate increases that most miners are likely to deploy.

Ok, so to get 409.942 TH/s of mining out of only 3060 cards mining at only a 42 MH/s rate, we will need just shy of ten million (9,760,524) cards doing the mining! Holy shit, ten million cards!

Getting precise unit shipment numbers of each dGPU is really difficult because nobody truly publishes precise numbers. What we must do is tease the numbers apart from various published market watch research groups, like this one: GPU shipments soar in Q1 year-over-year | Jon Peddie Research

What does that article tell us?

• Direct quote from the article: According to a new research report from the analyst firm Jon Peddie Research, the growth of the PC-based Graphics Processor Units (GPU) market reached 119 million units in Q1'21
• Of that 119 million, 15.17% of all shipped GPUs came from NVIDIA. 119,000,000 * 0.1517 == 18 million NV GPUs shipped in the first quarter of this year.
• You'll notice we also have quarterly sales data posted which goes back Q4'20, and the article notes NVIDIA's shipments in Q1 were 3.9% higher than last quarter. 18,000,000 / 1.039 == 17.4 million NV GPUs shipped in Q4 which covers the initial 30xx series launch.

A few things we don't have:

• No data on Q2'21 sales, which is the quarter ending in about 19 days. We can reasonably extrapolate Q2 sales will be at least equal to Q1 given the trendline, albeit we aren't done with the quarter just yet. Roughly we're about 5/6ths (83.3%) of the way through Q2, so let's swag 18,000,000 * 0.83 == about 15MM NV GPUs have probably shipped by this point.
• An argument could be made that shipped GPUs are not directly linked to sold AIB cards inside the same timeframe, because there is time neded to glue it all together. What that really means is we'd have to calculate GPUs shipped even before Q4'20 and have to figure out how to subtract it from Q2'21. I'm not sure it's going to materially alter the outcome here and we have literally no data to help create an estimate. The shift might affect the numbers by the couple-hundred-thousands, but isn't going to materially move the needle.
• There is no published comprehensive sub-breakdown of which of those shipped 50 million GPU's were 3080's, or 3070's, or 2080's, or 1660's, or 730's, or whatever. What about Quadro too?

We could try to extrapolate percentages of shipments correlated to the sold unit numbers we see on eBay as tracked by Tom's Hardware GPU Pricing Index: Tracking Graphics Cards Sold on eBay | Tom's Hardware (tomshardware.com). A reasoned argument could be made of total available cards getting scalped on eBay will be correlated to the availability of those cards in the retail sales channel. Funny enough, 3060's make up roughly 20% of the 3000-series cards sold on eBay. Not really sure what that tells us, the 20% number is funny as you'll see below...

What are we left with here?

The ETH rate increased by 409TH/s since September the 17th. If every card added to ETH mining was only NVIDIA Ampere (it wasn't), and was only the slowest-mining Ampere model without any tuning at all (it wasn't), we'd only need to consume 20% of the total GPUs shipped by NV in the last two quarters and the portion of this quarter (they didn't.)

What did we learn?

Miners aren't ruining the GPU economy. Let's stop perpetuating the bullshit already.

 

[trinibwoy]

Let's put some numbers together, shall we? Let's completly strawman the shit ouf of this.

Coinbase ETH Global Hashrate chart

Set the chart start date to September 17th 2020; the release date of the RTX 3080. You should find the global hashrate for ETH is 266.878 TH/s. As of the date I'm writing this post (June 11th 2021) the global hash rate I see as 637.818 TH/s. Let's make it even a more worst case for my argument, let's instead use the absolute highest value on the chart: 676.820 TH/s on May 21st of 2021.

Beginning value, before any 30xx series GPUs were in the public's hands: 266.878 TH/s
Peak value, after almost precisely eight months of 30xx series GPU sales: 676.820 TH/s
A total of 409.942 TH/s of compute was added to mining ETH in a time span of near-exactly eight months.

Let's also assume the absolutely worst case, that literally EVERY GPU added to mining ETH was an NVIDIA Ampere and not AMD. Let's then contrive the argument even moreso, we'll use the slowest mining Ampere: the RTX 3060 (even though it wasn't actually available the whole time.) Why are we doing this? Because it will result in the most video cards needing to be sold in order to make the argument most favorable to a statement of "miners are buying all the cards!!" According to Mistake Enables Full GeForce RTX 3060 ETH Mining - Over 50 MH/s - Legit Reviews Getting Over 50 MH/s on the GeForce RTX 3060 we could see around 42MH/s without any tuning. To be the most gracious to my opponent's argument, we will completely ignore the tuning potential for both power efficiency and mining rate increases that most miners are likely to deploy.

Ok, so to get 409.942 TH/s of mining out of only 3060 cards mining at only a 42 MH/s rate, we will need just shy of ten million (9,760,524) cards doing the mining! Holy shit, ten million cards!

Getting precise unit shipment numbers of each dGPU is really difficult because nobody truly publishes precise numbers. What we must do is tease the numbers apart from various published market watch research groups, like this one: GPU shipments soar in Q1 year-over-year | Jon Peddie Research

What does that article tell us?

• Direct quote from the article: According to a new research report from the analyst firm Jon Peddie Research, the growth of the PC-based Graphics Processor Units (GPU) market reached 119 million units in Q1'21
• Of that 119 million, 15.17% of all shipped GPUs came from NVIDIA. 119,000,000 * 0.1517 == 18 million NV GPUs shipped in the first quarter of this year.
• You'll notice we also have quarterly sales data posted which goes back Q4'20, and the article notes NVIDIA's shipments in Q1 were 3.9% higher than last quarter. 18,000,000 / 1.039 == 17.4 million NV GPUs shipped in Q4 which covers the initial 30xx series launch.

A few things we don't have:

• No data on Q2'21 sales, which is the quarter ending in about 19 days. We can reasonably extrapolate Q2 sales will be at least equal to Q1 given the trendline, albeit we aren't done with the quarter just yet. Roughly we're about 5/6ths (83.3%) of the way through Q2, so let's swag 18,000,000 * 0.83 == about 15MM NV GPUs have probably shipped by this point.
• An argument could be made that shipped GPUs are not directly linked to sold AIB cards inside the same timeframe, because there is time neded to glue it all together. What that really means is we'd have to calculate GPUs shipped even before Q4'20 and have to figure out how to subtract it from Q2'21. I'm not sure it's going to materially alter the outcome here and we have literally no data to help create an estimate. The shift might affect the numbers by the couple-hundred-thousands, but isn't going to materially move the needle.
• There is no published comprehensive sub-breakdown of which of those shipped 50 million GPU's were 3080's, or 3070's, or 2080's, or 1660's, or 730's, or whatever. What about Quadro too?

We could try to extrapolate percentages of shipments correlated to the sold unit numbers we see on eBay as tracked by Tom's Hardware GPU Pricing Index: Tracking Graphics Cards Sold on eBay | Tom's Hardware (tomshardware.com). A reasoned argument could be made of total available cards getting scalped on eBay will be correlated to the availability of those cards in the retail sales channel. Funny enough, 3060's make up roughly 20% of the 3000-series cards sold on eBay. Not really sure what that tells us, the 20% number is funny as you'll see below...

What are we left with here?

The ETH rate increased by 409TH/s since September the 17th. If every card added to ETH mining was only NVIDIA Ampere (it wasn't), and was only the slowest-mining Ampere model without any tuning at all (it wasn't), we'd only need to consume 20% of the total GPUs shipped by NV in the last two quarters and the portion of this quarter (they didn't.)

What did we learn?

Miners aren't ruining the GPU economy. Let's stop perpetuating the bullshit already.

Science, I like it. Are GPUs only being used for ETH mining or are there other cryptos that could be consuming inventory?

 

[Albuquerque]

Science, I like it. Are GPUs only being used for ETH mining or are there other cryptos that could be consuming inventory?

Absolutely there are, however they're a tiny fraction of the hashrate of ETH, mostly because it's been the profitable coin to mine. Using my same dates and absolute-worst-case method from above, Bitcoin moved from ~130TH/s to 190TH/s in the same timeframe and a 3060 mines BTC at what appears to be the same rate as ETH (with tuning, around 50MH/s.)

If we use the same contrived arguments of ONLY Ampere, and ONLY the slowest one, math says you'd need 1.2 million 3060's to make that BTC hash increase. It's 2% of the shipped GPUs in the last three quarters.

To roughly quote Richard Dawkins: "Science, it works. Bitches!" (Richard Dawkins - Science works bitches! - YouTube)

Miners aren't ruining the GPU economy.

 

[PSman1700]

Gamers are simply willing to pay thousands of dollars to be able to have the latest GPU's.

 

[DegustatoR]

Gamers are simply willing to pay thousands of dollars to be able to have the latest GPU's.

Yeah, whole four of them on the planet.

 

[Albuquerque]

Yeah, whole four of them on the planet.

Truly this is the crux of the matter. If you wanna play games on a modern-ish GPU in the last year(ish), then you get to fork out the moolah or you simply don't get to play.

And again, normally these prices couldn't possibly be sustainable for gamers. However, how much money have so many people saved in the last 16 months after not driving to work (and thus not paying for gas or maintenance items), not wearing out their nice clothes, not eating out, not taking vacations? I personally have easily saved $10k in all of the above, I can catalog it if you really want to know. Would I willingly pay $2000 for a gaming video card two years ago? Hell no! Would I today? Honestly, I've been strongly considering it, mostly because the single newest component on my gaming rig is four years old: 3930k at 4.5GHz, 32GB of DDR3/1600, a Sammy 850 EVO 1TB, a Sammy 860 EVO 2TB, a Dell U2711 to watch it all, and the newest piece: an Aorus 1080Ti OC card.

Would I try to ratioinalize taking the $10k I saved over the past sixteen months of COVID and spend half of it on a gaming rig? It's really tempting. I'm still trying to make myself wait until later in the year, just to see, since I really don't have any one specific game that's pushing me to the dark side (yet.)

God(s) help my wallet if TES VI comes out before then

 

[DegustatoR]

Truly this is the crux of the matter. If you wanna play games on a modern-ish GPU in the last year(ish), then you get to fork out the moolah or you simply don't get to play

Well that's the point. You don't need Ampere to play latest games. You don't need Turing even. So those who are buying these GPUs for X4 of their MSRP for gaming are few and less than that even. And this does actively hurt the market since those who would be upgrading now to a DX12U capable GPU don't, they still wait on their Pascals and Vegas. Which means that we're a year or so behind the point where DX12U feature set would become the baseline. And this is a big issue for both h/w manufacturers and gamers.

 

[Albuquerque]

You're right, you don't need the latest card. Yuo bet you kinda want it though, right? And how sure can you be of a used card, especially one that's four years old (anything Pascal?) Who is to say the prior owner didn't cook that thing to nearly it's death trying to get the highest possible overclock? I wouldn't be in the market for a used card at all.

Someone up in the thread wanted to beat me up about "nuance." Think about one popular segment of gamers who were most likely to keep their well-paying jobs during the push to a remote workforce during COVID: IT professionals, like myself. I make good money, I continued making good money, and yet I had almost nothing to spend it on. For the majority of the population who weren't in the six-figures salary club, at least here in the US, they also picked up a decent chunk of COVID stimulus money too.

What else are ya gonna spend that money on? Those two free COVID stimulus checks looks like a mighty fine 3080 I wager.

 

[pcchen]

Well that's the point. You don't need Ampere to play latest games. You don't need Turing even. So those who are buying these GPUs for X4 of their MSRP for gaming are few and less than that even. And this does actively hurt the market since those who would be upgrading now to a DX12U capable GPU don't, they still wait on their Pascals and Vegas. Which means that we're a year or so behind the point where DX12U feature set would become the baseline. And this is a big issue for both h/w manufacturers and gamers.

Well, it's not that bad even into this year. One of my friend who got his 3070 (urged by me, to replace his aging GTX 970) at MSRP, and that's in 2021.
Also one can take a look at Steam's hardware survey. It's not very accurate, of course, but 3070 is the most popular 30 series GPU on it, at 1.42%, which is on par with Radeon RX 570 and GTX 970. Even 3080 is at 0.86%, more than 2080 and 2080 Ti.

 

[Albuquerque]

Some day I'm gonna get lucky and find a 3080 FE at Best Buy and pay MSRP

 

[DegustatoR]

New You're right, you don't need the latest card. Yuo bet you kinda want it though, right?

Yeah, until you go and see the prices. Gamers don't buy these at these prices, it's ridiculous. I've got me a 3080 last year prior to things going stratospheric and I was thinking that I'm paying a bit to much for it. Now the same card cost 3X what I've paid for it and I can't see myself paying that any time at all - and I am an enthusiast. So I wouldn't count on gamers buying a lot of these at current prices really.

 

[Albuquerque]

Gamers don't buy these at these prices, it's ridiculous. < snip > So I wouldn't count on gamers buying a lot of these at current prices really.

And where are they going then? I've shown you the math, there's simply too many cards sold and not enough tangible evidence of mining hashrate to make up a story about it being miners. Since basic algebra has functionally disproven that hypothesis, where do you think those GPUs are going?

Also one can take a look at Steam's hardware survey. It's not very accurate, of course, but 3070 is the most popular 30 series GPU on it, at 1.42%, which is on par with Radeon RX 570 and GTX 970. Even 3080 is at 0.86%, more than 2080 and 2080 Ti.

I think these percentages of both 3070 and 3080 in the Steam survey, when compared to their older bretherin. says quite a bit, Yeah, there are statistical errors in the Steam survey; those errors apply across all generations, not just uniquely to the Ampere line.

 

[ninelven]

Probably to the people buying them by the pallet. Now, I don't know what one does with a pallet or entire shipping container of gpus, but it probably isn't playing crysis.

 

[CarstenS]

Let's put some numbers together, shall we? Let's completly strawman the shit ouf of this.

Coinbase ETH Global Hashrate chart

Set the chart start date to September 17th 2020; the release date of the RTX 3080. You should find the global hashrate for ETH is 266.878 TH/s. As of the date I'm writing this post (June 11th 2021) the global hash rate I see as 637.818 TH/s. Let's make it even a more worst case for my argument, let's instead use the absolute highest value on the chart: 676.820 TH/s on May 21st of 2021.

Beginning value, before any 30xx series GPUs were in the public's hands: 266.878 TH/s
Peak value, after almost precisely eight months of 30xx series GPU sales: 676.820 TH/s
A total of 409.942 TH/s of compute was added to mining ETH in a time span of near-exactly eight months.

Let's also assume the absolutely worst case, that literally EVERY GPU added to mining ETH was an NVIDIA Ampere and not AMD. Let's then contrive the argument even moreso, we'll use the slowest mining Ampere: the RTX 3060 (even though it wasn't actually available the whole time.) Why are we doing this? Because it will result in the most video cards needing to be sold in order to make the argument most favorable to a statement of "miners are buying all the cards!!" According to Mistake Enables Full GeForce RTX 3060 ETH Mining - Over 50 MH/s - Legit Reviews Getting Over 50 MH/s on the GeForce RTX 3060 we could see around 42MH/s without any tuning. To be the most gracious to my opponent's argument, we will completely ignore the tuning potential for both power efficiency and mining rate increases that most miners are likely to deploy.

Ok, so to get 409.942 TH/s of mining out of only 3060 cards mining at only a 42 MH/s rate, we will need just shy of ten million (9,760,524) cards doing the mining! Holy shit, ten million cards!

Getting precise unit shipment numbers of each dGPU is really difficult because nobody truly publishes precise numbers. What we must do is tease the numbers apart from various published market watch research groups, like this one: GPU shipments soar in Q1 year-over-year | Jon Peddie Research

What does that article tell us?

• Direct quote from the article: According to a new research report from the analyst firm Jon Peddie Research, the growth of the PC-based Graphics Processor Units (GPU) market reached 119 million units in Q1'21
• Of that 119 million, 15.17% of all shipped GPUs came from NVIDIA. 119,000,000 * 0.1517 == 18 million NV GPUs shipped in the first quarter of this year.
• You'll notice we also have quarterly sales data posted which goes back Q4'20, and the article notes NVIDIA's shipments in Q1 were 3.9% higher than last quarter. 18,000,000 / 1.039 == 17.4 million NV GPUs shipped in Q4 which covers the initial 30xx series launch.

A few things we don't have:

• No data on Q2'21 sales, which is the quarter ending in about 19 days. We can reasonably extrapolate Q2 sales will be at least equal to Q1 given the trendline, albeit we aren't done with the quarter just yet. Roughly we're about 5/6ths (83.3%) of the way through Q2, so let's swag 18,000,000 * 0.83 == about 15MM NV GPUs have probably shipped by this point.
• An argument could be made that shipped GPUs are not directly linked to sold AIB cards inside the same timeframe, because there is time neded to glue it all together. What that really means is we'd have to calculate GPUs shipped even before Q4'20 and have to figure out how to subtract it from Q2'21. I'm not sure it's going to materially alter the outcome here and we have literally no data to help create an estimate. The shift might affect the numbers by the couple-hundred-thousands, but isn't going to materially move the needle.
• There is no published comprehensive sub-breakdown of which of those shipped 50 million GPU's were 3080's, or 3070's, or 2080's, or 1660's, or 730's, or whatever. What about Quadro too?

We could try to extrapolate percentages of shipments correlated to the sold unit numbers we see on eBay as tracked by Tom's Hardware GPU Pricing Index: Tracking Graphics Cards Sold on eBay | Tom's Hardware (tomshardware.com). A reasoned argument could be made of total available cards getting scalped on eBay will be correlated to the availability of those cards in the retail sales channel. Funny enough, 3060's make up roughly 20% of the 3000-series cards sold on eBay. Not really sure what that tells us, the 20% number is funny as you'll see below...

What are we left with here?

The ETH rate increased by 409TH/s since September the 17th. If every card added to ETH mining was only NVIDIA Ampere (it wasn't), and was only the slowest-mining Ampere model without any tuning at all (it wasn't), we'd only need to consume 20% of the total GPUs shipped by NV in the last two quarters and the portion of this quarter (they didn't.)

What did we learn?

Miners aren't ruining the GPU economy. Let's stop perpetuating the bullshit already.

Thank you for your calculations, not so much though for spinning diverging opinions into the strawman realm.
That said, it did not take into account any of the points I would love to consider, before declaring victory and jumping to conclusions. To spell it out very clearly: I neither argue nor support your math yet, I simply am under the impression, that there are a lot of points not considered. Some of them, because we simply don't have the data - so, not your fault.

One example. there's still no data about the share of notebook GPUs. Notebooks themselves, AFAIK, have passed stationary PCs in sheer numbers. They most likely do have a lower attach rate of gaming-grade dGPUs. Yet, according to the JPR report you linked, there were 89M Notebooks shipped in Q1'21 ("notebook shipments exceeded 89 million units for the quarter"), out of a total Market of 119M units ("PC-based Graphics Processor Units (GPU) market reached 119 million units in Q1'21").

Can we infer from that a number of 75% of "GPUs" being sold in the Notebook market? I don't know for example if JPR does count android-based chromebooks into the 89M figure, if so how many of them and then, whether or not they also count toward the total number of "GPUs" sold. If we had to factor out Notebook GPUs, we'd also have to re-evaluate Nvidias 15% overall GPU market share and rather go by 80-ish percent JPR claims for PC-dGPU as baseline (~24M GF-units) and still figure out finer intricacies. For example the ratio of cheaper, older cards which most likely don't get bought now for mining (everything below 6 GByte, except you have free electricity and bet on Ravencoin) - how is the overyll mix? Why are GTX1600's with less than 4 GByte also ridiculously expensive?
dGPUs also include Quadros, which you also mentioned. Probably not that many though. And that's only one example of - IMHLO - very unclear data on which we would base other calculations.

Again, I am not arguing for or against miners being the culprits, but i am simply pointing out how hard it is to even guess educatedly.

edit: Looking at the ever controversial steam survey percentages, we have ~3,6 % RTX3000 users on that gaming platform. If - and that's a big IF - this is any indication from the 120M monthly active steam users, we'd be at 4,4M RTX3K-GPUs. Not every gamer installs steam though, some of it certainly comes from internet-/gaming cafees, some maybe even from steam installs in Nvidias GeForce NOW service. More uncertainties.

Here's what I gathered from the numbers steam showed me:
3090 0,38
3080 0,9
3070 1,48
3060 Ti 0,4
3060 0,27
3060 Laptop 0,21
(percentages)

 

[pcchen]

Can we infer from that a number of 75% of "GPUs" being sold in the Notebook market? I don't know for example if JPR does count android-based chromebooks into the 89M figure, if so how many of them and then, whether or not they also count toward the total number of "GPUs" sold.

From the wording I'd say it's more likely to have Chromebooks included.

Due the rush to try and equip school children with Chromebooks, combined with the work-at-home rules due to the COVID-19 pandemic, notebook shipments exceeded 89 million units for the quarter

I have to say, of course there are many uncertainties from the numbers, but the order of magnitude is probably right. Also the "GPU mining caused all the price increase" camp has even less evidence, so this is a good start.

 

[HLJ]

I don't look at the price when building a gaming PC.
I have disposable income and about to get more (accepted a new job)

Demand and supply ‍

 

[Sxotty]

I was going to get a new card finally this round then we all know what happen need. I don't really mind waiting. Just getting a huge backlog of games i got on cheap sales and now all the free games. Whenever I finally upgrade I'll have years to catch up on I think.

 

[Albuquerque]

I have to say, of course there are many uncertainties from the numbers, but the order of magnitude is probably right. Also the "GPU mining caused all the price increase" camp has even less evidence, so this is a good start.

This was precisely the point of my post -- not to provide rock hard numbers, but instead to provide the absolute worst case scenario I could construct for my own argument and then show it's still not a reasonable stance to say "Miners are ruining the GPU economy." At the end of the day, miners create obvious work that we can track -- there simply isn't enough work being generated to account for even half of the GPU sales, no matter how far down the rabbit hole we want to deconstruct those sales numbers.

Now, does this mean miners aren't part of the overall consumption of GPU sales? Absolutely they are! Anyone buying cards is going to feed into the demand side, which still feeds into the pricing. it is a certainty that miners do have an effect on pricing, without question.

However, prior statements of "basically 100%" are equally "basically 100%" unfounded and unsupported by logic.

 

[ninelven]

On the other hand, the techpowerup poll exists.