News Nvidia GPU Share Continues to Increase, According to Steam

Apr 1, 2020
1,390
1,046
7,060
It's not that Steam's data is SUSPECT, it's just that it's a sample of a sample of the world market. Valve reported back in January that Steam has 120 million monthly users worldwide, which is a fairly representative sample size in most countries Steam is available in, but remember that not all users participate in the hardware survey, and in countries outside the USA, especially APAC countries such as China and Vietnam, internet cafés are much more prevalent than here in the USA, and that's something that Steam directly targets (Steam PC Café). It was reported last year that internet cafés were using their machines to mine cryptocurrency while they were shut due to Covid restrictions, so if a cafe gets 100 nVidia cards in so they can flog away at ether, those machines are still going to be running Steam for gaming as well, so the user share of nVidia increases.

The real test is going to be when Intel gaming GPUs start getting on the market...
 

InvalidError

Titan
Moderator
It's not that Steam's data is SUSPECT, it's just that it's a sample of a sample of the world market.
Probably the biggest sample that anyone has access to short of other publishers who keep their stats for themselves.

I find it a little weird that many people periodically say that the Steam survey is biased because they claim they have never received a survey participation request or haven't received one in a very long time. I receive one every 3-4 months and have been receiving those for as far back as I remember the survey becoming a thing. I got my third one for this year, second one since upgrading my PC just yesterday.

As for Asian internet cafes multi-counting on the survey, I'd imagine this wouldn't be happening to the usual extent while most cafes are closed and I'd hope the survey data has some sort of fingerprint (ex: Steam install ID) to remove most duplicate entries from the final tallies.
 

GenericUser

Distinguished
Nov 20, 2010
294
139
18,990
Probably the biggest sample that anyone has access to short of other publishers who keep their stats for themselves.

I find it a little weird that many people periodically say that the Steam survey is biased because they claim they have never received a survey participation request or haven't received one in a very long time. I receive one every 3-4 months and have been receiving those for as far back as I remember the survey becoming a thing. I got my third one for this year, second one since upgrading my PC just yesterday.

As for Asian internet cafes multi-counting on the survey, I'd imagine this wouldn't be happening to the usual extent while most cafes are closed and I'd hope the survey data has some sort of fingerprint (ex: Steam install ID) to remove most duplicate entries from the final tallies.

I haven't been requested to do a hardware survey in about 6 or 7 years. Honestly curious about what their criteria is for deciding which users to ping for one (though not to imply accusations of bias).
 

spongiemaster

Admirable
Dec 12, 2019
2,273
1,277
7,560
It's not that Steam's data is SUSPECT, it's just that it's a sample of a sample of the world market. Valve reported back in January that Steam has 120 million monthly users worldwide, which is a fairly representative sample size in most countries Steam is available in, but remember that not all users participate in the hardware survey, and in countries outside the USA, especially APAC countries such as China and Vietnam, internet cafés are much more prevalent than here in the USA, and that's something that Steam directly targets (Steam PC Café). It was reported last year that internet cafés were using their machines to mine cryptocurrency while they were shut due to Covid restrictions, so if a cafe gets 100 nVidia cards in so they can flog away at ether, those machines are still going to be running Steam for gaming as well, so the user share of nVidia increases.

The real test is going to be when Intel gaming GPUs start getting on the market...
The internet cafe multi-counting problem occurred a couple of years back and Steam fixed the problem. Outside of that, the data is suspect simply because Valve has never said how it samples. If you don't do a pure random sampling — for example, if you bias toward sampling "new" or "unknown" hardware, or even PCs where the hardware configuration has changed since the last sample — then it screws up the statistics in a very bad way. Steam doesn't disclose how many PCs were sampled each cycle either. Theoretically, it could be "all PCs connected to Steam," but then why actually ask for people to opt-in on the HW survey? So, it's a possibly not random sampling, and that's by far the biggest issue.

As for Steam not representing all users, that's less of an issue. Presumably there's very little correlation between AMD GPU owners not using Steam and Nvidia owners not using Steam — though Intel GPU users not using Steam wouldn't be as surprising (eg, a lot of business PCs don't ever run Steam). We're mostly interested in gaming GPUs used to play games, and Steam is by far the most popular gaming distribution service. If Steam China takes over for regular Steam in that country and its users no longer count toward the totals, that would be a potential problem, but AFAIK regular Steam still works in China. If anything, I'd expect China to skew more heavily toward Nvidia (it's a more respected brand there, again AFAIK), with or without Internet cafes.

I should note that the overall percentages of the market, according to Steam's DX12 data, are 80.48% Nvidia, 14.81% AMD, 3.98% Intel, and then 0.73% "other." It's interesting that the Intel percentage is so low, as everything with UHD 500 or later should be DX12 compliant I think. I suppose most people with Skylake or later CPU aren't using the integrated graphics? Or maybe it needs to be Kaby Lake or later?

Another interesting fact: If you look at the Vulkan API numbers, things get really screwy. Go ahead and add up the first 15 or so percentages in the Vulkan GPU list. https://store.steampowered.com/hwsurvey/directx/ I'll help you out. For August, the first 15 GPUs in the Vulkan list sum up to 103.70%. Oops! In fact, the entire list of 384 GPUs sums up to 194.80%! Clearly something is fubar with the way Steam presents that data. The DirectX 12 GPUs, all 212 of them, sum up to 84.44% for August as well, which indicates a different sort of problem. I divided all of the numbers in the list by the total percentage in order to normalize things to 100%, but the data might still be wrong.

Fundamentally, Steam may collect data incorrectly (non-random sampling), and the API list doesn't properly sum up to 100%. We're looking at ALL GPUs that can run a specific API, so it should be 100%, and it's not. That's bad and reduces confidence in the underlying statistics. Basically, you can't trust the Steam Hardware Survey, even if it's the best publicly accessible set of data that we have.
 

InvalidError

Titan
Moderator
If you don't do a pure random sampling — for example, if you bias toward sampling "new" or "unknown" hardware, or even PCs where the hardware configuration has changed since the last sample — then it screws up the statistics in a very bad way.
I've been getting the Steam Survey every 3-4 months for as long as I remember it existing despite my computer not changing in any way whatsoever for 2-4 years at a time, so I can definitely say that hardware changes do not appear to be a factor whatsoever.

Maybe they're just using a bad PRNG that tends to spit out some subset of account numbers far more often than others, so you end up with some people like me who seem to get the survey at practically every single update period and others who hardly ever get it. They could also be applying negative bias to accounts that have declined participation in the past
 
doesn't matter what kind of numbers steam survey are showing. just look at JPR reports.
JPR reports can show total sales, according to their sources, but can't differentiate between gamers and miners and other users. And we do report on the data JPR releases, which was linked in this news bit.
I've been getting the Steam Survey every 3-4 months for as long as I remember it existing despite my computer not changing in any way whatsoever for 2-4 years at a time, so I can definitely say that hardware changes do not appear to be a factor whatsoever.

Maybe they're just using a bad PRNG that tends to spit out some subset of account numbers far more often than others, so you end up with some people like me who seem to get the survey at practically every single update period and others who hardly ever get it. They could also be applying negative bias to accounts that have declined participation in the past
I get it fairly regularly, but oddly it often comes in batches. I remember a month or two back, I had three or four different PCs all get hit with the HW survey all at the same time. And that has happened before. At least two of them were "new/unknown" GPUs — one was the Intel Xe DG1 card, another was using a new RTX 3070 Ti I think. Anyway, given I have five or so PCs regularly connecting to Steam, none of them got sampled for several months, and then several of them got hit at the same time... my anecdotal evidence suggests the survey is NOT purely random and is somehow tied to hardware or user ID or some other factor. In fact, if you get sampled once every 3-4 months on just one PC, consistently, that's again a sign that it's not random. That's a pattern.
 

InvalidError

Titan
Moderator
I get it fairly regularly, but oddly it often comes in batches. I remember a month or two back, I had three or four different PCs all get hit with the HW survey all at the same time.
That doesn't surprise me at all: the Steam survey is a quarterly thing, so I would expect a burst of survey requests to go out at whatever time the survey sampling period starts and people who have multiple PCs with Steam running to get up to as many survey requests.
 
That doesn't surprise me at all: the Steam survey is a quarterly thing, so I would expect a burst of survey requests to go out at whatever time the survey sampling period starts and people who have multiple PCs with Steam running to get up to as many survey requests.
How do you figure? It gets posted with new data every month, meaning it would be a monthly thing, not quarterly. I think it might sample specific PCs individually — again, that's NOT random! — but the data definitely changes every month. I should start keeping a record of precisely when I get sampled and on what hardware, just for kicks and giggles.
 
The internet cafe multi-counting problem occurred a couple of years back and Steam fixed the problem.
It's been "fixed" more than once. And it looks like it may be in need of fixing again.

The Language section can provide an indication of whether a certain region is getting polled to a greater extent than usual, and sure enough, we see the percentage of Windows systems running "Simplified Chinese" has skyrocketed from the previous month, going from 23.58% of systems in July up to 31.04% in August. Those 7.46 percentage points means the relative share of systems running that language has supposedly increased by nearly a third from one month to the next.

Korean systems also skyrocketed according to the data, going from 1.75% in July up to 3.04% in August. Though interestingly, their share in July had been more than cut in half compared to June, so perhaps Valve performed a correction then, only to have their fix break again the following month. Both countries have a large number of net cafes, which suggests there may once again be issues related to those.

The big change in the demographic has undoubtedly skewed results across the survey. Just look at the Windows version data as a good example. Windows 7 64-bit inexplicably climbed from 5.68% up to 9.00% between the July and August surveys. And back in the May survey, the OS accounted for only 2.04% of systems, with Windows 10 at 96.56%. So, over the course of just three months, Windows 7 64-bit's usage share has supposedly increased to over four times what it had been.

While the Steam survey can have some usefulness in showing trends over longer periods of time, it's probably not a good way to analyze short-term changes. It really needs a major overhaul to be more useful, such as by allowing data to be filtered to specific regions and other criteria.

That being said, I do suspect Nividia's share of graphics card sales relative to AMD is currently higher than usual. Nvidia is mostly just making graphics hardware on the 10nm node, whereas AMD needs to divide its 7nm production between CPUs, GPUs and console APUs. They are contractually obligated to make tens of millions of large console chips for Microsoft and Sony, and when it comes to deciding whether to dedicate the rest of their capacity to CPUs or GPUs, the CPUs are undoubtedly far more profitable on a per-wafer basis. So, only a relatively small portion of their production has likely going toward graphics card chips.
 

waltc3

Reputable
Aug 4, 2019
420
223
5,060
The Steam survey is the farthest thing from a scientifically reliable statistic because it is exclusively opt-in, and Valve never lets on to how many people per month opt-in to the survey--it is not automatically tabulated by Valve--the company is well-mannered enough to make the hardware survey opt-in only, so that it will not be compiling stats from people's machines without their permission. Lots of people are very sensitive about that sort of thing today. JPR results, even, are not furnished by the companies, but are estimates they compile from speaking to various retailers.

My theory is that people make a lot of noise about month-to-month Steam survey percentages on slow news days, basically...;)
 
i'm just tired seeing the debate that has been going on for years about the validity of steam hardware survey. that's all.
The thing is we sort of have to put in the disclaimer every time. If Steam used valid statistical survey methods, and said as much, we could simply point to the stats behind how the survey gets conducted. Again, any basic statistics course should make people question the Steam Hardware Survey, simply because there are far too many unknowns. It doesn't make the survey useless, but it's very likely a biased survey that should be taken with a nice scoop of salt.