News Steam survey suggests more people bought the RTX 4090 than the Steam Deck — along with millions of other RTX 40-series GPUs

[Admin]

The Steam Deck was projected to hit three million sales by the end of 2023, and based on the Steam Hardware Survey, the RTX 4090 somehow sold even more units — as did most other RTX 40-series desktop GPUs.

Steam survey suggests more people bought the RTX 4090 than the Steam Deck — along with millions of other RTX 40-series GPUs : Read more

 

[thestryker]

The topic of Steam hardware survey efficacy came up on one of the HUB videos recently and I thought they had a really good take. They basically viewed the data as mostly worthless due to the random nature of the survey. They also suggested a much better method would be for Valve to make it an opt-in program when you first load up Steam and register a system.

Neither one of them could rememeber the last time they'd gotten a survey pop-up and that got me thinking I couldn't recall either. I did get one on the first of this year though which had gotten me thinking about that video again.

 

[emike09]

Crazy how much the 4090 is in demand, especially for its price. I bought one early on. People want powerful cards and are willing to pay for it. It's a double-edged sword. For one, it tells Nvidia that people are willing to spend $1500-$2000+ for a card, which sucks. But that also means devs can eventually focus on building games that cater to and utilize those high-end cards, bringing us better visuals all around. Or it means devs can get lazy with optimizing knowing people have good cards.

 

[JarredWaltonGPU]

The topic of Steam hardware survey efficacy came up on one of the HUB videos recently and I thought they had a really good take. They basically viewed the data as mostly worthless due to the random nature of the survey. They also suggested a much better method would be for Valve to make it an opt-in program when you first load up Steam and register a system.

Neither one of them could remember the last time they'd gotten a survey pop-up and that got me thinking I couldn't recall either. I did get one on the first of this year though which had gotten me thinking about that video again.

If they said it was worthless because of the random nature, they don't understand statistics. My problem with it is that I worry it's not completely random. Because if the sampling is triggered by anything other than pure randomness, it's inherently skewed and untrustworthy.

And I often go in spurts on being sampled, which is also worrisome. I'll go for maybe five or six months without getting pinged, and then suddenly I'll get sampled on five PCs all in the same week. It's almost like they're basing the sampling on user ID rather than randomness, which again would be bad statistics.

What I do know is that the data tends to look internally consistent. That means it's not useless; we just can't extrapolate exactly how many people own various GPUs. Minor fluctuations on a monthly basis are expected if you're doing random sampling, because the confidence interval will always leave a margin of error. Randomly sample 1% of the population ten times and you'll get ten slightly different results. But the month-to-month fluctuations from the Steam HW Survey don't even seem to correlate with that sort of variability, so again: worrisome.

The thing that strikes me as odd is why Valve even bothers to make it opt-in sampling. Given how much data Valve already has access to, it seems to me they could just have all the data, all the time, and there's not much anyone could do about it. I would wager other launchers (Epic, Ubi, EA, etc.) that don't provide any public statistics may already be gathering all the data from all the users. Valve is just being slightly kind by providing some guidance to the marketplace.

What I really wish Valve would do is to provide more insight into the sampling, and also collect CPU strings and show that data as well! I'd really love to know how many people (percentage wise) have an i9-13900K, and how many are still holding on to an old Bulldozer CPU, and stuff like that. Just grouping by clocks and core counts means we know next to nothing about the various CPUs.

 

[thestryker]

If they said it was worthless because of the random nature, they don't understand statistics.

They were referring to the huge swings that randomly occur due to the poor sampling. It swings way outside what truly randomized sampling should do.

But the month-to-month fluctuations from the Steam HW Survey don't even seem to correlate with that sort of variability, so again: worrisome.

This is exactly my problem with it and I don't understand why Valve has never done anything to address it. As you say they've got all of the data anyways.

 

[JarredWaltonGPU]

They were referring to the huge swings that randomly occur due to the poor sampling. It swings way outside what truly randomized sampling should do.

This is exactly my problem with it and I don't understand why Valve has never done anything to address it. As you say they've got all of the data anyways.

That's the thing about random sampling. You set your confidence level and then run enough samples so that you can say, "We're 99% confident that the real value (i.e. if you sampled the entire population rather than only part of it) would have a value within a specific range." Even then, you'd have a 1% chance of falling outside of that range.

But 99% confidence levels are actually very uncommon. Most statistical surveys settle for a 95% confidence level, which would mean over the course of 20 months of sampling, you'd expect at least one instance of the sampling resulting in numbers that are outside of the expected range.

It's been a bit, but if you were doing a survey with a population size of 150 million as an example (ie, Steam users), and you sampled only 0.1% of the population (150,000 samples), you'd get a 99% confidence level that the real value would be within 0.33% of the surveyed value. Or alternatively, it would give a 95% confidence level that the real value was within 0.25% of the surveyed value. In other words, even with LOTS of samples from Valve, monthly swings of up to 0.5%~0.66% would be easily justifiable.

Instead, what we often get is very regular results that all seem to be similar, and then suddenly (like in August 2023) the calculations or sampling are changed and things swing by 5% or more in some cases. And then the swing is detected and Valve fixes whatever changed and slowly brings things back into the expected range.

 

[ekio]

AMD GPU percentage is near nothing because instead of selling their card for half the price of Ngreedia, they raised their prices to set a similar perf/dollar ratio.
Unsurprisingly, almost nobody is buying them since it costs the same with a much weaker brand and a much weaker software ecosystem.

AMD should really start selling their GPU at a MUCH lower price if they hope people buy them someday… a few years ago, cards where a third of the current prices so stop taking us for fools, reduce your crazy margins and stop being a joke.

 

[icraft]

Could also be that those that can afford a 4090 can also afford a Steam Deck to mess around with, and so can contribute both to the survey without increasing unique users.

 

[umeng2002_2]

Devs aren't catering to $2000 cards. They are catering to $200 cards. They then spam insane shadow resolutions and inefficient texture use or put in path tracing (if nVidia pays them) for the high-end cards.

 

[P1nky]

AMD GPU percentage is near nothing because instead of selling their card for half the price of Ngreedia, they raised their prices to set a similar perf/dollar ratio.

AMD should really start selling their GPU at a MUCH lower price if they hope people buy them someday… a few years ago, cards where a third of the current prices so stop taking us for fools, reduce your crazy margins and stop being a joke.

AMD used your strategy in the past for multiple generations and it didn't work. The GPU division was LOSING a lot of money because of it. The only thing it would accomplish it would be that NVIDIA would lower the prices and that would bring AMD in the read again.

 

[Phaaze88]

That trailblazing lead in machine learning has drawn in a crap ton of folks, huh. What's going on with Quadro?

But that also means devs can eventually focus on building games that cater to and utilize those high-end cards, bringing us better visuals all around. Or it means devs can get lazy with optimizing knowing people have good cards.

-designs around cash shops. Bugs and glitches will be littered throughout a game, that takes who knows how long to fix - if ever - but you can bet the shop works without a hitch. Being able to patch games later has been a boon, and it has also been abused.

-experienced devs are jumping or moving ship - some being purposefully let go by the studio's management or the publisher, in favor for cheaper, yet less experienced workers. So it's not entirely laziness, but some don't know what they heck they're doing, and the guide that could've helped them is no longer there.

-risk aversion has lead to some stagnation. I.e., certain yearly releases, remasters, and numerous sequels. One may even be able to feel that a title lacks passion/soul.

-games have/are becoming more business driven, and less art driven. By art, I don't just mean graphics.

-the mobile games market is much larger and more profitable. Doesn't require anywhere near the money or resources of a 4090 to play, plus, they're easier and cheaper to produce. That money being made and saved is not going unnoticed by the business-driven companies, and we're back to risk aversion, with less incentive to try anything new.

 

[hannibal]

I have partisipated the survey many times… and statistic works like this. You take random sample and you get result that is accurate enough!
But there is about 2% error margin… so the amounth of 4090 is smaller than that error margin! The survey works in things that are larger. It tells nicely those results that are level of 10% and higher. Anything that is smaller than the error margin are less likely to be usefull!

 

[JarredWaltonGPU]

I have partisipated the survey many times… and statistic works like this. You take random sample and you get result that is accurate enough!
But there is about 2% error margin… so the amounth of 4090 is smaller than that error margin! The survey works in things that are larger. It tells nicely those results that are level of 10% and higher. Anything that is smaller than the error margin are less likely to be usefull!

It's a 2% margin of error on the sample result, not on overall percentage. So if the survey says there are 0.78% of people with a Steam Deck, and the margin of error is 2%, that would mean the range would be 0.7644% to 0.7956% — not -1.22% to 2.78%. That's assuming correct statistical sampling in the first place, of course, which is still unknown WRT the Steam HW Survey.

 

[cirdecus]

I'm curious how the data measures multiple systems. For instance, many 4090 owners custom build their PCs and may swap some RAM or a CPU and reload windows often. Each time Steam is installed, they register their machine. For me personally, I've swapped my CPU and Motherboard twice the past year, all the while using the same 4090. I've reloaded windows at least 3 times. Does this show that i have three different PC's with a 4090? I don't know.

 

[umeng2002_2]

Go look at Cyberpunk or any game with a lot of ray-tracing. The 7900XTX is uncompetitive. A lack of good frame generation just hurts it further as path tracing needs frame generation.

 

[Avro Arrow]

I think that it's finally time for everyone to agree that the Steam Survey is for the birds. 😆