Steam provides data on the hardware their users run.
http://store.steampowered.com/hwsurvey/
The problem is that interpreting that list is a real pain in the butt since the gfx cards are itemized by exact model.
What are some of the good hardware breakpoints for lower or middle tier PC gamers?
Surprising that expecting DX12 alone means you’re dropping a quarter of the market directly.
This article has a good breakdown of common graphics cards and where they sit.
I also see a lot of charts with a graph showing their score with benchmark software, but they never really say much about what to actually expect from the cards.
I have a simple suggestion. Go to Steam and find some games which are similar to yours (same scope, same user demography, same price) and popular - check out their minimum requirements to get an idea of what would be right for your game.
There is one problem in general with setting requirements; if you are developing right now, what is important is not the average hardware right now, but the average hardware when your game is released. By then, “DX12 only” may cover a lot more hardware than it does now, the average VRAM may have increased etc.
This, very much! It doesn’t matter what the average or common user has, it matters what your users have.
That actually makes things a little trickier. Sure, it’s easy to figure out that your high-end FPS is going to skew towards people with GTX 970s and newer with pretty beefy CPUs, but when you start getting into niches, you may as well be writing your own hardware surveys to figure out what you need. Even checking out the Steam requirements for similar games can be a mixed bag. Like, when Supreme Commander came out, that thing required monster PCs, when the trend at the time was pretty low spec.
It makes things a lot trickier. Still, even if you can’t get better or more relevant information, it’s good to at least know the limits of the usefulness of the data you’ve got.
I’d estimate something like a 1060 or 470 is a pretty basic, pretty universal, new card you might aim for as a min spec. They’re both around 200 bucks if I recall (edit: barring the current cryptocurrency business), which means that in a year or two (which is Blood Sweat and Gold coming out?) gamers will probably be able to get them for $150 or maybe even $100.
I’d aim for 4GB, maybe 3 (nVidia went really weird with Pascal memory) as a “video memory” budget for an average user. BS&G didn’t seem like a graphically demanding game.
But one thing I hope you do is keep the top end pretty open. Not saying you need to include 4K textures, but what you CAN allow the user to amp up, please do. Unlocked framerate, unlocked resolution, all the AA a PC gamer could want, all that stuff.
Make the low end low, but keep the top open for people with more powerful cards.
Edit: And yeah, DX12 just hasn’t had a whole lot of adoption, either for gamers or developers. And from what I’ve heard most games don’t implement it very well so it’s not a very significant difference. Again, from what I’ve heard it requires building your game around it from the ground up.
I want to point out that the suggestions for what these cards are capable of are entirely dependent upon the game and the settings. A 3GB or 4GB card is absolutely not 1440p 60 FPS “ready,” unless we’re running at absolute minimum specs. It’s just not enough video memory for the screen size and asset data loaded in.
3GB or 4GB is 1080p territory. Which is fine, there’s nothing wrong with 1080p. But just keep in mind that the notes here aren’t exactly accurate.
Edit: Additionally, it says the RX 470 is the PS4 Pro equivalent, but the unmentioned reality is that console games can be optimized far more readily than PC games can, which means that you’re getting far more bang for your buck than a typical 470.
That’s a crazy minimum. More realistically people have a previous generation card because they’re still perfectly good for today’s games. A GTX 980 is still a bit of a beast.
Desktops are more likely to use GTX 400 too 600 series or Radeon up to the 7800 series than the latest GPUs. There might be many GTX 1000 series/R400 series users now, but not enough that you can rely on just them as your user base. Decent gaming latops might have 780m, 860m and 960m, and I’m not sure the super-cool GTX 1000 series-based ones are even available yet. You’ll have a very limited audience if you make a game with GTX 1060 as the minimum.
Go with Trexug’s suggestion. The bigger publishers have a fairly good idea what sort of specs their games run on, so if you’re making strategy games look at Paradox or Firaxis games, for shooters you can look at Valve or Activision games (assuming your aim is even near that level of graphics polish). Plenty of top-down (A)RPGs are even made in Unity, so pick any of them and simply copy their specs if that’s your thing.
Preferably you’ll have a number of different systems for testing among your in-house testers 
I’m probably being over-optimistic there. Maybe 460 and 1050. However, the fact that these are new doesn’t make them automatically better than a 980Ti or something (I was attempting to point out low-end new cards), and as already mentioned a minimum spec should be targeting future builds, not current ones.
Have to ask what “minimum” means, too. Minimum doesn’t mean “cards below this can’t run the game.” Minimum means some minimum standard you want, such as 1080p and 60 FPS on the minimum settings. (edit: and more devs should do what the Forza Motorsport team does, and actually specify what kind of performance “Minimum” gives)
Of course it’s possible (probably likely) I’m just used to being a graphics *****.
Wouldn’t 30fps be a good cut off point for minimum specs? It’s still playable, and a lot of console games run at 30fps anyway don’t they? I’d say 1080p or lower should be fine. Other than that, just smooth enough performance that it doesn’t affect gameplay. 60fps is my personal preference for a game, but that’d be better for “recommended” specs, I’d say.
I’d be careful with that. A lot of people just crank everything to max without understanding any of the implications, and then rant in their negative reviews about “terrible optimization”. Crysis 1 had the goal to have some headroom to keep hardware busy that wasn’t invented yet at time of release, and all that lead to was the “but can it run Crysis?”-meme in benchmarking. If you really want to provide such settings, I’d hide them away in manual .ini file editing, so that mainly hardcore enthusiasts who know what they are doing will mess with them. You can add comments there explaining that these are intended only for above high-end consumer grade systems.
I’m aiming for stable 30 fps on a GTX 660 and stable 60 fps on a GTX 1060 on max quality settings. Chosen mainly because those are the two cards that I own…
When they spell out minimum specs it’s often phrased as “minimum GTX 460 or equivalent”, implying something around that capability both in terms of speed and shader availability. Comparing generations is of course going to be a near-impossible job, with some generations translating evenly across three product lines to +20% or whatever, while others suddenly leap beyond at the low-end and barely match at the top. Then there are the Ti cards
But generally it’s all subjective from the people doing the spec comparison. They just have more experience at Blizzard, Valve or EA than most bedroom/garage/shipping container under a bridge programmers do, so if they say it’s playable on a potato you can expect less than 60fps, but still good frame rates at some average quality setting on said potato.
You’re of course correct that that can happen. However, if you signpost it very obviously, that decreases the chances of a user misunderstanding.
There I meant minimum in terms of performance. For instance, I definitely meet “recommended” specs on most Ubi games, but I am most definitely not getting 1440p 60FPS performance…with an RX 480…
These days it almost looks like gamers and critics want to misunderstand you. And even critics often don’t understand what optimization really means. They think it’s some magical fairy dust you can sprinkle over anything to make it go faster, and you just need to do that long and hard enough to make 60fps happen on any kind of hardware. And unless they happen to be devs themselves they just don’t understand things like e.g. procedural generation being incompatible with a good chunk of optimization techniques that are common in AAA games. All they know is “below 60fps on my computer = poorly optimized”.
If you still really really wanna offer settings to “futureproof” your game, I’d suggest to name your quality settings something like this:
highend PC from 2011 - low
highend PC from 2013 - medium
highend PC from 2015 - high
highend PC from 2017 - ultra
highend PC from 2023 (estimate)
highend PC from 2029 (estimate)
or “low, medium, high, ultra, do not use before invention of flying cars”.
Low, Medium, High, Ultra, Crysis. 
You really could do something like that. Like, make the ultra setting only use the most “reasonable” maximum settings, like say non-supersampling forms of AA, and have more powerful options only possible if you do a “custom” setting. And when the user selects the more intensive options, have a popup that points it out - “Hey, this setting is very graphically intensive. Be prepared for a major performance impact.”
It’s like Crysis. People KNEW that game was performance intensive, so it wasn’t really people complaining about optimization. And of course it helps that that game (or at least Crysis 2, which I’ve played) was very, very well optimized so it ran (and looked fabulous) on pretty weak hardware.
If your target audience is squarely made up of tech enthusiasts, sure. For most people, though, upgrading is something you do only when your current stuff isn’t good enough any more. If you want to maximise sales then your game shouldn’t be the one that shows people their stuff is no longer good enough, because then that upgrade they need is a barrier to entry for your title.
Good point. Though I have to wonder, in this specific circumstance, how “out of range” an RX 460 or 470 would be for the average user*. Again, I’m probably out of touch on this (because I’m typically pushing the envelope here - I’ve had a more powerful card than either of those for at least three years).
*a problem in and of itself. Are we counting laptop builds, which probably count for a significant portion of the survey?