RTX 4090 review: Spend at least $1,599 for Nvidia’s biggest bargain in years
arstechnica.com
arstechnica.com
I have a big Steam library that I don't want to give up, but I could see being all in on future versions of Steam Deck.
Consoles have been the better value for a long time.
That said, it's a mistake to look at $1600 flagship cards and AAA titles and think that's the only way to game on a PC. Even entry-level cards in the $200-300 range can play modern games at 1080p or even 1440p as long as you're not trying to max out the settings (max PC settings aren't available on consoles anyway)
The real upshot to PC gaming, IMO, is that there are hundreds of highly-rated PC titles over released over the past decade that you can access with a few clicks.
> Nowadays, I don't want to fiddle with Windows (which insists on forcing updates when I do power my PC up to play with friends),
Ironically, this was my experience with consoles. Since I play infrequently, I'd often be stuck on forced console software updates whenever I turned it on. AAA console games also need to be updated frequently, some times with 10s of gigabytes to download before I can play. Updates aren't exclusive to PCs any more.
But console generations are long, and after a couple years, even low end PC components will catch up and exceed console performance.
The PC component industry honestly does a great job of selling people things they largely don't need. Games mostly have to work on console, so it doesn't generally make a huge difference to have a GPU for instance which is 5-10x as powerful as the consoles, since the games are optimized for the consoles anyway.
Yes you can go for high resolution or high frame-rates, but how much would you really miss it if you didn't have it?
...but I already needed a computer. My desktop PC is my workhorse and still performs better than any laptop at a similar price point. I just look at that $700 GPU as a $500 upgrade to a more basic option - similar to a PC without a nice GPU plus a dedicated game box.
Doesn't hurt that I can also make use of it in photo/video/animation/etc applications as well as in games.
Still, I imagine that if you don't do anything that can't be done on an $800-900 laptop, it might just be simpler to buy that and an Xbox.
A $700 GPU needs expensive fussy cooling too, and it's hard to get peak performance from it, and it really doesn't add much value to the gameplay.
That GPU is plenty fast for most gaming, unless you're trying to push 4K or very high framerates at 1440p.
Nobody needs a $700 GPU to play video games. Look at the averaged charts: https://www.tomshardware.com/reviews/best-gpus,4380.html An RX 6600 still puts out over 100fps at 1080p medium settings on their 8-game average.
I always found it funny that PC gamers seem to think console gamers don't own a computer. We do, we've simply found that for the price of building a gaming PC, one can buy a decent laptop AND a console, and run them both without needing further upgrades for a decade. This has a huge bang for buck. I should know, I've done this twice with the x360 and PS4, paired with a mid range dell laptop. Both are flawless for their respective niche.
I'm probably building a PC next year, but even I acknowledge that I'm going pretty far out into diminishing returns of enjoyment curve to do so.
But to be fair, like cars from the 80s to now, they do a lot more these days. I never installed a Physx processor for Crysis but now that's on the GPU.
$1000 on a new machine will get you running basically any new release at 1080p 60fps with only a little tweaking in the settings. I usually do 4k @ 60fps with medium/high graphics on my now nearly 4 year old rig, and it's FINE.
Don't buy these stupidly overpriced cards unless you have a really compelling use case (as in - you're using these commercially) or you're overflowing with cash.
And that will run games at 4K.
Coming of age during the golden age of GoldSrc modding may have something to do with this. From Minecraft mods to Rimworld, or Elder Scrolls, Factorio, Fallout. Even tweaking rules.ini in older Command & Conquer games. I love modding and don't want to give it up by going to the walled garden of a console.
Just because NVidia went bonkers with their flagship pricing during COVID isn't a reason to throw out the best gaming platform out with the bathwater.
> I could see being all in on future versions of Steam Deck.
Steam Deck is just a computer. You're still gaming on a PC.
But I mean if you only care about games which are available on console, if you prefer a gamepad, sure buy a console.
Just like a console is a computer, just with a different OS and form factor.
But basically at this point, any platform's hardware should be able to play any game in existence.
Anyway, so new laptop is HP "Victus" for $1100. It includes Nvidia RTX 3050 Ti, so now she can play Stray. It seems like a good deal: 512 GB SSD, Core i7, 16 GB RAM.
The laptop is all fan :-) She discovered that it quickly overheats if you put it on the bed (blocks the vents on the bottom). Also the battery drains real-fast when using the GPU. Also the power brick is larger.
Steam will still self update which is annoying but usually fairly quick. But otherwise, keeping linux up to date is pretty easy and something you can control. With Manjaro, I run a "yay" once a week or so and it takes 10 minutes or so maximum and everything is up to date.
Unfortunately consoles have become normal PCs in this regard - if you wanna play online you'll be forced to update the console OS and applyvgame patches.
The 4090 clocks in at an incredible 80tflops. Compared to the 10tflops of the PS5's GPU. Now compare that to a Nintendo Switch at about 1tflop if we are being generous.
Still I wish it wasn't so large or hot.
Is it 80x as fun to play games on the 4090 than the switch? Or 5.5x as fun, if we go by price?
It's pay-to-win though.
I know you're probably joking, but I see this sentiment spoken seriously over on PCMR, and it's so cringe. Graphics quality between PC and console are so close now that the only people who care spend all their time pixel peeping at 4x zoom.
and there's still a backlog of older games which you can max out for not a lot.
The only spec of this card that I appreciate, is the 24 GB memory, but that's for machine learning and, according to NVIDIA, that's not even for the target audience.
I'm sorry you lost your sense of excitement and wonder, but this 4090 really looks quite great.
Granted, ray tracing is optional (I can't even distinguish between RT and non-RT in games like CP:2077), and DLSS is a game-changer for 4k gaming. But there are plenty of games that I've been playing would be served better by a 4080 over my 3080.
It's great for a movie displaying on a wall, not for a game on a desktop monitor.
If your 4k pixel count is based on 3-5 screen driving/flight simulator on big screens, then sure that's great to have
Mostly because fake reflections in triangle-based rendering is convincing enough unless you're looking for it or have just gotten good at spotting it.
It's like bad kerning. If you don't know what to look for, you probably don't notice it or pay it any mind. But once you do, you see it EVERYWHERE. [0]
But I see artifacts of fake reflections in games everywhere. The problem is that most games rely on screen-space reflections which don't work if the reflected object isn't on the screen. So imagine being over water, looking over mountains in the distance, and you see the mountains reflected in the water. But as the camera pans down and the peaks go off screen, suddenly the mountain peaks get cut off in the reflection.
The only way to render a mirror accurately in triangle-based rendering is to effectively use a technique similar to Portal, but then your transform engine has to would twice as hard, since it has to project the scene twice. Even then, you're limited to flat mirrors without taking shortcuts.
Curved objects can't be reflective in triangle-based rendering at all. They fake it well enough by using environment mapping of textures, but then they won't accurately reflect things around them.
[0] Relevant XKCD: https://xkcd.com/1015/
I build a new machine about once every 4 years for gaming/development, and I just don't really see all that much difference between the capabilities of the machine I built two generations ago (now ~9 years old) and the one I specced out last year.
I can still load up modern AAA titles on the 9 year old machine and they run just fine at 1080p 60fps with some minor tweaks to the graphics. Turns out the GTX 970 from 2014 is still more than capable.
Basically - these top end cards are a complete waste of cash for the average consumer. Hard to even claim you're future-proofing the machine with it - you're just plain wasting money.
Before the 30 series, everyone knew RTX Titan was not a gaming card, but rather for compute/professional work.
Nvidia rebranded their Titan cards as 3090/4090 and people lap that up for some reason.
Two reasons people are gobbling these really expensive cards:
* Inventory constraints of the 3xxx series pushed people to buy any available card and it just so happens the 3090 was there waiting for them as the only available card (or within parts bundles).
* Streamers normalize overpriced setups. It used to be cool to be thrifty with pc builds, now its just all about "performance" as games become less and less optimized on launch.
It was literally double the price of a 3080, but was at most 20% faster, and usually only 10-15% faster.
nVidia's claim that "It's for 4K gaming!" was marketing bullshit. With a such a small performance premium over the 3080, I'd argue that if a 3080 can't handle a game in 4K, the 3090 probably can't either.
The "1%" who have unlimited budget for consumer purchases like this, are millions of people.
Namely, there are 2 big issues: - Virtualization: Getting vGPU functionality out of them (most of the time, there's no point in having a full 1080 Ti for only 1 job, we want to have it split on more than 1 VM). - Kubernetes workloads: I'm running a simple workload that uses the GPU (for testing for instance) and k8s wants it all for only one pod.
You'll need to make sure to not overload gpu mem, otherwise you'll get oom errors. You can use k8s patch functionality to add custom resources, essentially use that to represent gpu mem and include the amount of gpu the pod uses in its definition, that way you don't go over the limit.
Then don't buy it?
These are flagship cards. It's like the $100,000 Corvette in Chevy's lineup. It's fun to watch the reviews and see on the showroom floor, but most people aren't in that enthusiast market segment. They're going to the dealership for the $40,000 Camaro or the $30,000 Colorado truck because those get the job done for them at a reasonable price.
If you're an enthusiast and this is your thing, you save up the money and buy the flagship. It's a small audience, but the product is there for those who want it. If you're a casual gamer, you get something like a 3060 and it works great as long as you're not maxing out settings and trying to play everything at 4K.
I honestly don't understand all of the criticism for these flagship parts. When auto makers release a ridiculous flagship car, everyone cheers it on and understands the purpose. When hardware makers release a flagship part, it seems everyone wants to criticize it for existing.
I think you're right. I imagine a lot of people don't like the feeling that something faster exists outside of their budget.
Ironically, I suspect these people would be happier if the faster cards simply didn't exist and the GPU lineup topped out at something like a 3060. It's funny that people get angry that faster parts simply exist.
These new GPUs and even CPUs are entirely about power and performance. The upfront cost is one thing, but the long term cost of the both the energy consumption and central air system having to handle the massive thermal load is getting out of hand, particularly in a hotter region.
It's severely overlooked or just lightly complained about as "this thing needs cooling" without really elaborating what that entails or means. It's one thing to cool down a large studio area with a commercial HVAC system for doing youtube videos, but it's another to try and cool a bedroom in a house with the normal central air system.
I just wish these PC component manufactures would have some more focus on making their products more energy efficient, do more with less. You can always do more by consuming more, that's less impressive. And to go back to cars. Take a huge engine, and one half its size, both output the same power, but one uses a third as much of the same fuel. Which one is more impressive?
I understand flagship is the extreme edge. But when the power consumption of the flagships keep increasing, it normalizes the lower and mid ranges to also keep increasing. I just want to see a focus on efficiency, stop doing more with more. Do more with less instead.
Flagship cars generally do get significantly worse gas mileage than mainstream cars.
But you're exaggerating the inefficiency of a 4090. In terms of performance-per-watt, it's not that far off from their lower level cards.
If low power is a primary driver, it's very easy to move a slider and reduce the power usage for any of these cards.
> I just wish these PC component manufactures would have some more focus on making their products more energy efficient,
They do! Efficiency and performance are one in the same. They wouldn't be able to squeeze out this much performance if they didn't optimize the efficiency. More efficient parts result in higher performance for a given thermal envelope.
If you wanted the most efficiency, you would buy a 4090 and turn the power limit down to 50%. This is very easy to do. Reducing voltage and power limits is extremely common in small case builds.
Efficiency should be better than previous gens, at least this was the case for the CPUs I looked at (looking at their power efficiency ratings at full TDP they're hard to tell).
Marginal efficiency is lower when the power envelope is expanded to make these higher performance numbers achievable.
I'd wager that performance matched, these "power hungry" current gen GPUs/CPUs would show better efficiency than previous gens.
Besides all that, releasing "consumer" cards with such overly inefficient power consumption is just being tone-deaf at a time where half of the Western world is facing rapidly rising electricity bills.
https://tpucdn.com/review/nvidia-geforce-rtx-4090-founders-e...
https://tpucdn.com/review/nvidia-geforce-rtx-4090-founders-e...
in contrast to the breathless rumors, those power numbers look incredibly reasonable... it's less than the 3090 pulled last gen, and it's 1.7x the perf/w of a 3080 or 2x the efficiency of a 3090. 3090 was always known to be pushed a little far, but, there is an easy 70% uplift in efficiency here, and none of the power-creep stuff ended up actually happening.
4090 Ti will eventually happen, and will pull about the same as 3090 Ti. Which, if it gets 1.7-2x the perf/w, sounds fine.
Overall looking very promising for the next 5-10 years, and Im hoping any potential recession doesn’t cause Intel or AMD to take their foot off the pedal, or in the worst case (more likely for Intel), even scrap the product lines.
Count me in as bearish, at least for aftermarket discrete GPUs. Desktops' share continues to slowly deflate as a percentage of PC shipments. Fewer and fewer consumers upgrade their built-in GPU, whether discrete or integrated. The GPU "middle class" is disappearing between "whatever came with my computer" and behemoths like the RTX 4090.
Picture with an itx motherboard for launghs: https://www.reddit.com/r/sffpc/comments/xjl017/asus_really_p...
That would make sense. But AFAIK the standards could be amended if there was enough commercial justification. So I'm just wondering if we're headed towards a day when that happens.
This might be a great card for a CAD or video rendering workstation, but AAA game studios aren't going to be able to afford to spend hours lovingly drawing intricate details if only 0.1% of players can play their games in 4k because it costs $1600 for just the graphics card.
They do this anyway for future proofing. The tools they use capture far more detail than can be used on modern hardware. So they down-sample for target hardware and keep the hi-res assets around for the eventual next-gen "remaster".
Publishers have realized that AAA titles have long legs. GTAV and Skyrim have seen releases on several generations of hardware because people will keep buying them. So now the concept of a "remaster" release on future hardware is baked into development.
For games, unless I was running a top-of-the-line res/refresh rate monitor, I'd probably just get whatever the 4060-4080 ends up being and call it a day.
A game on max settings looks the same on different cards. The thing that really changes is the framerate.
Game studio's bring in more demanding settings when people have the power to run them. So what will happen is even more raytracing and post processing effects that you can turn on or off.
Making a usable asset from that is almost always a task of reduction not iterative addition.
I've always thought that they start off making the best graphical version of the game and then just start cutting off bits and bobs to make the lower tier versions.
Game studios find these cards useful because their own processes get faster not because they expect people to run games on a 4090. The 3090 in my case has sped up my workflows quite significantly. I can render, play, model and bake textures at the same time.
So if anything this will improve the state of things.
That's not really how it works. The faster cards allow people to play at higher resolutions and higher framerates with full resolution textures, wider field of views, more detail rendered off in the distance, higher resolution shadows, and so on.
It’s just that their product line has a more or less linear CUDA core to price best line of fit.[1]
There’s a dip specifically at the RTX 3080 Ti, where you’re getting more cores for the money than elsewhere on the best line of fit.
Not sure what trade offs there are there except less VRAM and in one instance I read that cooling wasn’t as good.
[1]: https://docs.google.com/spreadsheets/d/1MHTVDAyU7i8oXYAlmGPE...
There's also been quite a bit of inflation since the days of the 1660 Super...
So 10x 4090 cuda cores are as powerful as 10x 2060 cores?
A graph that plots price per core and power usage per core would be awesome.
Google Sheets is so amazing, it just works and I can watch how somebody edits the document and collaboration is dimensions ahead of every service I know.
Why is there no 3090ti in there, which should be the top of the line card from the previous "generation"? Current generations top should be compared with previous generations top rather than third-to-top, otherwise it seems a bit disingenuous.
Are there any better comparisons out there?
https://www.youtube.com/watch?v=j9vC9NBL8zo and https://www.youtube.com/watch?v=aQklDR8nv8U both include the RTX 3090 Ti and RX 6950 XT.
See 13 game average at https://youtu.be/aQklDR8nv8U?t=888
https://www.techspot.com/review/2544-nvidia-geforce-rtx-4090...
13-game average (4K): https://static.techspot.com/articles-info/2544/bench/Average...
> The GeForce RTX 4090 is on average 59% faster than the RTX 3090 Ti and 71% faster than the 6950 XT at 4K gaming.
Still though, as indifferent as I am towards Nvidia I think this is a good deal as well. There's plenty to balk at with that price-point, but there's also lots of (good) cards on the market right now. Nvidia deciding to release the 40-series in the current market seems almost downright experimental.
Unless there's a need to buy immediately or to stick with nVidia, it might make sense to see the RDNA 3 cards' pricing, performance, and power requirements.
AMD is moving to chiplets and more advanced nodes at TSMC, and they've been executing well on CPU and GPU improvements. Who knows what advances they make, or how much more hardware they throw at the problem. They are likely to improve upon their own performance by at least 33%, but perhaps they'll also approach Nvidia's 60%.
Certainly if they only reach 33%, pricing will be key - they will not be competing against the RTX 4090, but only against lower tier cards if that's the case.
I think it's likely that some RDNA3 cards will be on that Pareto surface.
NVIDIA also announced toolkits that allow moders to apply DLSS to older games that aren't receiving updates anymore.
But you still can build the entire PC for $1600 and play most games well.
The Radeon RX 6650 XT is dipping below $300 and will run 1080p with great image quality settings.
Or spend $425 on a Radeon RX 6750 XT and play most games at 1440p with excellent image quality.
Maybe #novidia or #novidya as a hashtag.
> When a standalone GPU is as large as a modern video gaming console—it's nearly identical in total volume to the Xbox Series S and more than double the size of a Nintendo Switch—it's hard not to laugh incredulously at the thing.
Am I correct in understanding that DLSS fundamentally modifies the original rendered picture/game ? Do designers/artists compensate for that ? Do they still have a "final product" in mind ?
DLSS 2.x is a fundamentally different approach, it is standard TAA where the sample weights are set by a ML model based on previous frames and motion vectors.
DLSS3 takes this further and actually does interpolation (or, extrapolation?) of additional frame(s) based on the previous "concrete" frame and the motion vectors for objects inside the game.
I doubt artists thought about this with the current generation of games. The tech is so new that it was shoe horned into a lot of games, and it has evolved a lot in just a few hardware generations. Going forward, I'm certain artists will give it more consideration, since it is clear now that this is the future of gaming.
Artist always compensated to make games look and run better, even in NES era.
I wonder what allowance 3rd parties have to detune these cards, and perhaps produce a 4090 with a more modest cooling solution suited to a ~300W power level, maybe making a 2 or 2.5 slot size card, rather than the current monster cards we're seeing.
The Steam survey shows the GTX1060 is still the most popular card so even the most recent games need to perform adequately with that in mind.
https://store.steampowered.com/hwsurvey/Steam-Hardware-Softw...
I still use a GTX1070 for playing at 1080p or 4K at 60fps for older or non-demanding games and it's still absolutely fine.
On the other hand, if you're doing other things with your GPU than just gaming, this might be a very well-worth card, depending on how much AI/ML/video encoding/3d rendering/physics simulations/X you do.
Personally I'm waiting for some benchmarks related to CUDA and particularly Stable Diffusion to come out before I decide.
And I don't game that much anymore, maybe some hours per month at maximum. But I'm using my GPU a lot for video encoding and 3D rendering, so still interested in GPU development.
If you observe NVIDIA movements and actions, you'll see that they are starting to focus a lot more on B2B than they ever did before. Almost like their B2C is not the main focus anymore, just a good PR piece for them.
I'd be very surprised if gaming didn't represented the majority of Nvidia's revenue of selling GPUs.
[1] https://webtribunal.net/blog/gpu-market-share/
[2] https://store.steampowered.com/hwsurvey/Steam-Hardware-Softw...
Going by public information(https://nvidianews.nvidia.com/news/nvidia-announces-financia...), "graphics" is still the majority of their income at about 58% (other part being "Compute & Networking" at 42%). However, "graphics" also include "the Quadro/NVIDIA RTX GPUs for enterprise design, GRID software for cloud-based visual and virtual computing, and automotive platforms for infotainment systems.", none of which are gaming use cases. Question is if those represent more than 10% of the "graphics" revenue or not.
My guess is yes. And even if 1% of consumer graphic cards are being sold for the non-gaming use cases, it seems pretty clear that the demand in the non-gaming use cases seems even bigger than gaming ones.
Also note that from the press release linked above, there are only a few lines dedicated to gaming. The rest is all about non-gaming use cases.
But seriously, unless you need PCIe Gen 5 or have a 240hz monitor, it's not that huge of a deal. Yes, it's a $1600 card, but what are you going to do? Buy AMD (which presumably won't have quite as good performance), or wait until next year?
Same for DisplayPort 2.0. These benchmarks show games exceeding 120fps at 4K already. Displays that support those higher rates and resolutions will be common in 5-6 years.
It's more of a question of "why not?" rather than "does it need it" at this price point.
Even if you screw very hard the wall part, the inner zone is still floating on its own weight
(No affiliation, I just clicked on the first search result)
There are GPU support brackets for these behemoths.
Also known as ANTI-SAG brackets.
http://dunfield.classiccmp.org/s100/h/hzinside.jpg
We could bring that back.
Now there are two 4080 cards, with wildly different specifications. Basically, one of them is actually a 4080 and the other is what they used to call a 4070.
This is reflected in their name and price here[1].
[1]: https://www.nvidia.com/en-us/geforce/graphics-cards/40-serie...
This is going to be awful when buying pre-built machines. They'll just say it has a 4080 and stick the crappy one in.
Everyone who wasn't in the tank was telling you god's own truth: NVIDIA is going to leap two nodes, it was always gonna be a huge leap, bigger than Pascal. They were drowned out by the AMD hypetrain crowd as usual who desperately wanted a flop and were trying to will it into existence.
It does go to show just how far NVIDIA is ahead on architecture... they could fight RDNA2 to a standstill with basically a whole node of disadvantage, and they did the same thing with Turing vs RDNA1 where they were using TSMC 12FFN (which is literally 16FF with the "your mom-sized" reticle - there is no optical shrink in 12FFN) and managed to fight 7nm cards to a draw on both performance and efficiency.