Nvidia GeForce RTX 4090 Review: The Next Level In Graphics Performance

let's cut straight to the chase RTX 1490 is fast extremely fast to the point where it's out matching the best CPUs on the market making an actual balanced system with this GPU rather difficult it's best use case scenarios are for 4K High refresh rate screens or for extremely testing Ray tracing based titles that's in the here and now but just around the corner there's a new generation of games coming where we're going to need the kind of horsepower this product is offering that starts of course with epic's Unreal Engine 5 where we've only really had a taste of what this possible vision for the next generation is really about the Matrix awakens is based on a new rendering Paradigm it relies on Hardware accelerated raid facing it thrives on image reconstruction based techniques this is the engine of choice for developers working on the games of Tomorrow it seems once we've pushed past the current cross-genberlays however in the recent RTX 4 thousand launch beyond the pricing debate which is important of course we also saw a vision for the future of PC gaming cyberpunk 2077's RT overdrive upgrade is effectively the most challenging of high-end AAA Graphics showcases running effectively with full path tracing on all lighting elements old games RTX vmix will bring a range of older PC titles to The Cutting Edge via RT so that's the vision right consoles will always have their place but PC has the potential to offer a generational leap beyond the generation or leap we've yet to see but what we're going to need to get there are two things more Innovation via machine learning and wealth more horsepower a lot more horsepower in this review we will talk about that horsepower we'll be deploying our new test bench with an emphasis on Modern apis and forward-looking features and yes we'll be stacking up the art cx4090 against 3090 TI the most powerful of the last gen cards and we'll also be including the RTX 3090 for the genon Gen comparisons and the Radeon RX 6900 XT almost the best that team red has to offer in the here and now there is the slightly faster 6950 XT which the company did not widely sample to journalists we'll be looking at rasterization performance of course we'll be rolling out some of the most punishing raid-facing workloads and we'll be speaking a little bit about reconstruction yes dlss3 nvidia's latest AI upscaling technology a temporal upscaler if you like well if you've seen our dlss first look we can now dispense with the percentage multipliers and talk about actual frame mates first up let's focus on the RTX 4090 itself based on the new Ada Lovelace architecture it's using a new process node from Chip manufacturer tsmc has 24 gigs of gddr6xv van 16 384 Cuda cores a gigabyte per second of bandwidth over a 384-bit memory bus stated boost clock of 2.5 gigahertz or thereabouts so I think 2.7 is more in line with its usage as I see it and yes 450 Watts total Graphics board power perhaps not what you want to see as we are prepared headlong into an energy crisis but I would venture to suggest that doubling or even tripling of energy costs probably won't bother you too much if you're able to drop sixteen hundred dollars on a single PC component that allows to Circa 2 Grand in European currencies and around 1 650 pounds in the UK pricing and energy usage will be more critical on the 4 4080 and lower end parts but even so despite the high tgp some quick power consumption stats versus RTX 3090 offer up some interesting numbers I compared 3090 to 4090 across three titles using our Benchmark runs and here are the results cyberpunk 2077 at 4krt effectively forces the card into what you would think would be full power mode but 4090 power consumption averages at 406 Watts which at 46.82 FPS gives 8.67 joules per frame RTX 3090 at the same settings 13.84 joules per frame huge difference sticking with rt I went for Metro Exodus enhanced Edition next something a bit more sane but we're still seeing 3.63 joules per frame on 40 90 up against 5.3 on 1390. still though average power consumption of 436 Watts on the 4090 Forza Horizon 5 is fascinating though 3.74 joules per frame on the 3090 but only 1.83 on 40 90 and a remarkably low power consumption of 284 Watts so yes RTX 4090 is more power hungry than any Nvidia GPU I've ever seen before any GPU full stop actually but its efficiency is significantly better than last gen and while tgp is indeed 450 Watts it won't be consuming that flat out all the time you'll note on the 3090 stats there that the older card is running much closer to its 350 watt tgp in all benches okay with that diversion will be an important one out of the way back to the card I'm reviewing the founders Edition which appears to be one of the more restrained models compared to some of the gigantic aib models on Route look similar to the 3090 and 3090 TI not as long perhaps but it's certainly a fatto card based on the same design language in v-link allowing you to connect up a couple of RTX cards well that's now a thing of the past as is SLI in general really though some may say that dlss3 has some similarities with that old technology the pcie Gen 5 power input that made its debut on RTX 3090 TI is back which looks like a more discreet way to deliver power to the GPU but chances are you don't have one of those new power supplies so the somewhat ungainly adapter dongle requiring four yes four eight pin pcie power cables is included in the Box 3090 TI actually use three there beyond that aside from the sheer size of the card there's little more to say about the form factor ports layout Remains the Same meaning HDMI 2.1 and 3 display port 1.4 A's no DP 2.0 it seems which won't affect too many users but may have implications for future display Technologies and for high-end VR headsets let's dive into performance but just quickly I know it may seem fairly obvious to point this out but this card should not be paired with a 1080p screen unless you're into extreme super sampling anti-aliasing some weird and wacky things can happen for example here's remedies control maxed at 1080p resolution no RT here the game actually has a 240 frames per second frame rate limit which the 1490 hits when no other card does perhaps that sounds like a good thing but you can see the effect of the frame rate cap on frame times there it presents worse than the other cards in the lineup in many cases at 1080p lowest one percent and 0.01 scores will be lower than less capable gpus when you hit a CPU stutter the higher your GPU performance the harder it falls when it hits that CPU bottleneck you can check out all of the stats in the review at Eurogamer which has some nice interactive bar charts but one thing that did stick out is that when extremely CPU limited amd's driver often seems to be more efficient I guess you do see some decent scaling with raid facing at the high end at 1080p resolution but really I'd be considering using 4K Reconstruction from 1080p at that point so yeah basic 1080p full HD results there I still consider those largely academic 1440p again the scaling is not good even with a core I9 12900k paired with high-end 6000 Mega transfers per second ddr5 from g-skill your gains from the 4090 over the 3090 TI would not be that impressive for example take gears 5 here 30 90 to 40 90 at 4K we get a very reasonable 77.5 percent gen on gen gain not the best scaling but pretty decent I'd say however at 1440p that Advantage drops to a mere 30 percent I've not bothered to render it here but your gain at 1080p is a mere 8 ultimately then the more GPU power you have the higher end the workload you need to get your money's worth right this isn't anything new as such but arguably the performance differentials at lower resolutions have never been less impressive what this is really telling us is that the increase in graphics performance at the hardware level simply hasn't been matched by a similar increase in performance from the CPU so certainly not in gaming workloads anyway with that in mind then I don't consider 40 90 worthwhile for 1080P or 1440p gaming and even at 4K you will find many titles and workloads where you do hit CPU limits do maternal for example in this sequence no Ray tracing but there's just a 46 Boost from 30 90 TI to 4090 elsewhere in the game I'd expect it to be much higher so these are the reasons why the vast majority of our testing will be at 4K and why we've beefed up raid facing representation the toughest workloads are where this card is going to be truly put through its paces and let's start with rage racing then we had to work with Dying Light 2 because it's pushing Ray tracing hard really hard in the real-time info sequence here the 4090 is around 74 faster than the 3090 TI 94 faster than the RTX 3090 it has a 3 0.5 times performance multiplier over the 6900 XT interesting game really I have to wonder whether the RT mode as is on PC would even exist if the Investments Nvidia made in raid facing never came to pass you get RT shadows and ambient occlusion on consoles likely a better fit for the PC rdna2 cards but push on to ultra RT and it's all about RT Reflections and especially impressive RT Global illumination on top of that truly ambitious stuff and even without reconstruction the 4090 is delivering a 58 frames per second average in this sequence now let's switch things up by moving over to 4K performance modes of the various cards here using reconstruction dlss on Nvidia fsr2 on AMD we're internally rendering at 1080p on all cards then the Reconstruction algorithms kick in they have their own performance costs but the increase to Performance is transformative RTX 4090 increases its average to 115 frames per second up against 47 FPS on AMD at 70 and 77 FPS respectively 30 90 and 3090 TI are still in the mix Hitman 3 and other title where RT is pushed really hard across the bench RTX 4090 is twice as fast as RTX 3090 at 4K resolution comparing 3090 TI to 1490 it's about 74 percentage points to the better there's a 3.25 times performance multiplier comparing 40 90 to 6900 XT team red had different priorities with rdna 2 but it's the most punishing RT titles that emphasize the challenge facing radio on going forward a golf is still there if we shift to 4K performance mode reconstruction dlss versus fsr2 but the gaps close because I suspect at 1080p the CPU may have a bit more of a say in the results still an 83 percentage Point advantage to 1490 over 6900 XT dropping to 19 over the 30 90 and 15 over 30 90 TI there's a strong argument in many scenarios to actually move to a higher dlss quality level in 4090 in some titles even at 4K cyberpunk 2077 doesn't make life any easier but it's another example of showing what the RTX 4090 can really do not use the standard Benchmark here but rather a high speed run through night City a developer test tool of sorts still in the retail code at 4K we're seeing an 85 percentage Point increase moving from 3090 TI to 4090 that rises to a 2.13 times multiplier going from 30 90 to 40 90. the new am video card is nion three and a half times faster than amd's 6900 XT 4K May well be overwhelming but there are still big differentials as we scale down to lower resolutions at 1440p the 4090 is still 96 faster than 30 90 84 points to the better against 30 90 TI and it's three times faster than 6900 XT some titles do scale more gracefully to less capable kit however f122 with ultra RT across the board here 40 90 is 89 percentage points ahead of 3090 at 4K dropping to 77 at 1440p and just 40 at 1080p a 2.94 times performance multiplier between 40 90 and 6900 XT at 4K drops to a mere 83 percentage point lead at 1080p I mean the gap's still wide right but ultimately 1080p is important for RT as for many users even on higher end cards it's the native resolution before reconstruction kicks in for that 4K output just a shame this title doesn't support fsr2 we'll end with a couple of titles that run fast with rt where AMD users can get a really good experience and where either the 4090 will lose ground or it'll produce some stratospheric results we begin with Metro Exodus enhanced Edition the first AAA title built around an engine that demands gpus with Hardware accelerated Ray tracing functionality so console release too so tailored for good results on rda2 in the tiger stage here the 4090 is streets ahead of all competition though the Gap closes compared to some of our prior results at 4K here the 4090 commands are 58 percentage point lead over a 3090 TI rising to 77 over the non-ti 3090 it's a 2.87 times performance boost over RX 6900xt finally Marvel Spider-Man you think that swinging through Times Square would be the best Benchmark but weirdly I've found that a leisurely stroll through Feast aged you can absolutely Pummel performance in this one with full RT enabled at Native 4K there are some sweet results here 73 lead over 30 90 TI rising to an even more impressive 88 over 30 90. we're looking at a 2.44 times multiplier over 6900 XT here but I should stress that you can still get a great experience from this game on aod Hardware with rt as you would have seen from my prior Spider-Man PC performance video this is another game that supports fsr2 dlss and indeed xcss and I think that this is the model that all games supporting reconstruction should follow going forwards regardless of whether they're sponsored by an individual GPU vendor or not the gaps close up significantly now 4090s lead over 30 90 TI drops to just 13 at this point the game is running so fast it's hitting the CPU limit even in this demanding scene you're leaving GPU power on the table here with the 1490 so running at a higher dlss level or even a native resolution would be the way forward for 40 90 owners regardless 21 percentage points ahead of 30 90 and 51 versus 6900 XT at this point I think we're looking at one of the biggest gen on gen performance sleeps but there's also the matter of dlss3 to factor in nvidia's new frame generation technology that boosts performance still further well we should discuss whether performance is the right word an interesting point raised in the wake of our preview video DLS S3 works by rendering the current frame and the one ahead using machine learning and a revised Optical flow analyzer to interpolate a third frame that sits in between what this means is that at the fundamental level it's perfectly acceptable to say that your system is not gaining extra performance as such but with that said gaming performance is measured by frame rate or frame time and dlss3 certainly helps there so perhaps your games are not running faster but they are presenting more smoothly in a manner consistent with a performance boost at the cost of additional latency where Nvidia attempts to mitigate that via its reflex technology in the preview video frame rate numbers weren't shared but quotes unquote performance multipliers were now we can show all of the data these are preview builds though so do remember though Marvel's Spider-Man I think that's going to be as close to final as it's likely to get still let's dive in since we've just talked about how Spider-Man is indeed CPU limited on 40 90. let's look at that in closer Focus here's the Fisk construction Mission which takes place in the open world with full raid facing enabled everything maxed there and it is essentially CPU limited across the board there's barely any gain from dlss2 here versus native rendering that's how performant the 4090 is in this title even at 4K Max in fact owing to the nature of how unstable being CPU limited can be sometimes native versus dlss2 sees native rendering actually run faster for a short time simply down to run to Van variances the new GPU basically takes us close to the wire until we hit CPU limits so there's not much more to gain from using dlss2 DLS S3 essentially takes us up to a 235 frames per second average in this sequence to the point where we actually would need a 4K 240Hz screen to see all of the temporal resolution when CPU limited like this Fame generation can essentially double performance let's go to a heavier a more static scene in Times Square here home of many of the best Spider-Man benchmarks here we get a 14 to 15 frames per second bump from dlss2 over native as this is still relatively demanding graphically but we're still hitting the CPU limit there with dlss2 DLS S3 by interpolating between two frames extra smoothness is added to the scene and it's not reliant on the CPU at all for that game interesting stuff I did say in the first look video that you still need a performance CPU though for clean frame times if dlss3 amplifies poor frame times you end up with a stuttering mess and to illustrate here's some web swinging on the same top end settings with RTX 4090 but this time with a lowly ryzen 3 3100 CPU which is simply not equipped to run this game well at Max settings poor frame times without frame generation turn into a total mess when frame generation is enabled via dlss3 so there are ways and means to circumvent the CPU limit with dlss3 as we saw earlier but the bottom line is that good frame health is required on the input to get good frame Health on the output on the flip side garbage in garbage out Spider-Man results throw up another issue with dlss3 the fact that Nvidia recommends running with an unlocked frame rate vsync options are turned off in game to make that happen which essentially means that your game is always running flat out I have a 4K 120 hertz screen I don't need 200 frames per second frankly and frame rate capping doesn't work in the traditional sense here with a g-sync screen nvidia's setup would see screen tearing kick in above 120 FPS on my screen which is essentially all the time in this game again not good and not what g-sync is all about it's defined by a tear-free experience now Alex is covering this in more depth in his video but the bottom line is that control panel vsync does work but may have latency implications on some titles and may present canva stutter issues on others like flight simulator for example so this is still early days here for dlss3 more work needs to be done moving on though to GPU limited scenario cyberpunk 2077 the cherry blossom Market punishing on console and certainly super demanding on PC this is where we saw dlss3 beat out native rendering with a nion four times frame rate multiplier how that pans out is fascinating 4K RT rendering is hugely expensive so the first acceleration step is to use dlss2 performance mode which increases performance by 2.6 times an average 34 FPS jumps to Nikon 88. frame generation adds a further 54 percent of frame rate here taking you up to a 135 frames per second average in this sequence so in this scenario drastically reducing raid counts via dlss2 does much of the heavy lifting DLS S3 proving less additive but certainly very useful we're talking about psycho RT here running smoothly on a 4K 120 hertz screen after all the multipliers work in a similar manner in the extended driving section which across its length is less demanding than the marketplace DLS S3 actually offers a 166 frames per second average again remarkable stuff but here DLS S2 again shows how it's still crucial to the experience offering an initial 2.4 times performance multiplier versus native 4K with DLS S3 adding an extra 44 of frame rate on top of that we had some RTX 39 TI comparisons too in the first look video where the AI element tops out with dlss2 and where the native resolution rendering results will obviously be a good chunk lower in the Cherry Blossom Marketplace and average 54 FPS on 30 90 TI plays 88 FPS on 4090 both on dlss2 frame generation gives 4090 an extra push taking us to around a 2.5 times frame rate multiplier I also did the same test in the driving section where dlss-23090 TI delivers an average 69 FPS nice up against an average 111 FPS on 40 90 with frame generation taking you to 166 frames per second as I said in the first look it's the difference between a good g-sync experience on a 4K 60hz screen up against a much smoother one that looks great on a 4K 120 or even a 4K 144 Hertz display still though RTX 3090 TI still delivers good base numbers there I'd say frame generation is a really impressive bonus but anything from RTX 3080 upwards still a really good 4K GPU in my opinion you may remember from the first look that we also looked at an early build of portal RTX which takes the valve classic and uses RTX remix to deliver a full path traced experience the preview code we had did not properly support vsync off operating in something that looked like borderless Window mode running unlocked but only displaying non-torn frames not good for our capture tools with DLS S3 this game can actually run in excess of 120 frames per second but to get some frame rate metfix I used this static shot I set up a stress test area that ensures the game runs under 120 frames per second meaning I can capture all of the output frames path tracing at Native 4K is hard right which is why even the RTX 4090 is running under 20 frames per second here the load becomes much lighter using dlss performance mode which operates internally at 1080p reconstruct function here takes you up to high 50s and into the 60s DLS S3 frame generation does the rest taking us into the 90s there's a five times frame rate multiplier then from native to DLS S3 for the record in this level at least the native 4K version runs at around 24 FPS on average outside of this stress test area and dlss3 operates usually above 120 FPS so that's DLS S3 Alex has much more in His companion video to this one coming soon if it's not on the channel already it's another disruptive technology in the PC graphics space and a first generation effort much has been said about so-called fake frames and concerns about latency I actually think the bigger issue is the lack of an official way to cap frame rate when the Nvidia control panel vsync works well the job is done and the results are excellent but it's not always an option that's a problem for me in the here and now okay the last round of performance testing and its standard rasterization 4K is the focus here but 1440p and even 1080p numbers are in the accompanying Eurogamer text review personally wouldn't be playing control without the RT features but here we are with the initial bench run RT off 70 performance bump versus 3090 and a 52 percentage Point increase up against 3090 TI this test run has always been problematic for AMD and it is again RTX 4090 offers up literally twice the performance of the RX 6900 XT cyberpunk 2077 at Ultra 4K is bizarrely not that much faster than the RT version frame rates are very low here and the situation changes a lot at lower pixel counts performance uplifts are therefore disappointing a mere 43 percentage Point increase over 30 90 TI rising to 59 versus RTX 3090. once again though the 4090 is twice as fast as the 6900 XT things actually look somewhat more plausible at 1440p where the light of the land looks more in line with expectations 40 90s 60 faster than 3090 TI 77 points up on 30 90 while the performance lead over 6900 XT is reduced to just to 80 again similar to control RT is the way I'd be playing this one especially on this class of Hardware f122 sees 6900 XT move Beyond RTX 39t but just off the pace compared to 3090 TI however 4090 rules the roost as you would expect 49 faster than 3090 TI a 62 Point bump versus 3090 and a 61 increase over RX 6900 XT at this point you can see that even at 4K the overall increases in performance are lower compared to rate-based offerings but with F1 here inconsistent frame times on 40 90 suggest we're just hitting CPU limitations which will affect overall performance Forza Horizon 5 at 4K extreme is another title where AMD Ruled The Roost until the arrival of 4090 seven points ahead here of the 3090 TI at 4K and 10 ahead of the standard 3090 what this means is that the 4090 while comfortably ahead is 55 percentage points to the better 65 and 70 ahead versus 3090 TI and 3090 respectively and based on the power table at the beginning of this review it's doing so with a mass reduced powered volt on top of that again though we seem to be having inconsistent frame times here on the 4090s suggesting that we're on the cusp once again of hitting CPU limitations this would explain the reduced power draw and it may well be the case that if we could ramp settings up still further perhaps with higher levels of multi-sampling anti-aliasing the genome jet increases here could be even higher short look at gears 5 here which brings us into more established territory a 57 percentage point lead over 30 90 TI rising to 77 over its direct last-gen predecessor the RTX 1390 6900 XT is falling short again here though with the 1490 97 ahead there seems to be quite a wide range in results in rasterized performance then especially against 6900 XT so I'm really going to be interested to see how the rdna3 cards compare Hitman 3 next again no RT here these are rasterization tests so not the way I'd be playing it but still some strong results here for RTX 4090 80 ahead of 3090 TI 95 ahead of RTX 3090 across the whole bench 6900 XT inches ahead of the 3090 here meaning that the 1490s lead is in the 87 percentage point region a couple more faster results to come in before we wrap up and that starts with the Red Dead Redemption 2 uh The Benchmark one here sees a return to the less impressive performance differentials I guess 58 of the better versus RTX 1390 but only 45 faster than 3090 TI still 6900 XT is off Pace here so it's a 73 performance Advantage here for the new RTX 4090 and yes we end with shadow of the Tomb Raider our ever present benchmarking friend 67 ahead of 3090 TI across the three segments here 83 ahead of the 1390 with the 6900 XT falling short again against its contemporaries meaning a 107 lead for RTX 4090 so in wrapping everything up then there were big performance increases here but it's fair to say that the tougher the workload the better the bump but on a more General level my takeaway here is that for non-rt games in particular RTX 4090 essentially solves graphics for the generation of games that we have every 4K raster test that we carried out here by the weirdness of cyberpunk's ultra mode um it's giving us frame rates in excess of 100 frames per second many of them comfortably exceed 144 frames per second meaning that a 4K 144 Hertz screen will be well served by the 4090 and that's even without factoring in dlss with dlss this essentially makes the toughest of RT workloads good for a 4K High refresh rate screen too even games we consider to be exceptionally challenging like Dying Light 2's Ultra level RT in my time with the 4090 plan actual games as opposed to benchmarking them I think it's fair to say I don't really worry about settings anymore I don't need to think too much about the implications of weight facing either once dlss is in the mix and that's not even factoring in dlss3 all of which takes me back to the idea of a new generation of games that I started this review with I think the horsepower is there I think dlss3 has a key role to play once it's established itself across a range of titles I'm still curious to see how to run in say a fast-paced FPS title for example but dss3 is also a first gen effort I do want to see capped frame rates properly supported um but looking over the results here perhaps the biggest surprise was power efficiency yes it's notionally a 450 watt graphics card but it often runs lower than that occasionally a fair bit lower so I'm very curious to see how that translates to other less capable RTX 4000 cards running at lower power limits ultimately though RTX 4090 is a super expensive high-end card and it's not for everyone clearly in fact right now it's only really with a high refresh rate 4K screen that you really see the best of it by extension the Ampere cards well I'd say RTX 3000 still has a lot of life in it how the 4080 Stacks up against the 3080 could well throw up some interesting results based on what we've seen today but in the here and now yes there is a new GPU King and it's quite the thing so I think that covers everything I want to say about the RTX 4090 then so yes please do like subscribe share if you enjoyed the work and ring the bell for those notionally instant notifications another shout out of course for the DF supporter program lots of Early Access behind the scenes bonus materials uh a brilliant Discord community and some amazing retro stuff it's all happening there all in addition to high quality video downloads and yes I'll be adding all of the performance video data unabridged for supporters to grab but that's all for me for now thanks as always for watching and supporting digital Foundry

Channel: Digital Foundry

Views: 303,036

Keywords:

Id: glz5B-4IlKE

Length: 34min 8sec (2048 seconds)

Published: Tue Oct 11 2022