Nvidia GeForce RTX 5070 review: No, it’s not “4090 performance at $549” – Ars Technica
It’s not all bad news, but in many ways, this is barely an upgrade.
“4090 performance at $549.”That’s what Nvidia CEO Jensen Huang said of the GeForce RTX 5070 when he announced the card at CES in January. Thanks to AI, this new midrange GPU would be able to match the frame rates of what had been the fastest consumer GPU that had previously existed for around one-third the price.Let’s dispel that notion up front. No, the GeForce RTX 5070 is not as fast as an RTX 4090, not without some very creative comparing of non-comparable numbers. Per usual for the 50-series, Nvidia is leaning on its AI-generated interpolated frames for the bulk of its claimed performance improvements. In terms of actual rendering speed, the 5070 isn’t even as fast as a 4080 or a 4070 Ti. It’s barely faster than last year’s 4070 Super, and it has disproportionately higher power usage.For all that, it’s still not necessarily a bad card—it’s a little faster than the $599 4070 Super at the same $549 price that Nvidia used for the RTX 4070. It still represents Nvidia’s minimum viable 4K GPU, once you factor in DLSS upscaling, and it’s well suited for people with 1440p monitors.But it’s a hard card to get excited about, and unlike every other 50-series GPU released so far, it actually has to compete against something—AMD’s new Radeon RX 9070 series. I can’t talk in detail about how these cards all stack up until later this week, but suffice it to say that Nvidia could use something a bit more impressive than the RTX 5070 at this price.The 5070’s CUDA core count falls right in between the RTX 4070 and the 4070 Super’s, where the 5080 and 5070 Ti both included small increases from their previous-generation counterparts. To boost speeds over the 4070 Super, this means Nvidia is leaning on architectural improvements in Blackwell, the memory bandwidth increase from GDDR7 (a fairly significant 33 percent increase), and increased GPU clock speeds (the memory switch and clock speed boosts likely explain the higher power consumption).The 5070 gives you 12GB of memory, which still clears that crucial “more than 8GB” bar you’ll want for higher-than-1080p resolutions and ultra-quality textures, though as time goes on, it may be more limiting at 4K if that’s a resolution you’re hoping to hit with this card.The number that sticks out in a bad way is the power consumption—a maximum of 250 W under load, 30 W more than the old 4070 Super. The 5070 series is still far more power-efficient than older cards like the RTX 3070 or 3080, but efficiency is no longer as much of an advantage for the 5070 as it was for the 4070 series.We reviewed the Founders Edition of the card provided by Nvidia, and it uses a different two-slot design than the much larger, much more powerful RTX 5090 and 5080 cards. It’s the same physical size as the 4070 and 4060 Founders Edition cards, but with an updated power connector that’s set at an angle instead of sticking out directly from the top of the card. Nvidia has updated the cooler design, too, moving both fans to the same side of the card rather than placing one on each side.Whether it’s because of the updated cooler or the higher power consumption or both, the 5070 actually runs hotter than any other card in our current round of testing, including (by a few degrees) the 5090. The card runs a bit loud under load—that’s not a scientific measurement, but I noticed it more than I did when testing either the 5090 Founders Edition or the 4070-series Founders Editions. It’s something to keep an eye on for any smaller or lower-end 5070 cards from partners that use smaller double-fan cooler designs, though we doubt it will be an issue for the triple-fan monstrosities.Our testbed is the same as it’s been since we updated it for the RTX 5090 review—a Ryzen 7 9800X3D is the heart of the build to minimize the likelihood that CPU performance will cap any of our GPUs’ frame rates.We’ve tested the 5070 at 4K and 1440p. Cards in this $500 to $600 price range tend to be a better fit for a 1440p high-refresh-rate gaming monitor, especially if you’re hoping to play without DLSS or other upscaling technologies available.But if you play older or lighter games, if you don’t mind upscalers, and if you can turn down a few settings, hitting 60 frames per second at 4K is often within the realm of possibility. It’s certainly the least I would spend if I was hoping to play on a 4K monitor.The 5070 usually represents a mid-single-digit performance increase over the 4070 Super despite having fewer CUDA cores. It’s about 20 percent faster than the RTX 4070—a comparison Nvidia would probably prefer, since the 5070 is technically replacing the 4070 at $549 and not the 4070 Super at $599. But the launch MSRPs for the 50-series have been mostly imaginary so far due to a combination of supply, demand, and people buying cards just so they can resell them for more money. We’d assume the same will hold true for the RTX 5070 until proven otherwise.There were a couple of exceptions in our suite that performed strangely no matter how many times we re-tested them or how many times we tried clearing our shader cache or reinstalling drivers, and one game that benefitted from an outsized increase. Cyberpunk 2077 in Ultra mode with ray-tracing disabled showed a larger improvement than the other games we tested—24 percent faster than the 4070 Super, and 72 percent faster than the 4070. It was an outlier, but some games may benefit more than others.Horizon Zero Dawn Remastered and Cyberpunk 2077 in Overdrive mode with DLSS turned off both gave abnormally low or otherwise strange numbers—we’ve presented those here for the sake of transparency, and we’ve communicated our findings to Nvidia, but we’d take these with a grain of salt when comparing results. Horizon seems to have some kind of issue with DLSS frame generation regardless of GPU, with low frame rates and odd visual artifacts across all the GeForce cards we’ve tested with recent drivers.On the AMD side, we can’t compare the 5070 directly to the RX 9070 series yet, but it’s worth noting that the 5070 outperforms the old RX 7900 GRE by a small but consistent amount in our benchmarks without ray-tracing effects (the gap in ray-traced games is larger because of AMD’s typical performance deficit in these games).AMD says that the $549 RX 9070 and $599 9070 XT should be 21 percent and 42 percent faster (respectively) than the RX 7900 GRE on average. That could spell trouble for the 5070, but it remains to be seen whether AMD has fixed its ray-tracing performance problems and how much the cards’ power efficiency has improved.We ran our 1440p benchmarks on the GeForce RTX 3070, too, to give a bit more context for upgraders. The 5070 is usually between 40 and 70 percent faster than the 3070 here, a card that will find itself more and more limited by its 8GB of RAM as time goes on. The 5070 is still a good upgrade from the 3070; it’s just not all that much better of an upgrade than what you could get from a 4070 Super a year ago. As usual, the older your card is, the better the upgrade will feel, even if the generation-over-generation improvement is small.In exchange for five or six percent better performance, the 5070 uses 13.5 percent more power under load than the 4070 Super did. Nvidia was so far ahead of AMD and the RTX 30-series on efficiency with the RTX 40-series GPUs that one relatively inefficient upgrade cycle doesn’t really wreck things, but we do generally like to see power consumption either increase proportionally with performance or for performance to increase more than power consumption does.Improving efficiency was always going to be tough for the 50-series, since the Blackwell GPU dies are manufactured using a 4nm TSMC process similar to what Nvidia used for the Ada Lovelace GPU dies for the 40-series. But the 5070 ends up slightly worse on this metric than the other 50-series cards.When you turn frame generation on, you can see the source of Jensen Huang’s “4090 performance for $549” performance claim. The 5070 with DLSS Multi-Frame Generation (MFG) enabled in 4x mode has a similar average frame rate in Cyberpunk 2077 compared to the RTX 4090 with the old Frame Generation mode enabled (now labeled “2x” in games that support MFG). That’s because the 5070 is capable of generating three AI-interpolated frames for every rendered frame, rather than one.But the 5070 (and, when they arrive, the 5060 series) demonstrates the limits of FG and MFG more than the other 50-series cards do. With MFG enabled, Cyberpunk 2077 can hit over 80 frames per second on average with the Overdrive preset enabled. That sounds pretty good! Until you consider that the base frame rate (with DLSS upscaling enabled, but not any frame generation) is closer to 20 frames per second. With a base frame rate this low, user input will feel sluggish, and visual artifacts are plainly visible when objects are in motion.DLSS MFG can still have its uses—the 5070 can push some pretty high frame rates at 1440p, and MFG could help you take better advantage of a 240 Hz or 360 Hz 1440p monitor than you could without MFG enabled. But it remains a relatively niche technology that’s useful under certain circumstances, not the broad cure-all for mediocre performance that Nvidia’s charts imply it is.The GeForce RTX 4070 and 4070 Super were some of the best values in Nvidia’s lineup last generation, offering the same performance as high-end 30-series cards and outstanding power efficiency for a little less money than you’d spend on an RTX 3080 or 3090.Now that it’s here, it’s hard to classify the RTX 5070 as anything other than a disappointment. It’s just barely, by the skin of its figurative teeth, faster than the 4070 Super for the same $549 MSRP as the regular 4070. This technically makes it an improvement, along with support for multi-frame generation. But this nearly imperceptible performance improvement comes with a 13.5 percent increase in power consumption under load, which seems like a bad trade-off any way you cut it.The RTX 5070 feels like the kind of product you make when you’re not particularly worried about what your competition is doing. The entire 50-series has sort of felt like that so far, between the astronomically high price of the RTX 5090 and the so-so performance improvements for the 5080 and 5070 Ti. But it’s different for the RTX 5070 because AMD actually has an answer for it coming soon.AMD’s RX 7600, 7700, and 7800 GPUs may not have been smash-hit sales successes for its last generation, but stiffer competition at and under $600 did help pull Nvidia’s pricing down to earth a bit. Soon we’ll know whether the RX 9070 series will exert a similar pull—or if it can actually manage to wrest some market share back from Nvidia after years of trending in the wrong direction.Ars Technica has been separating the signal from
the noise for over 25 years. With our unique combination of
technical savvy and wide-ranging interest in the technological arts
and sciences, Ars is the trusted source in a sea of information. After
all, you don’t need to know everything, only what’s important.
