NVIDIA GeForce RTX 5070 review — Midrange power, priced to move

If you hit the Steam hardware survey and look at where the vast majority of PC gamers live in terms of GPU, it’s where we are for this review today – the midtier, in both price and performance. As you’ll see The NVIDIA GeForce RTX 5070 is slotting right in between the RTX 4070 Super and RTX 4070, right where you’d expect it to be, but during their presentation NVIDIA made the claim that this card could come close to the results from an RTX 4090, using DLSS 4.  At just $549 MSRP, that’s a very bold claim, but one we can easily test, so let’s get to it.

As always, our first stop on the tour is to head under the hood to see where the 5070 lands in terms of cores, speeds, and feeds. Before we do that, let’s just preface this very, very clearly – you cannot directly compare across generations with just the amount of cores or even speeds.  The generational differences between cards (e.g. third-gen RT Cores and fourth-gen Tensor Cores in the 4000 series, vs. fourth-gen RT cores and fifth-gen Tensor Cores in the 5000 series) is always less about the amount of cores, but more about what’s being done with them.  Let’s take a look: 

As you might note, this card is utilizing a different GPU than its larger brother – a GB205 versus the GB203. We do expect the GB206 and GB207 variants as we get into the cards further down the chain, but that’s for another day.  That chip is part of the Blackwell 2.0 family, and as you’d expect shares the same supported advancements like DLSS 4, DirectX 12 Ultimate, and more.  The GPU operates at a frequency of 2325 MHz, boosting up to 2512 MHz, with the 16GB of GDDR7 (the same amount as the 5070 Ti), but running at 1750 MHz for a 28 Gbps effective rate.

As you can see, this is a far smaller card than the RTX 5080.  Dual slot, the card is small enough to likely be used in a Micro-ATX case – check your specs.  Connecting it, as is the case with the rest of the line, can be accomplished with the included adapter, or a single 1x16-pin power connector, with a power draw rating of 250 W maximum.  Just like the RTX 5090, the RTX 5070 has 1x HDMI 2.1b, 3x DisplayPort 2.1b, operating in a PCI-E 5.0 x16 slot.  

As I mentioned, the NVIDIA team places the RTX 5070 in the same power band as the previous generation flagship, the RTX 4090.  That’s completely impossible using simple raw power as the 4090 has significantly more processing power, power band available, and pure speed.  What the RTX 5000 series cards has over its previous generation, however, is DLSS 4.  Let’s dive into the differences between DLSS and Frame Generation, because yes – those are two different things, folks. 

Before we get into the numbers, I want to talk about DLSS 4 and “Fake frames”.  If you are unfamiliar, GeForce cards since the 2000 series have a technology called DLSS – deep learning super sampling.  Newer cards also picked up additional technologies like Deep Learning Anti-Aliasing, Ray Reconstruction, and Super Resolution.  We’ve deep dived all of them in the past, but let’s focus on DLSS first.

You might see annotations of DLSS  as 2X, 3X, and 4X, but more often you’ll see it designated as DLSS 2.0, DLSS 3, and DLSS 4 (and yes, they dropped the dot after 2).  DLSS 2.0 introduced TAAU or Temporal Anti-Aliasing Upsampling (or more often just listed as TAA), as well as correcting some of the image distortion we saw in the original implementation of DLSS.  

DLSS 3 was a big step up as it combined DLSS Super Resolution (resolution upsampling), DLSS Frame Generation, and NVIDIA Reflex. Super Resolution uses AI to upscale a lower resolution image into a higher resolution and, for the first time, the image that came out of the AI was frequently better than the original image!  Frame Generation was an option to use AI to look at the previous frame and then attempt to create a new frame to be inserted into the display path courtesy of NVIDIA’s new Tensor cores.  By not having to fully process and render every frame, the resulting framerate could be upwards of twice that of using TAA or native resolution.  There was a small bit of latency introduced by this process, and Reflex used a similar AI prediction model to anticipate where the player’s cursor would be.  

DLSS Ray Reconstruction, Super Resolution, and DLAA now use a new transformer AI model trained by NVIDIA’s supercomputers to better render the most complex parts of a scene. The result is improved stability from one frame to the next, enhanced lighting detail, and more detail in motion. If you’re wondering what level you should use, I’ve revised my recommendations recently.  Before I was recommending the Quality setting as the sharpness and cleanliness of the resultant image has been better for prior generations.  Well, that seems to have changed.  Using side-by-side comparisons and seeing them in motion, I’m now revising that to Performance instead.  This comes from moving from a CNN-based model (convolutional neural network) to a Transformer-based neural network.  The team at Medium explained these models at a level well beyond what I could, so I’ll direct you to them if you want to learn more. In the end, running at 4K resolution, you’ll appreciate the extra frames, but the improvements with DLSS 4’s combination of technologies are nearly indistinguishable from native, especially in motion.  Best of all, this new model is also coming to RTX 40, 30, and 20 series cards.

If you’d like to look for any visual impacts of DLSS 4, I put together an extensive amount of capture when I reviewed the RTX 5080 on every possible DLSS settings, as well as rasterized performance.  While these were pulled from a stronger card than the one we’re reviewing, the point remains the same – the improvements to AI generated images is substantial.  Much of the blurry look that permeated the first iteration of DLSS is gone, as is much of the misinterpreted “mouse tails” issues we saw in DLSS 2.  DLSS 3 was a big improvement over 2, fixing repeated or jagged textures, and we see a similar uplift with DLSS 4.  Now that I’ve had a few weeks to spend with it, it’s readily apparent that the new transform mode is a marked improvement.  Judge for yourself:

The next bit of tech that lets the RTX 5070 punch above its weight is Frame Generation.  In DLSS 3, cards were able to utilize AI to look at a frame, pixel by pixel, and anticipate where movement will occur.  That’s where the game needs to do a fresh calculation, so that’s where the GPU will spend its precious cycles.  The AI chip, on the other hand, will generate 7/8ths of that frame, essentially introducing a frame that didn’t need to be fully calculated by the GPU.  This saves a lot of cycles, resulting in a faster perceived framerate.  

This brings me to an axe I have to grind about “fake frames”.  If you’re worried about this sort of thing, I’d ask which you’re really more concerned with – visually smooth results or numbers on a graph?  If you steadfastly cross your arms and won’t use DLSS, hey – that’s your choice.  We include rasterized results in our graphs for you.  I can tell you this – you’d be enjoying your experience a lot more if it was cleaner, faster, and with a more consistent framerate above 100 instead of hovering around 30.  If you tried it in the past and weren’t happy with the latency, you’ll appreciate that Reflex has gotten an update to solve that as well, solving that issue with a similar predictive AI method. 

DLSS 4 takes the 7/8ths frame tech from DLSS 3 and kicks it into high gear. Using the new Blackwell architecture, it uses a similar technique as before to interpolate up to three additional AI-generated frames into the display stream, driving framerates into the stratosphere.  Games like Cyberpunk 2077 are notoriously brutal on videocards, with the flagship of the 4000 or 5000 series barely able to pop above 30 fps when running without synthetic frames.  As you’ll see when you read our benchmarks, DLSS 4 makes a world of difference.  

At launch there were 75 games that supported DLSS 4 directly, and that number has grown to add the likes of Kingdom Come Deliverance II, Dune: Awakening, and DOOM: The Dark Ages just around the corner. You can also use the NVIDIA App to attempt to override games that don’t directly support it to see if it’ll work. Avowed, for example is one of these, benefitting greatly from DLSS 4 with an appreciable uplift and a smoother overall experience.

Outside of gaming, one of the nice surprises for the GeForce RTX 5000 series is a new rendering format.  AV1 has enabled blisteringly fast render speeds in DaVinci Resolve, and now we have AV1 Ultra High Quality.  This new format cleans up visual noise on large open spaces like skies and walls.  It also adds another notch on the speed belt – something I’ve seen in action as I render the videos for this review!  I had tested a video I was rendering on my RTX 4090 and saw videos that rendered in 90 seconds take less than 30 on the 5070 Ti, and we saw similar results on the 5070 at just 34 seconds for the same video.   

With all the tech explained in likely more detail than you need, let’s get into where the RTX 5070 ends up in benchmarks. 

As you can see, we are still testing at 4K for these titles, and somehow this card is still delivering, sometimes in triple digits, at least when DLSS 4 is employed.  Even utilizing previous versions of DLSS, the card is delivering solid performance numbers at a resolution above what you’d expect.  That said, you’re probably going to be more impressed with the 1440p numbers.

Rasterized or otherwise, the RTX 5070 is a beast at 1440p!  Throw in DLSS 4 and it’s able to step into the ring at 4K.  That's pretty impressive and speaks volumes about how powerful AI-assisted framerates can be.  I’m sure there will be dozens of talking heads telling you otherwise, but numbers don’t lie.

There is one area where we did see a pretty solid drop – enabling RT (raytracing).  Raytracing has always been hard on cards, especially as you move down the stack, but here it can max out the 5070 at 4K. 1440p resolution gives enough headroom for RT, thankfully, and that’s the sweet spot for this card.  For example, turning on RT in the Monster Hunter Wilds benchmark at 4K caused thready results that would whip violently between 100 and sub-15 fps.  Turning it down to 1440p it handled it just fine.  As such, the numbers you see above are measured with RT turned off.  In practice, there are plenty of levers to pull and push to hit precisely the visual quality and bells and whistles addons to get RT at 4K – just know there will be tradeoffs.  

With these numbers we can now answer the question “Is this more powerful than a 4090”.  Well, the answer is an unqualified “sometimes”.  Probably not as definitive as you were looking for, but in a lot of circumstances, when aided by AI, yes – this card can actually topple the last generation king.  Let’s be realistic, though – you aren’t replacing a 4090 with a 5070, are you?  More likely you’re running an RTX 3070, 2070, or even the venerable 1080 Ti.  Does this card outperform those cards?  All day, every day, rasterized or otherwise – by around 65 to 70% by our measurements for the 3070.  Throw in DLSS and it’s not even a conversation.  

The last stop is the price.  You might look at the cards above this one and say absolutely not, but the RTX 5070 is priced to move.  The RTX 4070 launched at $599, but NVIDIA is launching the RTX 5070 at $549 MSRP.  Improved rendering performance that cuts video production time in half, generative AI speed improvements, Reflex 2 to cut latency, and DLSS 4 support allows this card to punch well above its base capabilities.  It’s very easy to recommend this card if you can’t financially find your way to the Ti model.  That said, and based on the launch of the other 5000 series cards, these cards will suffer from some artificial scarcity at launch, but be patient – don’t give into scalpers.  The wait will be worth it.