The premium Nvidia GeForce RTX 50 Series launch is now complete. Brought to you in four parts – RTX 5090, RTX 5080, RTX 5070 Ti, and RTX 5070 – it’s arguably most important that the newest and cheapest entrant, RTX 5070, does well. Starting at $549 /£539, it’s relevant to a wider range of enthusiasts, and you can read all about it in our in-depth review.
Nvidia knows most punters don’t upgrade on a generational basis. Rather, gamers seek replacements when their present card is no longer able to cut the mustard in newer games with elevated hardware requirements.
Released back in October 2020, I’d hazard gamers rocking a GeForce RTX 3070 card are looking at their faithful GPU and wondering if it’s time to give it the boot. Surely, they surmise, the all-new GeForce RTX 5070 is a great step forward, especially with the work Nvidia has put into genuine value-adds such as improved ray tracing and Multi-Frame Generation (MFG)
I’m here to answer this very question for you by comparing GeForce RTX 3070 against RTX 5070 across a number of games and applications. Should the newer card be leagues ahead, it may well be time to give the PC a spring boost.
First off, I’ll examine the tale of the tape by listing vital specifications.
| RTX 5070 | RTX 3070 | Ratio | |
|---|---|---|---|
| Released | Feb 2025 | Oct 2020 | – |
| Codename | Blackwell | Ampere | – |
| GPU | GB205 | GA104 | – |
| Process | TSMC 4N (4nm) | Samsung (8nm) | – |
| Transistors | 31.1bn | 17.4bn | 1.79 |
| Die size | 263mm² | 392mm² | 0.67 |
| CUDA cores | 6,144 | 5,888 | 1.04 |
| Boost clock | 2,512MHz | 1,725MHz | 1.46 |
| FP32 Boost TFLOPS | 30.9 | 20.3 | 1.52 |
| SM count | 48 of 50 | 46 of 48 | 1.04 |
| RT cores | 48 (4th Gen) | 46 (2nd Gen) | 1.04 |
| RT TFLOPS | 93.6 | 40 | 2.34 |
| Tensor cores | 192 (5th Gen) | 184 (3rd Gen) | 1.04 |
| FP16 Acc TFLOPS | 246.9 | 163 | 1.51 |
| ROPs | 80 | 96 | 0.83 |
| Memory | 12GB | 8GB | 1.50 |
| Memory type | GDDR7 | GDDR6X | – |
| Memory clock | 28Gb/s | 14Gb/s | 2.00 |
| Memory interface | 192-bits | 256-bits | 0.75 |
| Memory bandwidth | 672GB/s | 448GB/s | 1.50 |
| Board power | 250W | 220W | 1.14 |
| Launch MSRP | $549 | $499 | 1.10 |
Architecture
Advancements in silicon process enable Nvidia to shoehorn almost 80% more transistors into a die size that’s 33% smaller, but it’s how the transistors are used that’s changed over time.
You see, Nvidia’s shifts focus from allocating precious die space to typical Cuda shaders and moves instead towards laying down silicon for ray tracing and Tensor Cores. I know this because the Cuda core budget has barely increased even though heaps more transistors are available, while ROPs have decreased – whether by design or accident. Nevertheless, a substantially higher RTX 5070 clock speed helps keep raster throughput solid.
Speaking to the above, notice how much more performant RTX 5070 is with respect to ray-tracing and tensor-core throughput.
Keeping the chip small, Nvidia’s decision to drop memory bus width from 256 to 192 bits is understandable when GDDR7 allows the effective doubling of clock rate. All told, there’s 50% more bandwidth, even before Blackwell’s better on-chip caching gets involved.
A higher-frequency card usually entails increased power draw. That’s also the case with RTX 5070, whose 250W TBP is a little higher than RTX 3070. I still expect any user to easily switch between one and another without having to change PSUs and such like.
It used to be the case a perusal of headline specs informed accurate judgement of overall GPU potential. That’s no longer the case when upscaling technologies and MFG are added to the mix. Separating the wheat from the chaff, let me guide you through head-to-head benchmarks.
But before that happens, it’s important to appreciate time value of money. $499 back in 2020 is around $620 now, meaning RTX 5070 is actually cheaper in real terms. Bear that in mind when comparing results.
How I tested
Our 7950X3D Test PCs
Club386 carefully chooses each component in a test bench to best suit the review at hand. When you view our benchmarks, you’re not just getting an opinion, but the results of rigorous testing carried out using hardware we trust.
Shop Club386 test platform components:
CPU: AMD Ryzen 9 7950X3D
Motherboard: MSI MEG X670E ACE
Cooler: Arctic Liquid Freezer III 420 A-RGB
Memory: 64GB Kingston Fury Beast DDR5
Storage: 2TB WD_Black SN850X NVMe SSD
PSU: be quiet! Dark Power Pro 13 1,300W
Chassis: Fractal Design Torrent Grey
Application and AI
Over four years on, I’m seeing a 64% improvement in a standard raster-heavy benchmark.
The gap in Steel Nomad is a slightly lower 57%. I put this down to the non-ray-traced nature of the test, meaning RTX 5070 can’t bring its full RT arsenal out to play.
Not everyone’s busy playing games – I’m looking at you, Sam! Using a muscular GPU’s smarts for rendering is a useful barometer of how well it runs non-gaming workloads. In this case, RTX 5070 is almost twice as quick.
There’s not so much in it for basic AI tasks, though.
A common use of GPUs, running local AI for applications such as text generation shows a near-60% gain.
Gaming
Yet you’re really here for the games, aren’t you? Run at maximum settings, RTX 5070 is around 50% speedier at all resolutions. Put another way, you can go up one resolution with little loss of performance.
Featuring a modicum of ray tracing, numbers all make sense until we get to 4K. RTX 3070 does surprisingly well, leading me to believe it may not be running exactly the same settings. Investigations are afoot.
Here’s where things get messy for RTX 3070. The reasons are two-fold: there’s a heap of RT going on and its 8GB framebuffer struggles to contain the Forza workload.
The usual 50% uplift is magnified to a whopping 213% at QHD, and you can forget about 4K gaming at ultra-high settings on RTX 3070.
Remember I said the need for an upgrade comes in many forms? Here’s one of them.
Yet it’s horses for courses. Games that don’t exact a great toll on the GPU work well enough on older hardware. If Mount and Blade II was my thing, I’d not feel the need to splash the cash on RTX 5070.
4K60 used to be the holy grail for GPUs. What do you know, venerable RTX 3070 8GB can do that just fine.
Jeez, Louise, why are the scores so low? That’s because I’m running the very best version of the game that includes path tracing RT, playing into RTX 5070’s silicon hands.
I still find RTX 5070’s FHD return somewhat sticky. Time to whack on the DLSS.
A couple of points here. Run with DLSS Quality, scores for both cards rise in similar fashion that enable smoother gameplay. FHD now feels good on RTX 5070.
The RTX 50 Series’ ace in the hole is Multi-Frame Generation (MFG). It’s plain faster than FSR 3 used for RTX 3070. That said, though the numbers are impressive, there are words of caution. One needs a decent base framerate from which MFG works its magic upon – questions of latency arise when the base is too low – and in this case, I’d only be happy using it at FHD.
| Game | 3070 to 5070 % uplift at 4K |
|---|---|
| Assassin’s Creed Mirage | 52.9 |
| Final Fantasy XIV Dawntrail | 9.9 |
| Forza Motorsport | 208 |
| Mount and Blade | 46.1 |
| Rainbow Six Extraction | 47.4 |
| Cyberpunk 2077 DLSS + FG (QHD) | 202 |
Looking at it in an easy-to-digest table, gains range from a little to a lot.
Vitals
Both cards are much of a muchness as far as vitals are concerned. Makes sense, too, as power budgets are similar.
Conclusion
Nvidia GeForce RTX 5070 comes into the premium GPU fold armed with an attractive $549 price point. It’s a GPU that goes heavy on ray-tracing hardware and doubles down on Tensor Cores to facilitate MFG.
Like other RTX 50 Series cards, it’s a somewhat Jekyll and Hyde character, insofar as rasterisation potential takes a backseat to the newer technologies I spoke about. I feel this is the right way forward, especially as more games are recipients of MFG love, but such an approach, due to technical reasons, can put RTX 5070 in a tight spot when native frame rate is low.
It remains very useful to compare overall performance against a card which you may upgrade from. GeForce RTX 3070 fits the bill nicely as it came out over four years ago but is now starting to show its age in titles with heavy RT demands. The 8GB framebuffer is also a concern at higher resolutions.
Putting it all together, RTX 5070 typically offers 50% higher rasterisation performance for a similar power budget. Though I would have preferred a 16GB framebuffer for a $549 card, even the 12GB present on the card helps mitigate against some bottlenecking issues faced by RTX 3070 at higher resolutions.
Run with heavy RT and MFG – as Nvidia sees gaming evolving – paints RTX 5070 in a better light, where there’s scope for >100% improvement.
The bottom line is that if I had a card similar to RTX 3070 and it was now creaking under the weight of new game engines, RTX 5070 is a solid if unspectacular upgrade card.
GeForce RTX 5070
“Gear up for game-changing experiences with the Nvidia GeForce RTX 5070 and AI-powered DLSS 4.” Read our review.
GeForce RTX 3070
“The GeForce RTX 3070 graphics card delivers the performance that gamers crave, powered by Ampere—Nvidia’s 2nd gen RTX architecture.
