Nvidia GeForce RTX 5070 Founders Edition review: more RTX 3090 than 4090

From the outset, Nvidia GeForce RTX 5070 has something of an uphill battle on its hands. Not only does it need to prove Blackwell architecture scales down to the midrange where competition is significantly tougher, but it also wears an albatross thanks to Team Green’s tongue-in-cheek nod to RTX 4090 performance wrapped in an itty-bitty price tag.

With MSRP set at $549 for Nvidia’s Founders Edition, it’s a tantalising idea that comes with many asterisks. First, your only hope to achieve such muscle is through the magic of RTX 50 Series’ new Multi Frame Generation (MFG) feature, which itself comes with a caveat or two. Second, maintaining a decent price point or indeed available stock has proven challenging for the company owing to high demand, thinner production, and partner premiums.

Only one thing’s certain before heading into benchmarks: Nvidia doesn’t have free reign over the midrange like it does with the high-end and enthusiast arenas. It needs to tread carefully in order to go up against AMD Radeon RX 9070 and RX 9070 XT, which we’ll see more of tomorrow, so keep your eyes peeled.

For now, let’s dive into the Club386 Tables of Doom™ to see how it fares against the rest of Nvidia’s graphics cards both past and present.

Specifications

	RTX 5070	RTX 5070 Ti	RTX 5080	RTX 5090
Released	Feb 2025	Feb 2025	Jan 2025	Jan 2025
Codename	Blackwell	Blackwell	Blackwell	Blackwell
GPU	GB205	GB203	GB203	GB202
Process	TSMC 4N (4nm)	TSMC 4N	TSMC 4N	TSMC 4N
Transistors	31.1bn	45.6bn	45.6bn	92.2bn
Die size	263mm²	378mm²	378mm²	750mm²
CUDA cores	6,144	8,960	10,752	21,760
Boost clock	2,512MHz	2,452MHz	2,617MHz	2,407MHz
FP32 Boost TFLOPS	30.9	43.9	56.3	104.8
SM count	48 of 50	70 of 84	84 of 84	170 of 192
RT cores	48 (4th Gen)	70 (4th Gen)	84 (4th Gen)	170 (4th Gen)
RT TFLOPS	93.6	133.2	170.6	317.5.1
Tensor cores	192 (5th Gen)	280 (5th Gen)	336 (5th Gen)	680 (5th Gen)
FP16 Acc TFLOPS	246.9	351.5	450.2	419
ROPS	80	96	112	176
Memory	12GB	16GB	16GB	32GB
Memory type	GDDR7	GDDR7	GDDR7	GDDR7
Mem. clock	28Gb/s	28Gb/s	30Gb/s	28Gb/s
Mem. interface	192-bits	256-bits	256-bits	512-bits
Mem. bandwidth	672GB/s	896GB/s	960GB/s	1,792GB/s
Board power	250W	300W	360W	575W
Launch MSRP	$549	$749	$999	$1,999

As the current entry-level to Blackwell short of any unannounced RTX 5060 Ti, GeForce RTX 5070 is the brand’s value offering built using the same TSMC 4nm node as its siblings. In order to trim prices down, it features a GB205 die that’s 30% smaller than its bigger brother.

All five GPCs (Graphics Processing Clusters) are up and active, with 48 streaming processors and ray tracing cores doing most of the heavy lifting. Concessions from its Ti alternative save you 27% on MSRP while sacrificing up to 32% of its core components. It’s a small hit to value, but all the more pronounced in this performance bracket.

The part that’ll raise eyebrows most is the drop to 12GB of memory. Moving to GDDR7 certainly helps speed things up with the same 28Gb/s bandwidth improvements as RTX 5090, but such a small buffer on a 192-bit bus will fill up fast with modern games. If nothing else, it has efficiency in its corner, relatively sipping power at no more than 250W.

	RTX 5070	RTX 4070 Super	RTX 3070
Released	Feb 2025	Jan 2024	Oct 2020
Codename	Blackwell	Ada Lovelace	Ampere
GPU	GB205	AD104	GA104
Process	TSMC 4N (4nm)	TSMC 4N (4nm)	Samsung 8nm
Transistors	31.1bn	35.8	17.4bn
Die size	263mm²	294.5mm²	392mm²
CUDA cores	6,144	7,168	5,888
Boost clock	2,512MHz	2,475MHz	1,725MHz
FP32 Boost TFLOPS	30.9	36	20.3
SM count	48	56	48
RT cores	48 (4th Gen)	56 (3rd Gen)	46 (2nd Gen)
RT TFLOPS	93.6	82.1	39.7
Tensor cores	192 (5th Gen)	240 (4th Gen)	184 (3rd Gen)
FP16 Acc TFLOPS	246.9		162.6
ROPS	80	80	96
Memory	12GB	12GB	8GB
Memory type	GDDR7	GDDR6X	GDDR6
Mem. clock	28Gb/s	21Gb/s	14Gb/s
Mem. interface	192-bits	192-bits	256-bits
Mem. bandwidth	672GB/s	504GB/s	448GB/s
Board power	250W	220W	220W
Launch MSRP	$549	$599	$499

Reflecting on the graphics cards it replaces; a smaller die ultimately means less space for transistors and cores. Instead, the focus is on better architecture. Blackwell comes with a wealth of improvements in the form of 4th Gen RT Cores, 5th Gen Tensor Cores, faster memory, a fancy new PCIe Gen 5 interface, and bespoke features such as DLSS 4’s Multi Frame Generation.

Nvidia also touts plenty for professionals. Support for FP4 improves GenAI performance, while dedicated hardware accelerated 4:2:2 support (NVENC) takes aim at speeding up video and graphics rendering via the GPU.

To compensate for fewer components under the hood, Nvidia rightly trims MSRP back by $50 compared to its predecessor. The main question is whether this is enough to entice people away from RTX 4070 Super, although Ada Lovelace availability is exceptionall thin, making this something of a moot point.

RTX 5070 will cost you a touch more to run with higher power requirements. It’s nothing out of the ordinary, actually maintaining a slimmer chassis that’ll slip right into small form factor (SFF) builds, but something to keep in mind when counting pennies.

Founders Edition

From Parm’s glowing GeForce RTX 5090 review to my own RTX 5080 breakdown, Nvidia’s gorgeous Founders Edition design is a fan favourite here at Club386. RTX 5070 follows suit, but in an almost miniaturised version.

All three are SFF-ready certified, but this midranger boasts the widest compatibility at just 242mm x 112mm. The numbers suggest it’s only 20% smaller than its siblings, but they absolutely dwarf it side-by-side. All in all, business as usual for a 70-class card.

The shrunken chassis isn’t entirely without compromise. Unlike the high-end models, you won’t see a lighting strip anywhere on this dinky device. It’s not much of a loss as most people wouldn’t notice the illumination in its default horizontal configuration anyway and even vertical users can bask in the glory of its two-tone gunmetal grey and black aesthetic.

Far from all form, shifting both fans into a more traditional configuration serves a functional purpose, too. It creates linear airflow past an all-new vapour chamber with heat pipes stretching from end to end. A concaved fin stack on the other side helps reduce pressure, curved to match the surface area of each of the fans’ seven blades. Any thermal overflow heads out the angled exhausts on the side, keeping the card cool.

Directly in the centre of it all lies a tiny 263mm² die, controlling Nvidia’s smallest-ever PCB. Of course, moving everything to the middle necessitates some forward-thinking to reach the I/O and PCB, calling upon proprietary boards and ribbon cables to make up the distance.

GB205 might be smaller, but Nvidia hasn’t skimped on the display connectors. Just like the rest of Blackwell, you get your pick of three DisplayPort 2.1b slots and a single HDMI 2.1b, each capable of up to 4K at 480Hz or 8K at 165Hz with display stream compression (DSC). You’ll have a tough time saturating such lofty resolutions even with the magic of MFG.

GeForce RTX 5070 carries the same 12V-2×6 power connector as its brethren, capable of delivering up to 600W across a single cable. For as mired in controversy as it is, there fortunately might not be cause for concern here.

Starting life as 12VHPWR, 12V-2×6 came about as a revision with 0.1mm shorter pins and 0.15mm longer terminals in an attempt to curb melting ports. Sadly, the same issues appeared on RTX 5080 and 5090, but their absence on RTX 5070 Ti spells good news for RTX 5070. Investigations are ongoing, but evidence points towards higher voltage across single pins generating too much heat to bear, which is less of an issue for a 250W model.

Nobody should have to seek out third-party solutions to fix pain points, but there are options for peace of mind. MSI adapters (sold separately) feature yellow tips so you can identify when your cable is properly seated, while custom cables now come fitted with cooling fans and pin sensors that warn you when things get a little too close for comfort.

It’s still best to pair your RTX 5070 with a modern power supply outfitted with a 12VHPWR header for a cleaner single-cable look, but there’s no inherent need to upgrade the PSU you have. Nvidia bundles a two-way eight-pin splitter that’ll work with most modern models, so long as you sport at least 650W by Nvidia’s recommendation.

Akin to the rest of Blackwell, you can likely get away with less than Nvidia’s baseline power requirements. Running the most demanding games never once saw RTX 5070 peak past 351W alongside our Ryzen 9 7950X3D processor and 64GB of DDR5 RAM. Of course, your mileage may vary if you have every SSD slot filled and RGB stacked to the nines.

Perhaps the most impressive thing about RTX 5070 is that it manages to keep fairly cool and quiet despite a tiny chassis with smaller fans. These little blowers are as silent as they come at 36.6dB while keeping temperatures at no more than 70°C.

Performance

Like all Blackwell GPUs before it, I’ve sat Nvidia GeForce RTX 5070 Founders Edition down in the Club386 test bench. Supported by the venerable AMD Ryzen 9 7950X3D and 64GB of DDR5 memory, we can determine what ceilings this card hits.

Application & AI

Putting its ray tracing chops to the test, GeForce RTX 5070 pushes past its predecessor by 12% in 3DMark Speed Way. We have new 4th Gen RT Cores to thank for this, but it is ultimately just half the uplift we’ve seen from Ada Lovelace to Blackwell with previous entries.

Fortunately, performance generally sits in line with its price tag. RTX 5070 is 24% behind its Ti alternative, which reflects its 27% cheaper cost.

The gap between Blackwell and Ada Lovelace 70-class cards reduces in rasterised performance, with just 9% in it between them in 3DMark Steel Nomad. It’s already clear that you’re better reframing this as an Ampere upgrade, benefitting from a 58% uplift compared to RTX 3070.

Practically neck-and-neck, creators will have a slightly better time with GeForce RTX 4070 Super than its modern alternative. That said, you’d be hard-pressed to find anything wrong with 6,188 samples per minute in Blender at this price tag.

Flexing its artificial intelligence brawniness, RTX 5070 nets an impressive half-precision score of 48,443 in Geekbench AI. Not only is this 9% ahead of its forebearer, but it also takes an 8% lead over RTX 4070 Ti Super and is less than a percent away from RTX 4080 Super – its 5th Generation Tensor Cores at work.

Blackwell alone isn’t enough to keep pace in Procyon AI Text Generation, as RTX 5070 slips right back into place between RTX 4070 Super and RTX 4070 Ti Super. Scoring 3,439 in Llama 3.1, it’s still a 12% improvement over the former, but large language models certainly benefit from beefier graphics cards.

Game Rasterisation

You can count on synthetic tests to paint a picture, but I know it’s games that you’re really after. Putting RTX 5070 to the test in our usual triple-A suite, I’ve not only benchmarked the Blackwell card at its intended 1920×1080 (FHD) and 2560×1440 (QHD) resolutions, but also 3840×2160 (UHD). The results are fairly surprising.

Even bouncing around 9th Century Baghdad in Assassin’s Creed Mirage, RTX 5070 finds itself snug between RTX 4070 Super and RTX 4070 Ti Super. Though, with no more than 11fps in it between the three of them at every resolution, it’s a lot like splitting hairs. Usually, we’d expect a 20% generational performance leap, but we see no more than 4% here.

It’s only when comparing the latest card against its Ampere alternative that you find rhyme or reason, offering a 53% leap over RTX 3070 at UHD and 45% at QHD.

GeForce RTX 5070 is almost indistinguishable from RTX 4070 Super in Final Fantasy XIV: Dawntrail. There’s just 1fps between the two cards at higher resolutions, while QHD stretches the distance to a mere 5%.

Its lower MSRP certainly takes out the sting of waning rasterised efforts, but there’s still a question of where this newbie belongs against graphics cards both past and present. At the very least, it’ll placate those coming from RTX 3070 with up to a 50% performance increase.

Looking at it in isolation, Nvidia’s latest x70 does well to keep pace in Forza Motorsport given how demanding the racing game is. Falling short of our 60fps goal at UHD is entirely forgivable since even RTX 5080 can’t manage that. It still nets 72fps at QHD, which is its target audience.

However, viewing the wider context introduces the same disappointment as previous benchmarks. Both RTX 4070 Super and 5070 are practically tied with a 1-5fps gap, leaving little to sing about. Still, RTX 3070 owners won’t scoff at the 213% boost at QHD.

Brace yourself for deja vu, as Mount & Blade II: Bannerlord tells the same story we’ve seen several times now. GeForce RTX 5070 achieves a tiny 4% leap over its predecessor, while slipping between 5-10% behind RTX 4070 Ti Super.

Once again, it sounds far better put up against Ampere being 53% stronger than RTX 3070. You can see where I’m heading with this.

RTX 5070 wavers in Tom Clancy’s Rainbow Six Extraction, slipping 10fps behind its antecedent at FHD. Barely regaining its footing, the card slips back into its RTX 4070 sandwich at QHD before taking a paltry 3fps lead at UHD. It’s all very much of muchness with triple-digit frame rates but it’s clear Blackwell’s rasterised performance struggles to scale downwards.

	Average increase or decrease at 1440p
RTX 5080 to RTX 5070	–28%
RX 7900 XTX to RTX 5070	-22%
RTX 4080 Super to RTX 5070	-22%
RTX 5070 Ti to RTX 5070	-18%
RX 7900 XT to RTX 5070	-10%
RTX 4070 Ti Super to RTX 5070	-5%
RTX 4070 Super to RTX 5070	+3%
RX 7800 XT to RTX 5070	+17%
RTX 3080 to RTX 5070	+20%
RTX 3070 Ti to RTX 5070	+66%
RTX 3070 to RTX 5070	+81%

GeForce RTX 5070 is certainly a capable UHD graphics card in some games, but it’s a midranger through and through. You’ll find most of its value packed in the 1440p ballpark, although it’s not a particularly shining example of generational improvements.

Comparing Blackwell to its Ada Lovelace predecessors, RTX 5090 sees a relatively ample 25% improvement over RTX 4090. Both RTX 5080 and RTX 5070 Ti sit comfortably with a 15% uplift. Meanwhile, RTX 5070 scrapes by with just over 3% more muscle than RTX 4070 Super at both QHD and UHD. Middling by designation, but not at all by nature.

Cast your mind back to 2020’s Ampere and the graphics card starts to make some sense, providing up to 81% more than its RTX 30 Series counterpart can muster.

Frame Generation

Most of Blackwell’s strength doesn’t come from rasterisation, but instead its Frame Generation feature set. This affords graphics cards the ability to insert AI frames in between natively rendered ones. Whereas RTX 40 Series was only capable of sandwiching one, labelled FG 2x, RTX 50 Series comes with Multi Frame Generation, adding a second and third artificial frame into the mix.

3DMark’s DLSS test is the perfect place to marvel at just how far Nvidia’s upscaling suite has come over the years. Using RTX 5070 as a prime example, it goes from a shaky 56fps riddled with unwanted visual artifacts to a much cleaner image with triple the frame rate at 154fps using MFG 4x. Multi Frame Generation means business.

Older cards, such as RTX 4070 Super, still benefit from other DLSS 4 enhancements, such as better standard Frame Gen, Ray Reconstruction, Super Resolution, and Anti-Aliasing (DLAA), but MFG is exclusive to Blackwell. It leverages improved hardware-based Flip Metering to more accurately pinpoint the timing of each synthetic frame. This means we can entirely attribute RTX 5070’s 88% lead to MFG 4x alone.

The first caveat is that MFG requires developer implementation, which means support isn’t guaranteed. Nvidia assures that there will be at least 75 games with the entire DLSS 4 suite, most of which have already released.

Sadly, Call of Duty: Black Ops 6 isn’t one of them, but it showcases that standard Frame Generation behaves identically on 70-class cards. Previously, we saw up to a 16% advantage moving from RTX 4070 Ti Super to RTX 5070 Ti, and a 10% leap from RTX 4080 Super to RTX 5080, making this 2% shuffle a bit tepid.

Cyberpunk 2077 is an incredibly demanding game using its Raytracing Overdrive preset, and it really puts RTX 5070’s hardware up against a wall. While both midrangers suffer with 12GB memory buffer and 192-bit bus at UHD, fewer transistors and cores are what see Blackwell fall behind its predecessor when each GPU is firing on all cylinders.

This makes MFG something of a double-edged sword. On one hand, it practically doubles the frame rate ceiling against anything RTX 4070 Super can produce. On the other, a lower native frame rate will incur latency penalties even as the visible fps climbs higher. As a result, Nvidia recommends a base of no less than 60fps so you can prevent a sluggish desynchronisation between your mouse and what you see on screen.

With path tracing, Alan Wake 2 is a game that relies on upscalers and AI wizardry to achieve workable frame rates. Like a monkey’s paw, MFG grants you up to a 236% uplift with a respectable 94fps at UHD, but it comes with the caveat of dreaded lag.

Usually, your latency decreases as your native frame rate increases, but introducing synthetic frames into the mix creates the inverse effect. Thanks to a comfortable 79fps DLSS Quality base, FHD with MFG running at full pelt still sees less lag than bog standard UHD without any Frame Gen magic.

	Native Latency	FG Latency	FG 3x Latency	FG 4x Latency
5070 @ UHD	95.6ms	114.7ms	120.9ms	126.6ms
5070 @ QHD	53.5ms	64.2ms	69.6ms	73.4ms
5070 @ FHD	37.6ms	45.8ms	49.3ms	51.4ms

Some people are more susceptible to the feel of input lag than others, meaning each reviewer will place a different weight on the metric. Personally, I think it’s an acceptable trade in immersive single-player games like Alan Wake 2 and would opt for QHD for a sharper image, but I’d steer clear of competitive games where reaction times mean life or death – even with Nvidia Reflex curbing the penalty.

It’s important not to get bogged down in the wrong numbers. I’ll stand by that MFG is the most innovative graphics card solution since DLSS first reared its head and I’ve no doubt it provides the foundations of the future, but it doesn’t automatically make lower-end graphics cards as good as their higher-end alternatives.

Overclocking

For the few willing to tinker, Nvidia has left enough headroom to overclock so you can achieve a better base frame rate to aid MFG. Memory easily scales a little higher than 28Gb/s, with scanners like MSI Afterburner identifying a decent overclocked curve, pushing in-game frequencies from 1,750MHz to 2,000MHz.

Using Final Fantasy XIV: Dawntrail at QHD as the guinea pig, I saw average frame rates climb up from 152fps to 158fps. Not enough to overtake RTX 4070 Ti Super, but plenty in the grand scheme of things. Frame drops weren’t very noticeable with a 93fps minimum to begin with, but this also raised to 97fps for added peace of mind.

A 4% leap isn’t tremendous, but it’s free performance that could very well make the difference between sub-60fps and the right frame rate for Multi Frame Generation.

Conclusion

GeForce RTX 5070 Founders Edition brings with it the weakest rasterised gains we’ve seen in Blackwell to date, confirming it’s not intended as an RTX 40 Series upgrade. Flip the calendar further back and there is value in taking the leap from Ampere, with sufficient overclocking headroom and a focus on efficiency.

Until more mainstream Blackwell iterations come out the woodwork, GeForce RTX 5070 Founders Edition remains the cheapest way to get Nvidia’s latest technology, and make no mistake: Multi Frame Generation is impressive. It’s just not a silver bullet that’ll turn your 70-class card into an RTX 4090.

You’ll need to be mindful of its drawbacks in this performance bracket to make the most of it, which isn’t small ask at $549. Add to that AMD Radeon RX 9070 starting at the same price, and Nvidia potentially has a tough fight on its hands. Keep an eye out for RDNA 4 reviews tomorrow.