Nvidia RTX 5090 is twice as fast as RTX 4090 but with a big caveat

Nvidia has finally lifted the lid on its Blackwell GPUs, giving us our first look at the long-awaited GeForce RTX 50 Series graphics cards at CES. Lofty claims surround the flagship RTX 5090 in particular, as Team Green promises unprecedented uplifts over its predecessor, but not all is as it seems at first glance.

Although I didn’t expect to see RTX 5070 Ti and RTX 5070 make the cut so early in the run alongside RTX 5080, there are few surprise specs thanks to the slew of leaks in the run-up. The range does indeed pack faster GDDR7 memory with a maximum of 1.8TB/s bandwidth, but sticks between 12GB and 16GB on its consumer friendly cards, only breaking ranks with 32GB on its eye-wateringly expensive RTX 5090. This warrants concern given how important capacity has become in recent years, but both capacities sit above the 8GB floor.

Much of the focus this year lies with AI capabilities, leveraging those Tensor Cores under the hood, but RTX 5090 isn’t quite full-fat Blackwell. Nvidia suggests the architecture handles up to 4,000 AI TOPS, 380 ray tracing TFLOPS, and 125 shader TFLOPS, each of which is a little lower in the flagship. There’s enough room here to wiggle in a future RTX 5090 Ti, Super, or Titan, but the more likely reason is to distance the consumer range from its server efforts.

Connectors aren’t quite as bad as we initially thought. RTX 5090’s power requirements demand a single 600W PCIe Gen 5 cable or four standard PCIe eight pins using a dongle. RTX 5080 scales down to a single 450W Gen 5 or three PCIe eight pins. Finally, both RTX 5070 and 5070 Ti are slim enough for a single 300W Gen 5 or two eight pins.

All support the PCIe 5.0 interface and come with DisplayPort 2.1b UHBR20 ports on the rear, creating parity with AMD’s previous generation graphics cards. This primes them for the latest monitors with unrestricted bandwidth, although faster cables are on the way for future models.

Nvidia’s Founders Edition cards aren’t worlds apart from what came before, but they relocate the prior generation’s rear fan to the front in a more traditional dual layout. This is to aid the design underneath, brushing air past the vapor chamber, which cools the centralised high-density PCB, memory, and MOSFETs. I’ll miss the outwardly unique design on RTX 30 and 40 Series FE cards, but these are still as sleek as ever and hopefully pack the performance to boot. Just keep in mind you’ll need up to 304mm of space to fit the new models in your system.

Performance

Nvidia barely touched on what the new cards are capable of during its presentation, but set expectations high stating RTX 5070 rivals RTX 4090 at a fraction of the price, and RTX 5090 offers double the performance. Sure enough, figures published separately highlight a monumental generational uplift between the two flagships across games like Cyberpunk 2077, Alan Wake 2, and Black Myth: Wukong. The only trouble is that these numbers don’t tell the full story.

Glance at the footnotes and you’ll see a wealth of caveats. Alongside comparisons throwing caution to the wind with different floating-point formats (FP8 for RTX 4090 and FP4 for RTX 5080), all performance boosts hinge on the newly introduced DLSS 4, exclusive to Blackwell. This isn’t inherently a problem since the days of purely rasterised performance uplifts are long gone, but comparing it to the wider space becomes difficult due to upscaler support limitations and sluggish adoption rates.

Doubling down on neural rendering technologies, DLSS 4 works similarly to Frame Generation, but instead of sandwiching a generated frame between two real ones, it makes a Big Mac out of it. Dubbed Multi Frame Generation, Nvidia has improved its AI Optical Flow process to slot up to three frames between traditionally rendered alternatives, which Nvidia claims creates up to eight times the performance over rasterised. Think 4K resolution at 240fps in Cyberpunk 2077.

DLSS 4 arrives with several enhancements to previous components, too. Frame Generation, Ray Reconstruction, Super Resolution, and DLAA (Deep Learning Anti-Aliasing) all get a facelift thanks to a new DLSS Transformer model working in the background. The result is marginally increased frame rates, reduced memory usage to alleviate that capacity pressure, improved temporal stability, and less ghosting, all adding up to something greater.

As you might expect, these are best leveraged by RTX 50 Series’ AI Management Processor and new Blackwell Tensor Cores, but older cards can enjoy the benefits too, depending on compatibility. Strange to think that 15 out of 16 pixels are generated by AI using all these features in tandem, but Team Green states they’re of higher quality than native rendering. We’ll be the judge of that when we get our hands on one to review.

The main issue is that these gigantic uplifts are only available in games with DLSS 3.5, where Single Frame Generation is available for Multi Frame Gen to piggyback. Given how nebulous DLSS has become over the years being a suite rather than a feature, it’s tough to tell how many games currently support Frame Generation in total or the speed of adoption, but I can hazard a guess.

Nvidia states DLSS 4 will have day one support for 75 games and apps, including Alan Wake 2, Cyberpunk 2077, and Portal with RTX. Assuming this is the total number of available Frame Generation titles, we’ve approximately seen less than five games and apps get support per month since its September 2023 debut, minus the three aforementioned originals. If there are more out there, then you will be able to use the DLSS Override feature in the Nvidia app to enable Multi Frame Gen over Single, but the relatively slow rollout puts a dampener on things.

Remove Frame Generation entirely, and both RTX 5090 and RTX 5070’s performance uplifts dwindle to anywhere between 1.25x and 1.4x their predecessors. Since Ubisoft’s shooter doesn’t support DLSS, Far Cry 6 spotlights the only fully rasterised comparison, showing the weakest increase of the lot. This could be due to CPU bottlenecks as we’ve seen limitations in the game before now, but you can observe similar performance in A Plague Tale: Requiem, which is confined to DLSS 3 without Frame Gen tech.

It’s clear to see the benefits of AI in gaming, but before you get sucked into the promise of RTX 4090 performance* that’s 63% cheaper, make sure you understand that big, bold asterisk.