When you buy through links in our articles, we may earn a small commission, which does not affect our editorial independence.
It’s official: Nvidia’s next generation of gaming graphics is here, and friends, the GeForce RTX 50-series looks pretty compellingly priced on paper. Nvidia CEO Jensen Huang announced eight different GeForce RTX 50-series graphics cards during his CES 2025 keynote – four for desktop, and four for laptops, all compatible with a new DLSS 4 generation.
The same GPUs were announced for both form factors: The GeForce RTX 5090, 5080, 5070 Ti, and 5070. It was a big difference from AMD’s keynote, where RDNA 4 and the Radeon RX 9070 weren’t even mentioned despite press receiving a high-level briefing.
Let’s start with Nvidia’s hotly anticipated desktop graphics cards, powered by Nvidia’s new Blackwell architecture. They’re actually cheaper than expected, with lower prices than their predecessors…except for the RTX 5090. But there’s a reason for that.
I was convinced the GeForce RTX 5090 would cost $2,500 or more, based on leaked specs. Well, Jensen didn’t actually mention product specifications, however, at $1,999, it’s a low price for AI researchers, if not necessarily gamers. With 32GB of VRAM and a rugged 512-bit bus, along with Nvidia’s flagship Blackwell GPU, this device will be fierce in gaming, but unprecedented in device learning. responsibilities if you can’t a beloved Nvidia business card.
The theme goes down the line: At $999, the RTX 5080 costs $200 less than the 4080 did at launch (and the same as the RTX 4080 Super). At $749 and $549, the RTX 5070 Ti and RTX 5070 cost $50 less than their predecessors (though to be fair, I called the RTX 4070 Ti “hobbled and wildly overpriced”). Better yet, Jensen said that the RTX 5070 delivers RTX 4090-class performance for roughly a third of the price.
Inside, the new Nvidia Founders Edition dual-flow dual-fan design, with a small PCB.
NVIDIA
If that’s true in games and not just AI workloads or benchmarks, it’s an exciting start to the new generation – but be warned that these starting prices for Nvidia’s overhauled Founders Edition versions may not reflect the same price you’ll see on custom third-party cards. Nvidia’s announcement post for the 50-series, published after the keynote, includes some high level performance details for each desktop GPU, but only in full ray tracing and DLSS-enabled situations. Since we have no tangible idea what sort of performance benefits DLSS 4 brings, take these with a bunch punch of salt for now. It shows what’s possible though.
Nvidia says the RTX 5090 is “2X Faster Than The GeForce RTX 4090;” the RTX 5080 is “Twice As Fast As The GeForce RTX 4080;” the RTX 5070 is “2X Faster Than The GeForce RTX 4070 Ti;” and the RTX 5070 is as fast as the RTX 4090.
Again, expect to hit those peak levels only in fast-paced games and scenarios that feature Nvidia’s full suite of ray tracing and DLSS features. The RTX 5090 offers many more CUDA cores than the 4090, so it deserves to be a beast no matter what, while the lower-tier RTX 50 cards have more modest CUDA upgrades, so Blackwell’s architectural tweaks (and DLSS 4) possibly wants to be modified. do the heavy lifting if Nvidia plans to deliver big functionality improvements. We’ll see!
Look for RTX 50 series graphics cards to start hitting the market in some form at the end of January. Nvidia specified which cards would be released and when.
Usually, Nvidia launches desktop GPUs well before their laptop counterparts, but not with the RTX 50-series! Well, kinda. They’re launching in March, while the desktop cards launch this month, but Jensen revealed both onstage.
Details were scarce beyond what is seen in the screenshot above. The GeForce RTX 5090 will be available on laptops starting at $2,899; the RTX 5080 in laptops starting at $2,199; the RTX 5070 Ti in laptops starting at $1,599; and the RTX 5070 on laptops starting at $1,299.
However, pay close attention to the TOP AIs indexed under the style numbers for computer and desktop GPUs. Laptop graphics tend to be scaled down compared to their desktop cousins, and AI TOPS suggests that’s the case with the RTX 50 series as well. The RTX 5090 desktop card offers 3400 AI TOPS, while the desktop edition it hits 1850, about the same as the desktop RTX 5080. (Not surprising, since the 4090 computer is necessarily a 4080 desktop built into a computer. ) The same trend continues. in the 50 series line of computers.
Update: Nvidia’s GeForce online page provided more specifications after the keynote, and yes, my initial suspicions were correct. Also be sure to pay attention to the memory capacity (especially for the RTX 5070) and bus width size, as well as the number of CUDA cores. This 5090 computer doesn’t have anywhere near the firepower of the 5090 desktop (Nvidia has wisely improved the memory capacity of the 5090 over that of the 5080 desktop), and the 5070 only offers 8 GB of VRAM on a 128-bit bus. – a general shortcoming that, frankly, 1080p resolution is rarely enough in the newer games of 2025, Nvidia can claim that the AI advantages of DLSS 4 reduce memory requirements.
That’s all fine and dandy, but how are those puppies behaving? Jensen didn’t go into details in the stream, but he conveyed some high-level Blackwell specs that are more important to AI researchers than gamers: AI TOPS, RT TFLOPS, “AI control processor” matrix, etc. You can see everything in the screenshot above.
On the plus side, at least some of Nvidia’s RTX 50-series will include bleeding-edge GDDR7 memory, helping to deliver up to 1.8TB/s of memory bandwidth, or twice what was possible with its predecessor. It will also offer dual shaders for INT and floating point calculations (the big two for traditional gaming graphics), the ability to intermix GPU and AI workloads as needed, and programmable GPU shaders that can process neural network tasks.
While it turns out that Nvidia is betting on AI, well, it has been for years. Haven’t you noticed DLSS and DLSS 3 Frame Gen? On that note, Jensen also teased DLSS 4, which features “neural texture compression” and “neural materials” that further reduce the need for classic GPU rendering.
Nvidia claims that while DLSS 3 can simply inject AI-generated frames between GPU-rendered frames, DLSS 4 can infer three full frames from a single classic frame. Hey, after all, AI scaling does away with local graphics. Jensen said a total of about 33 million pixels are generated for four frames of fourK images; with DLSS four, that’s still true, but the GPU only does about two million, and the AI does the rest of the heavy lifting. You can learn more about DLSS four and multi-frame generation on Nvidia’s online page if you’re interested.
As DLSS 3 users who forgot to allow Nvidia Reflex can tell you, injecting AI-generated frames also injects latency – Reflex can recover it. This is a much more pronounced challenge when employing AI to create multiple images. Enter Nvidia Reflex 2, which uses “frame warping generation” to lessen latency through up to 75%. It’s complicated, so watch the video and click this link for a detailed explanation of the generation if you need more details.
Bonkers – and potentially very, very game changing if it works as advertised. That’s a big if, though, and there are plenty of questions still swirling around the RTX 50-series’ capabilities after Nvidia’s detail-light keynote. Stay tuned to PCWorld (and PCWorld’s YouTube channel!) for all the latest updates and news coming from CES 2025.
Editor’s note: This article has been updated to include videos of MSI and Gigabyte 50 series graphics cards.
Brad Chacos spends his days poking around on desktop computers and tweeting too much. It specializes in graphics cards and gaming, but covers everything from security to Windows tricks and all types of PC hardware.
Business
Laptop
Mobile
Hardware
Storage
Offers
TechHive
Digital Magazine – Subscribe
Digital Magazine – Info
Gift Subscription
Newsletters