NVIDIA Unveils GeForce RTX 50 Blackwell, Up To 3X Faster Than RTX 4090
Announced tonight at NVIDIA's keynote in a surprisingly windy Las Vegas, the GeForce RTX 50 series represents NVIDIA's latest efforts at bringing its generally datacenter-focused chip development efforts back home to gamers demanding the latest rendering technology and the highest GPU performance. So without further ado, let's actually talk about the GPUs.
Obviously, first up is the GeForce RTX 5090, and the most important specification for a lot of you is the price: $1999 USD. That's an increase over even the $1599 of the GeForce RTX 4090, but it's not just inflation to blame, as the new GPU is a whopping 92 billion transistors. You might just get your money's worth, though, as the specifications are staggering: 170 Streaming Multiprocessors (up from 128 in RTX 4090), and a full 32GB of GDDR7 memory running at 28 Gbps. That delivers a total memory bandwidth just a hair under 1.8 TB/second, roughly an 80% increase from the previous generation.
Thanks almost entirely to the large size of the GPU and architectural changes, the GeForce RTX 5090 apparently offers 3400 AI TOPS—fully 2.5 times that of the GeForce RTX 4090, going by NVIDIA's own numbers. Of course, AI TOPS aren't that relevant to gaming workloads, which is what the majority of you dear readers are probably most interested in. NVIDIA didn't produce any
In addition to the top-end GeForce RTX 5090, NVIDIA also announced the GeForce RTX 5080, RTX 5070 Ti, and RTX 5070. The GeForce RTX 5080 is essentially one-half of a GeForce RTX 5090, with 16GB of RAM on a 256-bit bus, and it comes with a matching price tag of $999. It does feature slightly higher clock rates on its memory—a dizzying 30 Gbps—so the performance is probably a bit more than half of the flagship.
The GeForce RTX 5070 Ti is a further step down at $749; it also comes with 16GB of RAM, and then the GeForce RTX 5070 with 12GB of GDDR7 memory is attractively priced at $549. NVIDIA claims that even the GeForce RTX 5070 is faster than the GeForce RTX 4090, although we think that's likely to be both in a path-traced workload and taking DLSS 4 into account.
Indeed, DLSS 4 is real and NVIDIA says that it has moved away from convolutional neural networks toward a new new transformer model to increase both image quality and performance. However, the big feature this time around is multi-frame generation. DLSS 4 is capable of increasing frame rates considerably by generating up to three frames between the actual classically-rendered frames. Jensen Huang says that out of four full-resolution 4K output frames—approximately 33 megapixels of image data—only 2 megapixels have to be computed traditionally.
That's not the only AI feature NVIDIA premiered tonight. RTX Neural Shaders are small AI networks embedded within shader code. NVIDIA says that they can improve both image quality and performance with a smaller memory footprint compared to typical shader code, like with NVIDIA's Neural Texture Compression. Also, NVIDIA RTX Neural Face appears to be a new middleware library that uses AI to improve the realism of life-like, interactive faces rendered in real-time.
Arguably the most exciting new feature of 'Blackwell' is actually the introduction of something called RTX Mega Geometry. NVIDIA claims that this feature allows the GeForce RTX 50 series to ray-trace one-hundred times the triangles of Ada Lovelace GPUs. We're eager for more details on Mega Geometry, but Huang didn't actually mention it in his brief GeForce RTX 50 presentation.
Finally, NVIDIA announced the second generation of its Reflex input lag reduction technology. Intended for use in conjunction with DLSS 4, this is supposedly the secret sauce that allows the multi-frame generation technology to work without making every game feel like a complete lag-fest. The original version of Reflex offers tangible benefits in terms of input lag reduction, so we're eager to test NVIDIA's claims here.
Unusually, NVIDIA also revealed the mobile GeForce RTX 50 lineup tonight. These range from the top-end GeForce RTX 5090, equipped with 24GB of memory and delivering 1824 TOPS, down to the GeForce RTX 5070 with 8GB of RAM and producing 798 TOPS. As with the previous generation, it looks like NVIDIA is essentially using GPUs one tier down for the mobile parts, although the GeForce RTX 5090 (mobile) doesn't map directly to any of the desktop parts, so it will be interesting to see for sure.
As far as the desktop GPUs go, all of these parts are on the way to store shelves this month. Meanwhile, laptops equipped with the new GeForce RTX 50 mobile GPUs will be coming around toward the end of this quarter, or in March.
