There’s just something about a new graphics processor architecture that gets everyone’s juices flowing, either to covet it or complain. Or oddly, sometimes both. So happy CES from Nvidia!

The company on Monday announced four 50-series desktop graphics cards, plus their laptop equivalents, using its new Blackwell architecture. That doesn’t preclude an even lower-end model arriving later — it’s likely we’ll see RTX 5060 and RTX 5050 options, at the very least mobile versions for entry-level gaming or creative laptops. The market for the latter has changed a bit this year, though, with graphics integrated into the CPU getting beefed up in order to handle AI and basic gaming.

GeForce RTX 5090GeForce RTX 5080GeForce RTX 5070 TiGeForce RTX 5070 Memory 32GB GDDR716GB GDDR716GB GDDR712GB GDDR7Memory bandwidth (MBps) 1,792960896504Memory clock (GHz) n/an/an/an/aGPU clock (GHz, base boost) 2.01/2.412.30/2.622.30/2.452.16/2.51Memory data rate/Interface 512-bit256-bit256-bit192-bitRT cores 170847048CUDA cores 21,76010,7528,9606,144Texture mapping units 680336280192Streaming multiprocessors 170847048Tensor cores (trillion operations per second) 680 (3,352 AI TOPS)336 (1,801 AI TOPS)280 (1,406 AI TOPS)192 (988 AI TOPS)Process 4nm4nm4nm4nmMax thermal (degrees) 194F/90C190F/88C190F/88C185F/85CBus PCIe 5.0×16 PCIe 5.0×16 PCIe 5.0×16 PCIe 5.0×16 Size 2 slots; 12.0×5.4 inches (304x137mm)2 slots; 12.0×5.4 inches (304x137mm)2 slots; 12.0×5.4 inches (304x137mm)2 slots; 9.5×4.4 inches (242x112mm)TGP/minimum recommended PSU (watts) 575W/1,000W360W/850W300W/750W250W/650WPower connector 600W Gen 5 or 4×8-pin w/ bundled adapter450W Gen 5 or 3×8-pin w/ bundled adapter300W Gen 5 or 2×8-pin w/ bundled adapter300W Gen 5 or 2×8-pin w/ bundled adapterDisplay support 4K/480Hz or 8K/120Hz (with DSC)4K/480Hz or 8K/120Hz (with DSC)4K/480Hz or 8K/120Hz (with DSC)4K/480Hz or 8K/120Hz (with DSC)Connectors 3x DP 2.1b, 1x HDMI 2.1b3x DP 2.1b, 1x HDMI 2.1b3x DP 2.1b, 1x HDMI 2.1b3x DP 2.1b, 1x HDMI 2.1bProcessor model GB202-400-A1GB203-400-A1GB203-200-A1GB205-300-A1Launch price $1,999$999$749$549Ship date Jan. 30Jan. 30FebruaryFebruary

What’s better about Blackwell?

While Nvidia has been putting AI-focused Tensor cores in its cards since the RTX 20 series (Turing), this year it’s kicked it up a notch because, well, AI is now The Thing and it’s for a lot more than just DLSS (Nvidia’s ML-based upscaling). Plus, the new fifth generation of Tensor cores bring new capabilities to DLSS over the 40-series’ fourth-gen, and in conjunction with the new fourth-gen RT cores, improvements to render quality. Specifically:

• DLSS 4: The new generation of the company’s upscaling and optimization algorithms adds multiframe generation to its bag of tricks, using AI to extrapolate three frames rather than the single frame in DLSS 3. Nvidia claims it improves performance by about 4x. (It does make me wonder when the frames in games are created more by AI than actual designers.) The increased bandwidth allows it to use more data to generate detail, improve antialiasing, reduce ghosting and improve stability. DLSS 4 also adds the next generation of its Reflex latency reduction, but that isn’t Blackwell-dependent.
• Neural shaders: Blackwell implements tiny neural networks to work with the veteran programmable shaders to improve the quality of lighting and textures.
• Neural faces: It uses basic rasterized faces combined with 3D pose data to create digital faces in motion in real time. An increase in the number of ray-traced polygons the GPU can handle — RTX Mega Geometry — thanks to Blackwell, allows for more detail on elements that usually get shortchanged because of their complexity. In this case, hair and skin.
• Nvidia ACE: The company has been talking about this for a while — the technology stack that enables interactive nonplayer characters that answer questions and otherwise contribute to the gameplay without preprogramming. It needs the AI power to handle models and perform in real time that Blackwell brings.
• Better gen AI and training performance: The thing about generative AI is you can do it on relatively low-power hardware, and while you want it reasonably fast, that throws another wrench in. Because AI’s not just about speed, it’s about accuracy, and accuracy is subject to the type of data the processing device can handle and how much memory you can store interim generations and models in. In other words, hardware can make a difference when you need more variations generated or five-fingered (rather than four or six) hands. That’s why using a discrete GPU matters, and why a powerful GPU matters more. Blackwell brings some of Nvidia’s highest performing hardware for consumer and prosumer AI.

There’s also the generally-faster-than-the-previous-generation angle thanks to more of everything and optimizations that occur from year to year. We won’t really know specifics until we get a chance to test the cards. But Nvidia makes the usual claims that “this year’s model performs like the higher-level of last year’s.”

Watch this: Everything Announced at Nvidia’s CES Event in 12 Minutes

11:47

What should I know before I buy a 50-series card?

The 50-series imposes the usual year-over-year growing pains. Namely, they can be bigger and require more power. For instance, Nvidia specs the 5090 and 5080 as dual slot, but I’ve seen several much fatter listings; for example, the Asus ROG Astral GeForce RTX 5080 OC Edition lists as requiring 3.8 slots, which really means four. And even the midrange RTX 5070 models may hit the triple-slot mark.

The revitalized Alienware Area-51 line will offer some configurations with the RTX 50-series GPUs.

Plus, make sure your power supply’s up for it. The RTX 5080 requires the same level as the RTX 4090; the 1,000-watt recommended PSU for the RTX 5090 may number among any records Nvidia has broken.

The good thing about the laptop chips is that you don’t have to worry about making sure anything fits or will run. On the other hand, do check to see how much power the manufacturer allows the GPU to pull. Because if you need the power of an RTX 5090, you don’t want a system that will throttle it back when you need it most; you might as well save money on a lower class of card.

Also, think about your motherboard and CPU (as always). The new cards support PCIe 5, and it’s possible that if you stick it into a last-gen motherboard paired with a year’s-old processor then one or both may bottleneck the performance.