Nvidia rolled out its new Ampere-based GeForce RTX 3000-series graphics cards, with the top-of-the-line RTX 3090 that the company claims can handle 8K/60 frames-per-second gameplay -- using upscaling in what's called DLSS 8K mode. We also heard about the RTX 3080 and 3070, the usual better-faster replacements for their predecessors, the RTX 2080 and 2070. Essentially, the 30-generation of the GeForce cards is intended to provide the performance with ray tracing and DLSS that we expected when Nvidia first announced the RTX line. 

GeForce RTX 3090 cards are slated to ship on Sept. 24 starting at $1,499, while the RTX 3080 is coming on Sept. 17 starting at $699 and the RTX 3070 in October beginning at $499. 

The RTX 3090 is really more of a Titan RTX replacement, with HDMI 2.1 and AV1 decoding support, plus 24GB GDDR6X (G6X) and "a giant Ampere" processor. It's a three-slot card and almost all of it's cooling and heat dissipation. Games launching with DLSS 8K support include Control, Death Stranding, Minecraft with RTX Beta for Windows 10, Watch Dogs: Legion, and Wolfenstein: Youngblood.

The RTX 3080 and 3070 seem to deliver about a 30% improvement over the Super versions they replace, which is pretty good in gen-over-gen terms; Nvidia claims the 3070 is faster than the 2080 Ti at the price of the 2070 Super in 4K. 

For the GeForce Ampere cards, Nvidia claims the updated shader engine, RT core and Tensor cores deliver much better speed: 2.7x for shaders, 1.7x for RT and 2.7x for Tensor compared with the first generation. Nvidia showed it off with a pretty impressive sequel to the Marbles demo it showed off in the data-center Ampere announcement, running at 1440p/30fps with depth of field and 130 area lights. 

Before the big reveal, the company announced that Fortnite would be getting RTX support, including real-time shadows and ambient occlusion, and showed off a clip of Cyberpunk 2077 running the latest GPU. Nvidia's Broadcast software will now offer background blurring (automatically masking you) and audio noise removal.

You'll also be seeing its Reflex technology in an upcoming Game Ready driver; it works in conjunction with upcoming 360Hz monitors from Acer, Alienware, Asus and MSI to analyze the latency between mouse and keyboard to the display. You'll get extra info if you use one of the partner mice from Razer, Logitech or Asus, because complete system analysis requires some hardware or firmware support. It's not clear if it actually does anything other than provide an overlay with the timing so you can tell if you're getting better or if there's some bottleneck in the system. 

Nvidia launched its Ampere technology with the A100 data center processor. While the Ampere has a lot of new capabilities that aren't strictly relevant in a consumer graphics card, it does build on Turing in several ways that can impact gaming performance. 

Ampere architecture uses a 7nm process, like AMD's Navi-powered graphics, as compared with Nvidia's 12nm Turing predecessor. A smaller process node enables denser chips, which means you can either fit more capable processors in the same space as before or make smaller chips with the same capabilities as the previous chip.

Turing added Tensor Cores, which accelerate operations frequently used in ray tracing and AI deep learning and inferencing. Ampere's second-generation Tensor Cores and enhanced streaming multiprocessors incorporate a lot of new algorithms for accelerating mathematical operations, including mixed-precision and larger (64-bit) floating-point operations. 

FLOPS are big for real-time rendering performance -- that's gameplay rendering and video editing, among other things -- and the increased efficiency should speed up the real-time ray tracing to what we expected it to be when Nvidia first announced the RTX line. And AI has become important in making game interaction and rendering more procedural and less predetermined, and includes Nvidia's DLSS AI-based upscaling technology.

The Ampere-driven architecture also enables support for PCIe Gen 4, with its much higher throughput, in this case between the GPU and the CPU and main memory. It can also now perform asynchronous copy operations, which allows data to flow between the GPU and main memory without needing to use the CPU as a middleman; it adds support for more and faster video memory as well. With the new generation of its CUDA (11) programming interface, it's much more efficient at launching operations. 

So, faster.

NVLink, Nvidia's technology for connecting multiple GPUs in a system, also enters its third generation. The company claims an almost doubling of throughput between GPUs. Not so important for gaming, unless more than a handful of games decide to support it, but important for professional real-time rendering and high-res video editing.