New Nvidia GPUs Announced: The RTX 4090 and 4080

Goodbye, Ampere; hello, Ada Lovelace: Nvidia revealed its new generation of GeForce graphics cards at its annual Graphics Technology Conference, featuring the company’s biannual upgrade to the GPU’s architecture, this time named for the mathematician and programming pioneer.

Nvidia promises the flagship RTX 4090 can perform up to two times better on raster-heavy games and up to four times faster on fully ray-traced games than the former flagship RTX 3090 Ti, while the RTX 4080 should be up to three times faster than the RTX 3080 Ti. The RTX 4090 is slated to ship on Oct. 12 starting at $1,599, while the RTX 4080 is coming in November starting at $899; the latter will come in two versions, 16GB and 12GB.

In addition to Nvidia’s Founders Edition cards, versions using the GPUs will be available from Nvidia’s usual partners (except EVGA), including Asus, Gigabyte, MSI, PNY and Zotac. The RTX 3080, 3070 and 3060 will remain in the line, and I’d look for big discounts on them during the holiday shopping season. 

When the RTX 4090 ships, there will be some free proof-of-pretty-and-performance DLC: Portal with RTX and a new Ray Tracing: Overdrive Mode for Cyberpunk 2077. There will also be a home-built real-time-rendering app, Nvidia Racer RTX. 

The chips’ Ada Lovelace architecture — it’s not clear whether it will be referred to as “Ada,” which is shorter, or “Lovelace,” in keeping with tradition — introduces seven hardware upgrades, some of which are derived from its datacenter sibling, Hopper:

• Built on a 4nm process. Optimally, smaller processes make it possible to cram more silicon into the same space, with (usually) improved power and performance efficiencies.
• Third-generation ray-tracing cores double the throughput of calculations for how the rays interact with the polygons defining the image, which Nvidia says boost floating point operation speed by up to 2.8x.
• Fourth-generation Tensor cores, which drive Nvidia’s DLSS upscaling software, allows DLSS to jump from generating individual pixels to more efficient frame-based algorithms for its new DLSS 3. More than 35 games and apps are signed up to incorporate DLSS 3 starting in October. 
• Updated Streaming Multiprocessors deliver more than 2x the throughput of the Ampere SMs.
• Shader Execution Reordering optimizes the handoff of ray-traced data to the SMs for rasterizing, which the company says boosts shader performance by up to 2x and frame rates by up to 25%.
• Optical Flow Accelerator uses pixel motion data to DLSS in order to generate new frames; Nvidia says this improves performance in CPU-intensive games (such as complex simulation games, like Microsoft Flight Simulator). 
• Dual AV1 Encoders should improve on-the-fly video processing quality for streamers. OBS is expected to release a new version of its streaming software in October, and Discord is slated to support it later this year.

To go with the new processing capabilities, Nvidia introduced some new software algorithms to improve performance as well.

• Opacity Micro-Maps create and save masks in a world’s objects in advance, so they don’t need to be recreated.
• Displaced Micro-Meshes are meshes of smaller triangles that the new RT cores can process without needing to be stored or preprocessed somewhere else; they improve speed (up to 10x, per Nvidia) and require up to 20x less video memory in certain cases.