Nvidia officially announced its 4000 series of GPUs, based on the Lovelace architecture. The RTX 4090 arrives October 12th for $1,599. The RTX 4080 comes later in November for $899.
The 4090 ships with 24GB of GDDR6X memory, enough to claim it is 2 to 4 times faster than the 3090 Ti with the same power consumption. Nvidia recommends a PC Power Supply of at least 850 Watts when using a Ryzen 5900X class processor.
The RTX 4080 will come in 12 GB and 16 GB GDDR6X options for $899 and $1199 respectively. All three cards include updated ShadowPlay support, that can capture 8K video at 60 fps in HDR. It also supports hardware AV1 encoding.
Both cards in the 4000 series will support PCIe Gen-5 16-pin connectors without the need for a custom solution required in the previous generation. The cards will also include an adapter to connect with three standard 8-pin power connectors. Power supplies are coming in October from Asus, Cooler Master, FSP, Gigabyte, iBuyPower, MSI, and Thermaltake. Expect to see RTX 30-series still available for awhile as Nvidia says it made too many. The 4090 and the 16GB 4080 will also come as Founders Editions.
Nvidia also announced a processor for autonomous vehicles called Drive Thor, based on Nvidia's Hopper GPU platform that's optimized for processing algorithms at 2 quadrillion operations per second. That's 8 times Nvidia's Orin processor. With 77 billion transistors, Hopper can replace multiple chips, saving on expense and power consumption. Nvidia said it uses CPU cores from Nvidia's Grace processor and borrows some elements from the Lovelace architecture.
Thor will be able to run Linux, QNX, and Android simultaneously to serve different parts of the car. Drive Thor will also have lower-end versions meant for driver-assistance systems that don't need all the processing power that fully autonomous systems need. It will ship in 2024 and show up in cars in 2025, starting with China's Zeekr 001 EV.
And Nvidia also gave us all a look at DLSS3, the next version of its "Deep Learning Super Sampling" technology that can upscale graphics and supposedly quadruple performance over native resolution. It's an algorithm that can add bits to either increase frame rate at the same resolution or upscale the resolution without losing performance. So a game could run at 1080p but DLSS can use machine learning to make it look 4K.
DLSS3 can generate entire new frames using its Optical Flow Accelerator to track and calculate on-screen object motion vector, not just pixels. That should reduce stutter. And it will work with Nvidia Reflex technology to reduce latency and improve responsiveness.
In a demo they boosted Cyberpunk 2077 from less than 30 fps to around 100. DLSS2 could only get that to 60.
Only the new RTX 4000 series cards will support DLSS3 because it needs the new 4th-gen tensor cores and Optical Flow Accelerator. But more than 35 games are integrating support for DLSS3, with some launching as early as October.
Roger Chang contributed to this report.