What’s Important:

• DLSS 4 introduces Multi Frame Generation, exclusive to RTX 50 Series GPUs, increasing frame rates up to 8x.
• A new transformer-based AI model replaces the convolutional neural network, improving stability, motion clarity, and reducing ghosting.
• Enhanced Ray Reconstruction is compatible with all RTX GPUs, improving ray-traced lighting and reflection quality.
• Broad game and application support, with 75 titles confirmed at launch.
• Project DIGITS is a personal AI supercomputer starting at $3,000, delivering up to 1 petaflop of AI performance.
• Equipped with the GB10 Grace Blackwell Superchip and 128GB of unified memory.
• Supports up to 200 billion parameter AI models and scalable to handle 405 billion parameters when linking two systems.
• Includes access to NVIDIA’s AI software stack and supports popular frameworks like PyTorch and Python.

The RTX 50 is an interesting card, not just because of the Blackwell silicon improvements on the TSMC 4NP, but because we’re seeing the limitations of silicon improvements being fixed with AI.  In terms of silicon improvements, it seems like generation over generation, it’s around 20-40% depending on the card and workload. That’s pretty good, especially as the prices have help improve the overall value of the cards.

Of the 4 RTX 50 series cards that Nvidia has announced, 3 of them are cheaper than the earlier generations. The RTX 5070 starts at $549, compared to the RTX 4070 $599. RTX 5070 Ti is $749, vs RTX 4070 Ti $799. The RTX 5080 is now $999, down from $1,199 from the RTX 4080. The RTX 5090 now starts at $1,999, up from the $1,599 of the RTX 4090. I think these are fair prices for Nvidia to charge. The overall market for GPUs tends to go towards the lower end, with the Nvidia __50, __60, __70 series being the most common video cards in the Steam PC parts survey, with most being 30 and 40 generations.

The best part of these new cards is the net effective performance on the cheaper ones. The RTX 5070 now matches the effective and perceived performance of the RTX 4090, thanks to DLSS 4.

This trend is bringing AI outside of just LLMs, AI agents, and the alike into a practical application of AI in gaming: using a transformer model to generate the vast majority of the pixels in a scene, increasing fidelity and net performance while keeping the prices lower. The raw performance may not be equal, but the net perceived performance is. I think this is an important distinction, because if it looks the same or better (I saw a demo of Marvel Rivals with the RTX 5070 and RTX 4090 side by side, and it does) then it shouldn’t matter if the actual performance is equal.

In terms of actual performance in the silicon, Nvidia has new FP4 datatype accelerators, allowing for native FP4 performance. This means nearly the same quality of inference output in some of the examples, Nvidia showed me a Flux.1 Dev demo with a native FP4 model, and said the quality was nearly identical to the FP8 native model while using around half the memory. This push from Nvidia for FP4 data type expands into the data center as well, and the Project DIGITS supercomputer. It’s more or less the stable of the Blackwell architecture.

Given the vast majority of cards in the Steam PC Spec survey are currently on RTX 30 and 40 series, I think these could be compelling upgrades for quite a lot of gamers. I expect these cards to be incredibly supply restrained, especially since it the TSMC 4NP node, which is the same as the Nvidia B200 data center GPU. With how much demand Nvidia has for the B200 in data center, I’d imagine most of the wafers will be dedicated to the B200 rather than RTX 50 series. It might be a while, and difficult, until gamers are able to get a Blackwell RTX GPU. Gamers are historically very loyal to Nvidia, so even if it takes 6-9 months to get a card, I expect most would rather wait and keep trying than switch to a card from AMD or Intel.

The other announcement, that I frankly think is bigger than the RTX 50 series, is Project DIGITS with the new GB10 chip. It’s a mini-super computer from Nvidia, with a 20-core Arm CPU and Blackwell GPU, connected with NVLink C2C (chip to chip). This was made in partnership with Mediatek, and has 10 Arm Cortex-X925 cores and 10 Arm Cortex-A725 cores, 128GB of Unified Memory, and 4TB of storage. The GB10 looks like a great supercomputer for AI developer and those looking to play around with Nvidia’s tensor acceleration and CUDA architecture.

It launches in May and starts at $3,000, and seems to hint at Nvidia’s broader goal to enter the PC market. There have been rumors for ages of Nvidia working on a PC chip for Windows, and while this is not that chip, it seems like the sort of architecture and what Nvidia will be working on for that market in the future. It could bring some high end competition with Intel, AMD, and Qualcomm to the Windows CPU market and Arm entry into the PC gaming market.

There were a few more launches from Nvidia including the new Thor automotive platform for AI and autonomy, as well as demos of a few robots and partners towards world models and autonomy as well as synthetic data generation for these robots, Nvidia is working on quite a lot towards getting into a more intelligent future.

Join the newsletter and stay up to date