Renowned leaker @kopite7kimi believes Nvidia will launch the upcoming Blackwell RTX 50-series GPUs this fall, starting with the RTX 5080 rather than the flagship 5090. As with any unofficial rumor, sprinkle liberally with salt. Still, it could make plenty of sense, depending on the specifications and configurations Nvidia uses for what will inevitably be some of the best graphics cards when Blackwell arrives.
A bit of historical review is in order to set the stage. Two years ago, Nvidia launched it's RTX 40-series Ada Lovelace GPUs by starting with the top RTX 4090 on October 12, 2022. It was a tour de force for the green team, with a massive leap in generational performance — and only a minor $100 increase in price relative to the prior generation RTX 3090. The step down RTX 4080 arrived on November 16, 2022, one month later. The RTX 4070 Ti (after a rebranding) launched on January 5, 2023, with the rest of the GPUs following over the coming months. But while that's what Nvidia did with the RTX 40-series, it's by no means the precedent.
The RTX 30-series Ampere GPUs for example began with the RTX 3080 on September 17, 2020, with the top-tier RTX 3090 coming one week later. RTX 3070 arrived on October 29, 2020, with the RTX 3060 Ti launching on December 2, 2020. So, everything was basically in order from the top down, except the first two GPUs were swapped. The RTX 20-series Turing GPUs also had the RTX 2080 launch one week ahead of the top-tier RTX 2080 Ti, while earlier generations often didn't have the "ultimate" GPU come out until nearly a year after the x80-class card — GTX 1080 Ti launched on March 5, 2017 while the GTX 1080 came out on May 27, 2016.
Kopite7kimi tends to have accurate information about Nvidia's plans, and was making a comment on a different rumor that claimed Nvidia would only release the flagship RTX 5090 this year, with the less expensive parts arriving in 2025. But Kopite7kimi suggests that Nvidia's RTX 5080 is a 2024 product, and that it will come out before the top-of-the-line RTX 5090. That agrees with earlier rumors that indicated both GPUs would launch in 2024.
The range-topping RTX 5090 is projected to be based on the GB202 GPU, while the RTX 5080 will use the GB203 GPU. Both of those are merely codenames, and there's no clear indication of what other specifications might exist. Other rumors say that Nvidia decided not to create a GB204 chip, but instead it will have GB205 this time around. But here's where things get a bit interesting.
Nvidia's next-generation Blackwell GB200 data center and AI GPUs have been announced, and there are some big changes in store. The GB200 processors will be made on TSMC's custom 4NP manufacturing technology, a refined version of the current 4N fabrication process that used for Hopper H100/H200 and Ada Lovelace RTX 50-series GPUs. As a refinement of an existing node, it won't enable significantly higher transistor density, nor will it provide a full node performance uplift. To create a higher performance flagship solution, Nvidia opted to link two full reticle size chips together via a new NV-HBI link (Nvidia High Bandwidth Interface) that provides 10 TB/s of throughput.
Given what we know of the data center Blackwell solutions, it's entirely possible that Nvidia could take a similar approach with the top consumer and professional GPUs. Specifically, GB202 could be two GB203 chips linked via NV-HBI, doubling the potential core counts, memory capacity, and memory bandwidth — without any of the messiness of previous dual-GPU solutions. Note also that rumors of a 512-bit memory interface for RTX 5090 have popped up, with a large gulf between it and the rumored 256-bit RTX 5080.
This is speculation on our part, but if the RTX 5090 does indeed use a dual-die solution like GB200, having RTX 5080 launch first would be completely reasonable. In fact, it would almost be foolhardy to try to push out the dual-die cards before the single-die solution.
Either way, we have credible rumors that Nvidia plans to launch Blackwell consumer GPUs this year. With a smaller die on GB203 and the possibility of linking two chips together for GB202, Nvidia could be trying to kill two birds with one stone — or chip. Smaller chips are easier to debug and manufacture, and having a halo product that doubles the specifications in every area relative to the penultimate GPU would certainly attract some deep pockets. We'll find out more about the RTX 5080, RTX 5090, and other Blackwell GPUs later this year, and it seems likely we'll see at least two of the GPUs launch in 2024.
