So far, we don’t know much about Nvidia’s next-generation GPUs other than they will be code-named Blackwell. We assume they will be made using a 3nm process from either Samsung or TSMC, with the latter being the more likely candidate, and that they’ll probably feature GDDR7 memory and arrive in 2025. Now, a reliable leaker from previous Nvidia launches has filled in more details, posting what they think are tentative spec targets for the flagship RTX 5090 SKU. Since the cards won’t launch for another year, some salt is required here, but the details sound reasonable.
The source of this information is a Chiphell forum member named Panzerlied, who has previously successfully predicted Nvidia hardware details, indicating they might be an Nvidia mole. This user made a post on the forum detailing the percentage increases Nvidia is targeting for the RTX 5090, giving us the first hints about Nvidia’s plans. Overall, the company reportedly expects a 1.7x uplift, so a 70% boost in total performance. That level of performance will be achieved with boosted specs across the board for cores, cache, and clocks.
It’ll be interesting to see if Nvidia can keep GPU size down to a reasonable 3 slots with Blackwell, or if it’ll go even bigger.
According to the post, flagged by Videocardz, Nvidia is planning on a 50% boost in core count, bringing it to 24,576 CUDA cores. For context, the Turing-based RTX 2080 had just 2,944 cores, which seems laughable in comparison. Additionally, Nvidia is projecting a 52% increase in memory bandwidth, which could be accomplished by upgrading to 32Gb/s GDDR7 memory compared with the current 21Gb/s GDDR6X chips on the RTX 4090.
As far as clock speeds go, a 15% boost is expected, putting them at 2.9GHz for the maximum boost clock. With Nvidia’s GPU Boost technology, that number could easily hit 3GHz in gaming, making it the first “stock” GPU to cross that clock speed Rubicon without additional overclocking. The RTX 4090 currently offers a 2,520MHz clock speed in stock trim.
Finally, Nvidia will triple down on adding L2 cache to its Blackwell flagship to help boost memory bandwidth, similar to AMD’s Infinity Cache. The red team offers 96MB of that cache on the Radeon RX 7900 XTX, and Nvidia plans to go beyond that by increasing L2 by 78%, resulting in 128MB total.
Ultimately, whatever Nvidia does with the RTX 5090 could make it the de facto market leader. AMD is rumored to be sacrificing its high-end Radeon 8000 GPU dies for high-margin AI accelerators. However, Nvidia might take that route, too, depending on where the market is a year or so from now. We will have to wait and see if the AI market goes the way of the metaverse and dies down after a year or two, but all indications are that won’t happen (famous last words!).