Throughout this week, there has been a notable increase in speculations surrounding the RTX 40 SUPER series, reaching a peak with comprehensive specifications revealed by a well-known NVIDIA insider, Kopite7kimi. The disclosed details encompass CUDA core specifics and potential GPU models for the range of RTX 40 SUPER cards that have been a focal point of discussion in recent weeks.
As per Kopite’s tweet, the RTX 4080 SUPER is anticipated to incorporate the complete AD103-300 GPU, which boasts an impressive 10,240 CUDA cores. Although the details concerning memory configuration are still undisclosed, it’s important to point out that the AD103 GPU suggests a 256-bit memory bus. Achieving 20GB specifications without reducing the memory bus to 160 bits seems unlikely, but Kopite has not confirmed the specifics of memory and the speed for this SUPER upgrade.
Specifications for the RTX 4070 Ti SUPER and RTX 4070 SUPER have also come to light. The RTX 4070 Ti SUPER, despite its name being somewhat perplexing with the combination of “Ti” and “SUPER,” may be prone to naming inaccuracies given the extended duration of this leaked information, spanning several weeks. Nevertheless, this specific model is expected to include 8,448 CUDA cores and the potential utilization of two GPU variants, either the AD103-275 or a notably reduced AD102-175.
As for the RTX 4070 SUPER, it signifies a minor tweak in specifications compared to the RTX 4070 Ti, featuring a reduced CUDA core count of 7,168. Similar to the previously mentioned model, this version might also offer two GPU options, either the AD104-350 or the AD103-175.
It’s important to note that these specifications haven’t received recent updates, leaving room for the possibility that some SKUs on this list may have been put on hold. NVIDIA has yet to reveal any new specifications to its board partners, and the likelihood of significant releases for the current year remains relatively low. Nevertheless, there are clear signs that an update is impending, possibly encompassing one to three models.