NVIDIA Next-Gen RTX 5090 Leaked to be 1.7x Faster Than the RTX 4090, Packed With 192 SMs, 128MB of Cache & 2.9GHz Clock Speeds
A thread over at Chiphell has revealed a lot of premilimary information regarding NVIDIA’s next-gen Blackwell series. Allegedly, Blackwell is 1.7x faster than Lovelace whilst packing 50% more cores and 78% more cache.
RTX 5090, 50% More Cores & 1.7x Higher Performance
Please note that these are very early rumors, so take them with a grain of salt. Panzerlied, well-reputed Chiphell tipster claims that Blackwell will offer 50% more cores than Ada Lovelace. Since the RTX 4090 packs 128 SMs, that equates to 192 SMs for the RTX 5090 or 24576 FP32 cores.
Interestingly, the GTX 1080 Ti had 3584 FP32 units. Despite not scaling linearly due to architectural reasons, just 4 generations later, NVIDIA is now packing 7x more cores in their GPUs. The term ‘scale’ in the leak (via Google Translate) is interpreted as FP32 Units, as detailed by the comments.
Moreover, the memory bandwidth sees a 52% bump. This nets the RTX 5090 1,532 GB/s of effective bandwidth, still a bit lower than our initial expectations. Furthermore, the cache is uplifted by 78% for 128MB of L2 Cache. Hence, the RTX 5090 is rumored to have 16x more L2 cache than the RTX 3090, which is absolutely insane.
The frequency should go up by 15%, which does lead to a 2.9GHz early estimate. With time, as the architecture matures, Blackwell may break through the 3GHz barrier with ease.
All these spec-bumps add up to a 1.7x boost in performance, as claimed by the leaker. It is noteworthy that Kopite reports that there is no significant increase in the core count. Although, Blackwell will see drastic modifications in the unit (SM?) structure.
The data-center counterpart of Blackwell, AKA the GB100 series will shift to an MCM approach, akin to AMD. After all, even Jensen cannot surpass the physical retical limit so a multi-chip design should be a viable solution. The GeForce or client counterpart of Blackwell will come under the GB20X wildcard, with GB10X reserved for HPC Blackwell.
Release Date & Expected Performance Estimates
It is mentioned that Blackwell will tape out sometime during the end of 2023. By then, we can only guess the estimated performance through simulations and a few calculations. So, whatever you saw above was effectively a guess or a rumor and nothing conclusive.
Blackwell, in-line with NVIDIA’s own schedule should launch in early 2025. Therefore, we still have a lot more to uncover before we can make any final estimates. Stay tuned at Appuals for more articles like this.
Source: Chiphell
