In July this year, there were reports that NVIDIA's next-generation flagship GPU (RTX5090 or 5090Ti) may use a 512-bit memory bus, which is 32GB of VRAM. Today, it was revealed that NVIDIA aims to increase the overall performance of its next-generation "Blackwell" architecture GPUs by 1.7 times. The latest leaked specifications of the flagship RTX5090, codenamed GB202, have surfaced, showing NVIDIA's bold vision if it achieves its goals.
Visit the purchase page:
JD.com NVIDIA series product summary
According to a post on the Chiphell forum by well-known leaker Panzerlied, NVIDIA is aiming to increase the RTX5090's CUDA cores by 50%, bringing the total to 24,576. That’s more than 8 times more cores than the Turing-based RTX2080. Memory bandwidth will also be increased by 52% by using 32Gbps GDDR7, which is higher than the current RTX4090’s GDDR6X.
Clock speeds are expected to be increased by 15%, which is one of the biggest upgrades. This will boost the RTX 5090's frequency to 2.9GHz, easily exceeding 3GHz in gaming workloads. The highest frequency of the existing RTX4090 is 2.52GHz.
According to reports, in order to further improve performance, NVIDIA plans to triple the level 2 cache to 128MB, far exceeding the 96MB on AMD RX7900XTX. This additional cache is designed to take maximum advantage of the higher memory bandwidth brought by GDDR7.
If NVIDIA can achieve these ambitious goals, the RTX5090 will perform more than 70% better than the RTX4090 based on specs alone.
Comparison between NVIDIA RTX5090 and RTX4090:
The number of cores is 50% higher than that of RTX4090
Memory bandwidth increased by 52%
78% increase in L2 cache (possibly 128MB)
The frequency is increased by 15%, increasing the clock frequency to over 3.0GHz
Overall performance improved by about 70%
While it's not confirmed yet, leaked specs suggest that the new Blackwell-based GPU, codenamed "GB202," could push gaming performance to new heights. We'll see if NVIDIA can maintain a reasonable 3-socket form factor for such a powerful chip.