In July it was reported that Nvidia’s next GPU (RTX 5090 or 5090 Ti) will likely have a 512-bit memory bus, meaning 32 GB of VRAM. Today, it was revealed that Nvidia is targeting a 1.7x overall performance improvement for the next generation of its “Blackwell” architecture GPUs. New leaked specs of the RTX 5090 flagship, codenamed GB202, have surfaced, showing Nvidia’s bold targets if met.

According to a post by noted leaker Panzerlied on the Chiphell forums, Nvidia is aiming for a 50% increase in CUDA cores for the RTX 5090, bringing the total to a massive 24,576. That’s more than 8 times the Turing-based RTX 2080 cores. Memory bandwidth should also get a 52% boost by using 32Gbps GDDR7, compared to GDDR6X in the current RTX 4090.

Clock speeds could see one of the biggest generational jumps with an expected 15% increase. This will see the RTX 5090 boost up to 2.9GHz, easily exceeding 3GHz during workloads. The existing RTX 4090 goes up to 2.52GHz.

To further boost performance, Nvidia is reportedly planning to triple the L2 cache to 128MB – well above AMD’s 96MB in the RX 7900 XTX. This additional cache is designed to maximize the bandwidth of the higher memory powered by GDDR7.

If Nvidia can meet these ambitious targets, the RTX 5090 will deliver 70% more performance than the RTX 4090 based on specs alone.

NVIDIA RTX 5090 vs. RTX 4090

• 50% more cores than RTX 4090
• 52% increase in memory bandwidth
• 78% larger L2 cache (probably 128MB)
• 15% higher frequencies, pushing boost clocks above 3.0GHz
• Overall performance is about 70% higher.

Although not confirmed, the leaked specs so far point to a new Blackwell-based GPU codenamed “GB202” that could push gaming performance to new heights. It remains to be seen if Nvidia can maintain a reasonable 3-slot build for such a powerful chip.

source: Chipel Forum (Chinese) and VideoCardz