Nvidia is rumored to increase the number of CUDA cores in the RTX 5090 by 50%

Recently, a discerning netizen on ChipHell unveiled some details pertaining to NVIDIA’s forthcoming flagship, the GeForce RTX 5090. Indications suggest that compared to its predecessor, it will boast a 50% surge in CUDA cores, a 52% enhancement in memory bandwidth, a substantial 78% uptick in cache (presumably referring to the L2 cache), coupled with a frequency increment of 15%. In essence, this amalgamation of advancements propels the RTX 5090 to perform at an astounding 1.7 times the capability of the GeForce RTX 4090.

This extrapolation implies that the GeForce RTX 5090 will encompass 192 SM units, equating to 24,567 CUDA cores, with frequencies hovering around a robust 2.9 GHz. It’s poised to feature an L2 cache of 128MB, and the memory width might be discerned at 512 bits, armed with a GDDR6X memory operating at a swift 24 Gbps, rendering an approximate memory bandwidth in the vicinity of 1.53 TB/s. On a related note, Samsung has pioneered the development of the industry’s inaugural GDDR7 chip, with each data I/O interface operating at a rate of 32Gbps. The premier 16Gb GDDR7 chip, given a width of 384 bits, proffers a staggering bandwidth of upwards of 1.536 TB/s. One speculates that NVIDIA might pivot towards integrating GDDR7.

According to NVIDIA’s strategic timeline, the unveiling of the GeForce RTX 50 series GPUs isn’t slated until 2025, rendering any current conjecture seemingly premature. The aforementioned netizen had previously posited that GPUs based on the Blackwell architecture would encompass five chip models: GB202, GB203, GB205, GB206, and GB207. A notable departure from convention is the omission of the x04 chip, in this case, GB204, which has been supplanted by the x05 model, GB205.

Recent reports intimate that data center products founded on the Blackwell architecture will transition to a more compact chip design, embracing MCM multi-chip packaging. However, consumer-centric products appear to remain steadfast in their commitment to a singular chip design, which offers the dual advantage of temporal predictability and minimized risk, ensuring timely product fruition. Rumors are rife that NVIDIA intends to instigate sweeping modifications to the Blackwell architecture. While it isn’t a complete overhaul, there will be a novel structural approach to SM and CUDA. There’s speculation that RT units might be superseded by PT units, and ray tracing capabilities are set to undergo further refinement and fortification.