When NVIDIA launched its RTX 4090, there was nothing that could compete against it. A year later, we are in the same situation, as AMD at best rivals the RTX 4080 and Intel doesn’t even surpass the mid-range. NVIDIA will continue to be untouchable if this continues, as now more details are revealed about the RTX 50 and how these will be the first MCM GPU (GB200).
It has been just over a year since NVIDIA launched its RTX 4090, so we expect the next generation to be launched in 2024. Here, we would be talking about the first GPUs arriving for data centers and servers, followed by an RTX 5090 since the company always starts with the highest range and then moves down. We discussed the RTX 5090 a few days ago, as it would use a GB202 chip, with a 384-bit bus and GDDR7 memory.
NVIDIA will launch an RTX 50 in MCM format (GB200) to beat AMD RDNA 5
Although we do not know NVIDIA’s official specifications, we knew the obtained bandwidth, which is considerably higher than the current generation. With 384-bits using GDDR7 running at 32 Gbps, we have 1,536 GB/s, 52% more than the RTX 4090. Now, RedGamingTech offers updated specifications, showing the 384-bit memory remains, but GDDR7 will be used at 36 Gbps, providing an impressive 1,728 GB/s.
We also saw earlier that this GPU could have 204 SMs, giving a total of 26,112 CUDA Cores. However, the YouTuber states that it will use 192 SMs, which equates to 24,576 CUDA Cores. In terms of VRAM, a total of 36 GB will be used with 24 GB modules. It would have a 96 MB L2 cache and use TSMC’s 3 nm N3E node. Another interesting novelty mentioned is that NVIDIA will start by launching two top-of-the-line GPUs for Blackwell.
The GB202 (RTX 5090) is slated to arrive late next year. The problem is that these GPUs would be slower in rasterization than the future AMD RDNA 5, which would be released 9-12 months later. To address this, two options have emerged. The first is NVIDIA’s decision to launch a monolithic GB201 with more SMs to beat RDNA 5. The second option would be to launch the GB200, the company’s first gaming GPU with an MCM design.
The future GB200 could have 384 SMs
According to RedGamingTech, if NVIDIA chooses to create this GB200 in MCM format, it could have a total of 384 SMs (49,152 CUDA Cores), making it a true powerhouse. However, it considers that not all SMs would be activated if such a GPU were created, reducing the number of cores. We are talking about a GPU consisting of two RTX 5090s on a single graphics card, thus doubling the number of cores. It is more likely that NVIDIA will launch another monolithic GPU, although the probability of seeing RTX 50s with an MCM design remains.
Lastly, the YouTuber shows us novelties of the Blackwell architecture, which will bring the next generation of GPUs to life. These GPUs will have a unified L0 and L1 cache and a new bus connected directly to the L2 cache. Blackwell promises high performance, as each SM delivers twice the computation performance compared to Lovelace. It is surprising to see GDDR7 memory used in non-top-of-the-line GPUs, but this is due to a smaller bus. Future entry-level RTX 50s will use 64/96 bits, providing 256/384 GB/s of bandwidth. Finally, Frame Gen latency will be reduced, and path tracing reconstruction will be improved using denoisers.
The article “NVIDIA is not worried about AMD RDNA 4 GPUs, but RDNA 5: RTX 50 (GB200) MCM design and 49,152 Shaders in 2025!” first appeared on El Chapuzas Informático.