NVIDIA’s next-gen Blackwell GPUs are planned for a 2025 launch. With a slew of next-generation technologies, they will significantly improve ray-tracing and rasterization performance. The RTX 50 series GPUs will feature GDDR7 memory and a bus width of up to 512-bit. Thanks to these, the RTX 5090 will offer exceptional memory bandwidth, equaling, if not straight up, beating the pricier HBM memory solutions.
This implies the use of a chiplet architecture, probably disaggregated memory controller dies similar to the Radeon RX 7900 series GPUs. The graphics processing core should remain monolithic as an MCM design is still not feasible for low-latency rendering workloads.
In other news, NVIDIA won’t launch the GeForce RTX 4090 Ti after all. According to the well-reputed @kopite7kimi, the AD102-based GPU has been scrapped for good. This follows the cancellation of the RTX Titan, which was also going to be a behemoth of a card with a TBP of 600W. Unlike the RTX 4090 Ti, the Titan was supposed to leverage the full-fledged AD102 die with 18,432 shaders, 576 Tensor and 144 RT Cores, and 96MB of L2 cache.
The RTX 4090 Ti would feature 18,176 shaders, 568 Tensors, and 142 RT Cores. While the Titan would be paired with 48GB of GDDR6X memory, the 4090 Ti would feature just 24GB.
On the bright side, we may get a mid-season “Super” refresh based on cut-down AD103 and AD106 dies. A beefier RTX 4060 and 4070 are just what the 40 series family needs. They offer the least value with minimal generation uplifts from a price-performance perspective.