The first graphics processing units (GPUs) that use GDDR7 memory might continue using 16 Gbit chips, although there's a possibility for 24 Gbit chips as well. An article by 3DCenter highlighted that GDDR7 might offer chips with 24 Gbit and 48 Gbit densities, providing more memory options for GPUs. For example, a 24 Gbit chip could allow for 18 GB of memory over a 192-bit bus or 24 GB over a 256-bit bus. This flexibility would enable GPU makers to balance memory size and board costs effectively, especially important for keeping high-speed memory like GDDR7 affordable. However, higher densities like 32 Gbit or 64 Gbit might be too costly for consumer graphics cards and reserved for professional or high-end models.

The first gaming graphics cards (GPUs) that use the new GDDR7 memory might continue using 16 Gbit (2 GB) memory chips, similar to what we see in today's RTX 40-series cards. This info comes from kopite7kimi, who is known for accurate NVIDIA GeForce leaks. Typically, a 16 Gbit chip allows a graphics card with a 256-bit memory bus to have 16 GB of video memory. The top-end model in the current lineup, the RTX 4090, achieves its 24 GB of memory by using twelve of these chips. The leak suggests that NVIDIA might not apply the GDDR7 standard across all its products. Instead, like with previous series, it might work with a memory company to develop a specialized standard that's better suited for its graphics cards. This was seen when NVIDIA and Micron Technology co-developed the GDDR6X. The first GDDR7 chips, expected at the end of 2024 or start of 2025, could reach speeds up to 32 Gbps. A second wave of GDDR7 chips might push this to 36 Gbps by late 2025 or 2026, following the pattern seen with GDDR6's evolution.

GDDR7 is also significant because it supports "odd" memory capacities, like 3 GB chips, not just the usual powers of two (e.g., 2 GB, 4 GB). This flexibility could lead to graphics cards with more tailored memory amounts, fitting better with the graphics chip's capabilities and avoiding the need for costly memory doubling.

Sources: kopite7kimi (Twitter), 3DCenter.org