RTX 50: GDDR7 Delivers Bandwidth Boost, But Not Maximum

Recent discourse surrounding the Blackwell architecture for gaming GPUs appears to be burgeoning, with rumors suggesting that the GeForce RTX 50 series will employ GDDR7 memory modules in 16Gb (2GB) configurations, with potential support for 24Gb (3GB) modules. Furthermore, it is speculated that the GB20x series will maintain the same memory interface configuration as the AD10x series, with flagship graphics cards continuing to feature a maximum of 384-bit width, consistent with existing products.

An online disclosure revealed that although the initial GDDR7 modules reached speeds of 32 Gbps, the GDDR7 employed by the GeForce RTX 50 series will operate at 28 Gbps, resulting in marginally slower speeds. When the memory bus width remains unchanged, 28 Gbps GDDR7 will offer a bandwidth that is 55% higher than that of 18 Gbps GDDR6, and 33% greater than 21 Gbps GDDR6X.

The decision to launch products with a lower memory speed is not unprecedented. Nvidia’s first foray into GDDR6 graphics cards was with the GeForce RTX 20 series, and while memory manufacturers offered modules capable of 16 Gbps, Nvidia opted to set the speed at 14 Gbps.

Last year, Samsung announced the completion of development work on the industry’s first GDDR7 chips. Unlike the existing GDDR6’s NRZ/PAM2 or GDDR6X’s PAM4 signaling mechanisms, GDDR7 employs a PAM3 signaling mechanism. NRZ/PAM2 transmits 1 bit of data per cycle, PAM4 transmits 2 bits per cycle, while PAM3 achieves a data transmission of 3 bits every two cycles.

Additionally, it is rumored that the flagship chip within the GB20x series, the GB202, may still offer a 512-bit bus width, suggesting that graphics cards will possess larger memory capacities. With GDDR7 operating at a rate of 28 Gbps, the memory bandwidth could exceed that of the AD102 by nearly 80%.