Samsung introduced its latest memory solution last month and said that 8.5Gbps LPDDR5X and GDDR7 memory chips are on their way.
Now, Dr. Ian Cutress has introduced some of the technical specifications of the future GDDR7 memory. The next generation of GDDR7 memory will reportedly have twice the bandwidth and density of current GDDR6.
First of all, GDDR7 memory will use PAM3 signaling, with a rate of up to 36 Gbps. The traditional GDDR6 uses NRZ (not to zero) or PAM2 signaling, with data rates of 14 to 24 Gbps. It is worth mentioning that Intel’s previously announced 80G Thunderbolt interface and USB interface will also use PAM3 signaling.
Although there are currently some 18 Gbps, 20 Gbps and 22 Gbps GDDR6 chips, they did not make it into production during the development phase of the RTX 30 series “Ampere” GPUs, so NVIDIA and Micron have worked together to develop a GDDR6X standard based on PAM4 signaling, with speeds ranging from 18 Gbps to 23 Gbps.
According to the report, the traditional NRZ signaling can provide 1bit per cycle, while PAM4 can achieve 2bit per cycle, while PAM3 uses a more advanced waveform, including more “eyes” (the gap created by the crossing of waves), thus achieving to 3bit per cycle.
Samsung said that PAM3 is 25% more efficient than NRZ signaling, and GDDR7 will be 25% more energy efficient.
Samsung says that the new generation of GDDR7 memory can bring 36 Gbps data rate, which is double that of GDDR6, but of course, GDDR6 also has the same improvement for GDDR5.
A GPU with a 256-bit memory bus will get 1152 GB/s of memory bandwidth when using 36 Gbps of GDDR7 memory, while a high-end GPU with a 384-bit interface can reach 1728 GB/s and a mainstream 128-bit interface only 576 GB/s.
