Samsung announced GDDR6W memory at the end of last month, saying it doubled its bandwidth and capacity. Now, Samsung China has released an article with technical information about the new memory.
Samsung says that GDDR6W is based on Samsung’s GDDR6 product and introduces fan-out wafer-level packaging (FOWLP) technology to greatly increase memory bandwidth and capacity.
The diagram below illustrates how twice as many memory chips can fit in the same size package, and how manufacturers can develop graphics cards and notebooks with twice the bandwidth and graphics DRAM capacity without changing the actual size of the GPU. In other words, the same amount of memory can take up 50% less space than in previous designs.
GDDR6 vs. GDDR6W package comparison
FOWLP technology is described as stacking the memory die directly on the silicon wafer instead of on the PCB. To achieve this approach, Samsung uses Redistribution Layer (RDL) technology to form the expansion pattern. In addition, the thickness of the package is reduced and heat dissipation is improved because no PCB is involved.
The GDDR6W with FOWLP is 0.7mm thick, which is 36% thinner than the previous generation’s 1.1mm package thickness. In addition, the thermal properties and performance of the chip remain the same as the existing GDDR6, despite being split into several layers. Unlike GDDR6, the bandwidth of GDDR6W with FOWLP is doubled due to the larger I/O in a single package.
In terms of performance, the newly developed GDDR6W technology can support HBM-level bandwidth at the system level. HBM2E provides up to 1.6TB/s of system-level bandwidth (based on 4K system-level I/O) and a transfer rate of 3.2Gpbs per pin. The GDDR6W provides up to 1.4TB/s of bandwidth (based on 512 system-level I/O) and up to 22Gpbs per pin.
Samsung said it is currently moving forward with the standardization of GDDR6W products. Samsung also announced that it will work with GPU partners to expand GDDR6W applications to small package-size devices such as notebooks and new high-performance gas pedals for applications such as artificial intelligence (AI) and high-performance computing (HPC).
