11/30/2022 at 9:28 am by Oliver Jäger – With GDDR6W, Samsung is presenting a new memory standard for graphics cards that will be able to catch up with the high-bandwidth HBM memory. Thanks to a new stacking process called “Fan-out Wafer-Level Packaging” (FOWLP), this new memory is said to deliver double the capacity and transfer rate of GDDR6 with the same size.
Memory expert Samsung is stepping up a gear when it comes to graphics memory innovations for GPUs. Only in July of this year did Samsung bring out a 24 Gbit/s fast GDDR6 memory, which, according to the manufacturer, is the fastest graphics DRAM in the industry to date. Now the company is going one step further and presenting the GDDR6W memory, in which Samsung promises twice the capacity and bandwidth of GDDR6 with the same size
GDDR6W competes with HBM thanks to new stacking
This is to ensure “fan-out wafer level packaging” (FOWLP for short), a new stacking process for the memory chips, while saving manufacturing time and costs. With the same package size, twice as many memory chips can be accommodated, which should double the capacity of the graphics DRAM from 16 Gbit to 32 Gbit, while at the same time the bandwidth and the connections (I/Os) have doubled from 32 to 64 bit. According to Samsung, the area required for storage has been reduced by 50 percent compared to previous models.
This is where the FOWLP process comes into play, with the help of which the memory chips are packed directly on the silicon wafer instead of on a PCB. At the same time, RDL (Re-Distribution Layer) technology is also used, which ensures a much finer wiring pattern. Without a PCB, the package thickness is said to be reduced by 36 percent (from 1.1 mm to 0.7 mm) while at the same time increasing heat dissipation.
Package comparison: GDDR6 versus GDDR6W. Source: Samsung
GDDR6W memory should even be able to operate at HBM level in terms of bandwidth. HBM2E achieves a system-level bandwidth of 1.6 TB/s, with a speed of 3.2 Gbit/s per pin. The GDDR6W is said to have a system-level bandwidth of 1.4 TB/s while delivering 22 Gb/s per pin on a 512-bit basis. This corresponds to only one eighth of HBM2E (4,096 bits).
Also interesting: 3 nm chips for Nvidia, Qualcomm, IBM and Baidu: Samsung wants to dethrone TSMC
According to Samsung, the GDDR6W memory was already going through the JEDEC standardization process in the second quarter of 2022. On the one hand, this new form of memory should shine in graphics cards and accelerators for the upper performance end, which are used for artificial intelligence and HPC applications. On the other hand, Samsung would also like to expand its GDDR6W memory to devices in a smaller form factor, which primarily includes notebooks or laptops.