The NVIDIA Quadro RTX GPUs feature Samsung's 16 Gigabit GDDR6 memory
Samsung has announced that the NVIDIA Quadro RTX GPUs use its 16 Gigabit GDDR6 memory. The successor to GDDR5X offers improved performance and energy efficiency for demanding graphics applications as well for VR/AR and AI. The Quadro RTX-series is the first to implement Samsung's GDDR6 solution.
Samsung's GDDR6 was supposed to launch in 2017 itself but got pushed to 2018; Samsung was in no mood to sell GDDR5X memory at that time. Details then started to emerge during end-2017 that NVIDIA Volta would be using Samsung's 16 Gb GDDR6 VRAM. Samsung started mass production of 16 Gb GDDR6 in January 2018. Although the chips have a pin speed of 18 Gbps and data transfer rate of 72 GBps, the implementation in NVIDIA's Quadro RTX-series performs at a 14 Gbps pin speed with a 56 GBps data transfer rate. It is still a 75% increase over the 8 Gbps pin speed of GDDR5. Samsung's GDDR6 memory is highly power efficient and the company says that it requires 35% less power than GDDR5. The new memory operates at 1.35V against the 1.55V consumed by GDDR5.
Other vendors such as Micron and SK Hynix are also heavily invested in manufacturing GDDR6 memory. Micron wants GDDR6 to be implemented not just in GPUs but also in diverse applications such as AI neural nets and networking devices. Micron's GDDR6 is slower than that of Samsung with a 14 Gbps pin speed. SK Hynix is offering both 10 Gbps/12 Gbps and 12 Gbps/14 Gbps GDDR6 RAM that come in two voltage variants of 1.25V and 1.35V respectively.
Samsung's 14 Gbps GDDR6 seems to be a good choice for the Quadro RTX. Perhaps, the best benefits of the new Turing GPUs and the high speed GDDR6 memory can be realized by combining two of the Quadro RTX 8000s via NVLink that effectively doubles the memory buffer up to 96 GB.