Demystifying Micron's GDDR6X Memory Chip
Updated: Dec 24, 2022
A tear down blog of the NVIDIA's GeForce RTX 4090 revealed 12 pieces of Micron GDDR6X memory chips (in red circles of Fig. 1)are used in a single GPU motherboard. By far, Nvidia is the only GPU supplier using this latest memory technology.
Fig.1 : GDDR6X chips on motherboard
During a press release back in April 2022, Micron announced the mass production of latest GDDR6X memory chips which can transfer data at up to 26Gb/s. There are a few key take-aways from the press release:
The use of PAM4 signal interface .
The performance improvement compared to GDDR6.
PAM4 Signal Interface
The secret sauce of doubling memory bandwidth in GDDR6X is to use a multi-level signaling in the form of four-level pulse amplitude modulation (PAM4). The innovative signal interface is developed between Nvidia and Micron together to achieve higher bandwidth in AI applications.
In a two-level signaling modulation scheme, the data bit 0 or 1 is represented either by a voltage LOW or voltage HI signal, so each voltage signals carries 1 bit information. By using a four-level voltage modulation as shown in Fig. 2, each voltage level can now represent 2 bits of data thus each clock cycle a 2-bit data can be transferred.
Fig.2 : PAM4 driver implementation
By consulting Table 1 provide by Micron, the major improvement is that GDDR6X has 30% higher bandwidth per pin compared with GDDR6 technology. Both GDDR6 and GDDR6X achieve higher power efficiency, higher data rate than the GDDR5 and GDDR5X technologies.
Table.1 : GDDR5-GDDR5X-GDDR6-GDDR6X Feature Comparison
The most significant limitation with GDDR6X technology is its operating temperature range cannot drop below 0C while GDDR6 can operate at -40C. This limitation means that GDD6RX current cannot be used in automotive applications even though automotive application is the future growth driver for Micron.