GDDR6X VRAM: Specifications and Technical Features

It surprised locals and strangers when 4 months ago we talked about a new version of GDDR6 that not even the JEDEC registered. And it is that this new type of VRAM has been designed exclusively for the RTX 3080 and RTX 3090 by NVIDIA, so it is outside of any official standard and at the same time allows greater performance, but what news do these bring? GDDR6X VRAM?

The truth is that with these GDDR6X VRAMs, Micron scores much of the marketing in terms of being the first company that is capable of offering a discrete graphics memory that exceeds the TB / s speed, although it is not really demonstrable for now .

GDDR6X VRAM

And it is that this technology has been coming for 14 years, where at that time new ways to increase the speed in VRAM were already being studied.

Micron VRAM GDDR6X, more speed and in high production volume

The biggest secret of Micron’s GDDR6X is PAM 4 or 4-level Pulse Width Modulation. A technology in constant development and whose culmination is reaching up to 21 Gbps in this type of dedicated memory.

When Micron engineers managed to work with PAM 4 they accessed a new level of complexity, because a technology is useless if you can’t cut costs and manufacture it in volume. Manufacturing GDDR6X VRAM with PAM 4 is not a problem for Micron, but it was a problem to do it in the sufficient volume that NVIDIA required.

GDDR6X Micron

But what makes GDDR6X so novel outside of the speed achieved? and above all, how do you do it? There are three key points to address:

  • Double bandwidth with a lower energy cost per transaction to achieve 1 TB per second of memory bandwidth.
  • The already named PAM 4 that uses four voltage levels to encode and transfer up to two bits of data for each clock.
  • It can be designed and included in GDDR6 systems with mass production.

GDDR6X-vs-GDDR6

To this must be added the fact that the DRAM timings have been mitigated by double the data prefetch. This caused the demands of the memory arrays to relax, but the signals inside the chips increased in speed, where for example at 16 GB / s the time to capture and transmit the information dropped to 62.5 picoseconds.

Therefore, increasing the speed has brought with it that the frequencies give rise to a greater complexity and precision of the circuits where they exchange data. And it is that transmitting 2 bits of data at the same time results in that at the same speed compared to GDDR6 the circuits can work at half and with greater precision.

PAM 4, or how modulation affects speed

Micron GDDR6X PAM 4

As we have said, handling 2 bits per cycle requires a series of minor, but complex architectural changes. PAM 4 largely solves this complexity by adding 4 different signal levels known as symbols and whose unit of measurement is symbols per second.

Thus, where GDDR6 needs a burst length of 16 to transmit data via NRZ, PAM 4 with GDDR6X will have half the chain logically using fewer cycles. Also, GDDR6X needs two different modes of operation to work at higher clocks.

GDDR6X vs GDDR6 vs GDDR5X vs GDDR5

With PAM 4 the memory clock is the same as the command clock, which greatly simplifies the way to send data without the need for NRZ. The data driver is 2 UI wide, where every 8 symbols contain 2 data bits in 2 command cycles per clock , giving a tail of 8 Brust Length.

This has a direct impact on the so-called WCK or CK frequency. From GDDR5 to GDDR6X, the type of memory has maintained a very clear standard for the four versions of VRAM that exist:

  • The read and write accesses have two clocks per cycle, or put another way, they have a tCCD of 2 tCK . Therefore, we will only achieve 100% VRAM bus utilization when a read and write are executed every cycle per second.
  • Although the data in CMD, ADDR, and DATA are the same in GDDR6 and GDDR6X, GDDR6X requires half the CK frequency for the same performance.
  • It also shares with GDDR6 the fact of receiving commands and addresses in Double Data Rate (DDR) instead of in SDR.

If this is not enough, we must also deal with what has been seen from the point of view of efficiency.

Better energy efficiency at higher speeds

Micron has not specified much in this section, but it should be mentioned as it is representative in view of the GDDR6X temperatures on the Ampere cards.

According to the company, GDDR6X at 21 Gbps requires 15% less energy per bit transferred compared to GDDR6 running at 14 Gbps, which makes a difference in terms of bandwidth of 50%.

We have to assume that the fact of needing a lower CK frequency, together with the two data bits per UI have a lot to do with this, because from the point of view of the Vpp we are talking about the same 1.8V output and the Vdd and Vddq are the same when compared to GDDR6: 1.35 volts and 1.25 volts respectively.

GDDR6X eficiencia

So we cannot shed more light on this, and we will have to stick with the figures offered by Micron and NVIDIA, where in any case it seems that the GDDR6X are hotter as a rule.

Therefore, and summarizing, the improvements for this new type of VRAM are key, where they have also been included in specific sections that do not reduce compatibility with its previous version and instead, provide many improvements in speed and efficiency for graphics cards of NVIDIA. Will AMD be able to include them in its new RDNA 2 GPUs?