GDDR6: Features and Utilities on GPUs and Consoles

In this article, we are going to explain the operation of GDDR6 memory, as well as what are its applications and in which markets this type of memory has been implemented and what is the reason for it. We will also comment in passing on its future and the problem that exists to create new generations of GDDR memory.

GDDR6 memory features

The sixth generation Graphics Dual Data Rate, or better known as GDDR6, differs completely from previous generations by the fact that although each chip has a maximum bandwidth of 32 bits per clock cycle we now have two 16-bit channels running at same time instead of a 32-bit one. Which is what happened in previous generations.

This means that the GDDR6 can serve two memory requests at the same time, but the counterpart of this is that in order to supply the information with sufficient bandwidth, the bandwidth has had to be increased compared to GDDR5, this has led to Because the bitrate has doubled and therefore the GDDR6 transmits or receives twice as much data through its pins as GDDR5.

The advantage of having two channels is that it avoids the phenomenon of contention, which occurs when memory access is occupied by a different process, adding latency to those in the queue. So GDDR6 not only has a higher bandwidth but also a lower latency thanks to it.

GDDR6 memory specifications

GDDR6 memory can run at speeds of 12 Gbps, 14 Gbps, and 16 Gbps. At least those that are currently on the market in different products, but especially new generation graphics cards and consoles. Although the JEDEC JESD250B standard on which GDDR6 is based specifies speeds of up to 18 Gbps and even speeds of up to 22 Gbps have been reached in the laboratory, at the moment we have not seen products with GDDR6 at that speed.

Each one of the GDDR6 memory chips uses a 180-pin BGA package, its voltage varies between 1.25 V and 1.35 V. When the memory works at 12 Gbps it does so under the voltage of 1.25 V, it is also called mode of low power consumption, which can be activated at any time by the processor through the memory controller to reduce energy consumption at times when not so much bandwidth is needed or in systems that require little power such as laptops.

Regarding its capacity, this is 1 GB or 2 GB per chip, with 1 GB being more common than 2 GB in different products. Although the 2 GB configuration is gradually becoming standardized and it is possible that we will see configurations with a higher density such as 3 GB and even 4 GB in the lifetime in which GDDR6 is used.

In terms of bandwidth, GDDR6 can achieve the following data transfer speeds:

Interface bits	12 Gbps	14 Gbps	16 Gbps
128	192 GB / s	224 GB / s	256 GB / s
192	288 GB / s	336 GB / s	384 GB / s
256	384 GB / s	448 GB / s	512 GB / s
320	480 GB / s	560 GB / s	640 GB / s
384	576 GB / s	672 GB / s	768 GB /

It must be taken into account that the bandwidth in terms of Gbps can be moved in any space between 12 Gbps and its maximum speed, but we have chosen the speeds of 12 Gbps, 14 Gbps and 16 Gbps.

X8 mode or clamshell mode

This mode was already available in GDDR5, it is based on making two memory chips share the same controller data bus, so as not to have to increase the number of pins and thus the interface. This makes it possible to increase the storage capacity of the GDDR6 without increasing the bandwidth and thus not the memory controller pins on the periphery of the chip.

For this, what is done is that instead of taking a GDDR6 memory chip, connect its 16 pins to the data bus, one bit per pin, of the memory controller corresponding to a channel. What is done is to connect 8 pins to the data bus a GDDR6 chip and the other eight pins to a second chip.

The x8 mode along with the standard mode, x16, allow the following storage capacities:

Interface bits	No. of Chips (x16)	Capacity (x16)	No. of Chips (x8)	Capacity (x8)
128	4	4 or 8 GB	8	8 or 16 GB
192	6	6 or 12 GB	12	12 or 24 GB
256	8	8 or 16 GB	16	16 or 32 GB
320	10	10 or 20 GB	twenty	20 or 40 GB
384	12	12 or 24 GB	24	24 or 48 GB

It is not a RAM for low power systems

Given that it has been designed for systems with enormous graphic power, whether they are the new generation consoles or the most advanced graphic cards, this is not a type of memory designed to be used in low-power systems. The reason for this is that its energy consumption is higher than that of other types of memory.

For example, a GDDR6 memory chip at 14 Gbps and with a 32-bit bus when transmitting data that has an energy consumption of 7.5 pJ / bit, which translates into a consumption of 3.36 * 10 ¹² pJ /, which is 3.36 J / s and therefore 3.36 W. This for a single memory chip, so we are not facing a type of RAM that is designed to be used in low-power systems.

GDDR6 on next-gen consoles

GDDR6 is also used in next-generation consoles, not exclusively as video memory, but also as system-wide RAM in a unified or UMA memory configuration. This means that not only the GPU and its coprocessors have access to GDDR6 but also the CPU and its coprocessors.

Because the three next-generation consoles are based on AMD Ryzen architecture SoCs and use GDDR6 memory, they bring with them a number of common features when it comes to accessing RAM. In the first place, in terms of addressing, the CPU and GPU space is separated and depending on the memory address the memory controller points to, then the bandwidth will vary.

This leads to GDDR6 being handled differently when it does RAM for a next-gen console than when it does VRAM for a graphics card.

The case of Microsoft’s Xbox Series S and X

In the Xbox Series X and S for example we have two speeds in the bandwidth, it is not that there are two different memory wells, something that can be seen by opening both consoles, but that there are two different ways of accessing the memory.

If we talk about the Xbox Series X GPU, it directly accesses the RAM, we have 10 GB of the total memory that the GPU can access exclusively making use of all the available bandwidth and therefore occupying both channels. As for Xbox One S, the amount allocated only for the GPU is 8 GB, out of a total of 10 GB. As for the rest of the memory, it is accessible by the GPU but using only one of the channels, since the other channel is used by the CPU.

In both cases, the CPU accesses the RAM using the classic AMD Ryzen UMC, which communicates at the clock speed of the memory controller and with a 32-byte bus. In GDDR6, this speed is obtained by dividing the GDDR6 Gbps by eight, as in the console it is 14 Gbps then the memclk is 1.75 GHz. Therefore, in the next generation consoles, the CPU does not use the entire bus of GDDR6 channel memory, but it stays at 56GB / s.

What about PlayStation 5 versus GDDR6?

As for PlayStation 5, SONY has not given differentiated clock speeds like Microsoft. But, in terms of performance, it is to be expected that the SONY console will have separate access regions in order to avoid contention between CPU and GPU, which reduces effective bandwidth and thus total performance, especially of the GPU.

So in the absence of information, our assumption is that the behavior is very similar to that of the Xbox Series. At the end of the day the three consoles with Zen 2 with RDNA 2 and GDDR6 and we do not believe that AMD has opted for different solutions for common problems in the different consoles.