When NVIDIA unveiled its new Ampere architecture graphics cards, one of the big news was the inclusion of GDDR6X graphics memory . Those in green wanted a faster memory, so they opted for this “exotic” technology that Micron currently makes exclusively. In this article we are going to tell you what is the secret of this type of graphics memory to provide the best possible performance .
We say this type of memory is “exotic” simply because it is actually a Micron Technology project, it is not a JEDEC standard. This means that only Micron controls both the design and the production and distribution of this memory, as well as its costs. This GDDR6X memory is therefore not an overclocked version of GDDR6 as initially thought, but a new VRAM technology.
The fastest VRAM is exclusive to Micron
We cannot forget that since Micron is the only manufacturer that builds GDDR6X memory chips, it completely controls the production of it and also its price. Thus, this VRAM has the same problem as the HBM2: it is expensive and quite scarce.
Of course, this situation is not ideal for a range of consumer-oriented graphics cards, where not only costs but also production volume have to be covered, so NVIDIA has really risked a lot with this launch.
The secret of GDDR6X graphics memory
There is no doubt that this type of memory is superior to the GDDR6 that has been on the market for some time, since it offers higher operating speeds and this, of course, has an impact on greater bandwidth (and therefore on higher performance). How has Micron gone about increasing the effective frequencies in this new memory technology?
The company has not actually reinvented the wheel, but has used a technology that was already used in other markets such as Ethernet network cards (and in fact it is expected that it will also be used in the PCI-Express 6.0 standard) . This technology is called ” Pulse Amplitude Modulation Signaling “, and more precisely it is of the PAM4 type.
Until now, all memory technologies used the classic NRZ (non-return to zero) encoding in their data transmission interface. In this coding, the interface switches between two signal levels (ones and zeros) in such a way that only 1 bit is transmitted in each time unit or in each clock cycle.
For its part, PAM4 encoding generates different binary values using the amplitude of the signal (and, as its name suggests, in this case it uses 4 different values). Thus, the control unit can distinguish these four different levels and, as a result, is able to encode 2 bits of information for each clock cycle. PAM4 therefore transmits exactly 2 bits per cycle, doubling the data transmission bandwidth with respect to NRZ (you can see it very clearly in the diagram we have put above).
Advantages and disadvantages of this type of VRAM
We have already explained that GDDR6X uses a new type of data encoding, which allows the data transmission bandwidth to be literally doubled, and this is certainly its greatest advantage. This change means that to reach 21 Gbps it is not necessary to raise the clock speed of the VRAM to 2.625 GHz (memclk), but it makes it possible to reach the same bandwidth with half the speed.
Since GDDR6X has the same consumption as GDDR6, operating at a lower frequency to achieve the same bandwidth, it translates into a considerable reduction in consumption and generated heat , so here we have two additional advantages.
Now, this new type of graphics memory also has its disadvantages, and the first of them we already mentioned before: they have a higher manufacturing cost and they are also more scarce since they are only manufactured by Micron. Also, the problem with using PAM4-encoded VRAM is that the memory interface is also different, as the way of communicating has changed, and NVIDIA has had to implement it in a new memory controller .
And what about AMD? Will they have GDDR6X memory?
Starting with the second question, most likely not. In theory Big Navi will have pure GDDR6 memory, and this means that although they will achieve the same 21 Gbps as NVIDIA in its Ampere graphics, it will do so with a different memory configuration. AMD will have to raise the frequency of its VRAM to reach the same bandwidth, which means that the consumption and the generated temperature will be increased with respect to the NVIDIA GPUs.
AMD is aware of this fact, and they know that DLSS 3.0 can do them a lot of damage in terms of performance. Therefore, AMD will have no alternative but to invest in R&D and launch RDNA3 as soon as possible to be able to face NVIDIA, although they could simply incorporate a greater number of CUs in their RDNA2 GPUs to alleviate this difference in performance.