NVIDIA Ampere graphics cards were the first to incorporate GDDR6X graphics memory, the next step in terms of performance to GDDR6. Thanks to this memory they have managed to equal the 21 Gbps of bandwidth that we had at that time, but they have achieved it in an innovative way that provides multiple advantages ... and some disadvantages. Was the change worth it? Let's see it.
Yes. Although currently we are not taking advantage of GDDR6X full potential in terms of performance, there is the potential to double the memory bandwidth, and for now they have already managed to considerably reduce consumption and the heat generated, which are already perks that are well worth it, especially for high power consumption graphics cards such as the RTX 3080 and RTX 3090.
GDDR6X graphics memory is not a JEDEC standard, but is a design that in this case is exclusive to Micron Technology, a new memory technology that allows you to raise performance to new heights. We are not going to go into assessing the advantages and disadvantages of the fact that a manufacturer has exclusive rights to the technology, but we are going to see if the change with respect to GDDR6 has been worth it or not.
When we ask ourselves if GDDR6X is worth it we are truly asking for a graphics card that has it vs a close one (inprice) that doesn't. Right now NVIDIA has launched 2 graphics cards with GDDR6X: The RTX 3090 and the RTX 3080. While the RTX 3070, ' the graphics card that beats the 2080 for half the price', does not include GDDR6X. With a price tag of $499 its closest competitor is the RTX 3080 ($699). But what are we getting for those 200 bucks?
While the RTX 3080 costs $200 more it also has almost 3000 more CUDA cores, supported by a memory configuration of 10 GDDR6X against the RTX 3070's 8 GBDDR6. Both graphics cards are beasts when it comes to newer titles and they will keep being it for a while, but if you want to play at 4K for 3 years or more having everything maxed out, the RTX 3080 is the way to go.
The RTX 3090 costs $1,499 and you're getting the best NVIDIA has to offer as of now, but keep in mind that if you want to play games at 4K max settings the RTX 3080 is a more than capable card for half the price. If you want to go for that price/performance sweet-spot, then the RTX 3070 is ideal for you.
At this point no one will have doubts that the GDDR6X is a leap in many respects over GDDR6 graphics memory, so if anyone had any doubts they can already dispel them: in terms of performance, the GDDR6X is superior, and this is the reason why which NVIDIA has risked introducing in its Ampere architecture graphics cards, no more, no less.
To achieve this increase in performance, Micron has used a technology that was already used in network electronics without going any further, and that is to make use of the PAM4 encoding as opposed to the NRZ encoding of GDDR6. PAM stands for Pulse Amplitude Modulation, and what it does is that it is capable of generating various binary values using the amplitude of the electrical signal.
Thus, while NRZ encoding is only capable of generating ones and zeros, PAM4 is capable of generating four binary values per clock cycle, so it is capable of encoding 2 bits per cycle compared to only 1 of NRZ, thus doubling the width band.
Despite the fact that with PAM4 encoding, GDDR6X memory is capable of doubling the bandwidth compared to GDDR6, actually for now they have left their bandwidth at the same 21 Gbps that were already achieved with a high frequency in GDDR6. Why do this? Because they get the same bandwidth with a lower speed, which translates into many advantages:
Micron boasts that it has reimagined memory with this new technology, and NVIDIA has bet very hard on it. Of course, the potential it provides is enormous, and you can reach memory performance levels never seen before.
Now, using this type of graphics memory also has a series of technical disadvantages, in addition to the one already mentioned, which is nothing more than the fact that Micron manufactures it exclusively, regulating its price and production (this makes the GDDR6X more expensive and scarce than GDDR6).
By using the PAM4 encoding, we have already explained that 4 binary values are generated per clock cycle, which allows encoding 2 bits per cycle versus 1 of the NZR encoding. This makes the voltage controller much more important, and a fault in its sensitivity could cause memory errors, causing at least read / write errors in the VRAM and leading to artifact problems, hangs, blue screens , etc.
This is not to say that the GDDR6X is going to have memory errors far from it, but it will simply be much more sensitive and prone to them and will be much more dependent on the controller. The good part is that since it works at a lower voltage and generates less heat, this disadvantage is partially offset.