One of the problems in system design is the energy cost of data transfer, which is calculated by the formula P = V 2 * C * f, where V is the voltage that is squared, C is the capacity of the transistors to hold a charge and f is the frequency or clock speed.
As we increase the clock speed, the voltage rises, and the energy consumption also. This is a problem that both RAM and VRAM memory designers are facing and that GDDR6 has negatively affected, causing the number of chips that reach high speeds to be few of those that go out on each wafer.
GDDR6 problems
When GDDR6 was unveiled, three different bandwidths were announced: 12 Gbps, 14 Gbps, and 16 Gbps. Of which the 14 Gbps configuration has been standardized in a large number of devices, since it is the most used, either in the new generation consoles or in graphics cards of medium-high cards.
However, the question that comes to mind is: what happens with the GDDR6 configuration at 16 Gbps? As it is the fastest it should be the one that will be used the most, however, it is almost completely unpublished and outside of the AMD RX 6000, we have not seen it be used in other devices. The problem? The AMD RX 6000s had a huge stock shortage during their launch.
If we consider that TSMC does not seem to have problems with its manufacturing node at 7nm, then the only logic left for us is that GDDR6 memories at high speeds have a problem that makes them scarce.
The GDDR6X is a sign of it too
The fact that NVIDIA has decided to do without GDDR6 memory at high speeds for the RTX 3080 and RTX 3090 gives us many clues about the problem that GDDR6 could have and that is that said VRAM could not withstand a clock speed on its interface Communication.
GDDR6X has a much lower bus clock speed than GDDR6, so the power consumed by power transfer is much lower. But the key is not only in that point but in the fact that it is very likely that GDDR6 memory is not capable of maintaining a high enough clock speed on its interface over high periods of time.
All GDDR6 memory chips are performed in the same way, but stress tests are performed to see if a memory chip is valid for a given speed, if not then it is tested with a lower speed. The bottom line is that there are few GDDR6 memory chips at 16Gbps or faster, which has affected the stock of the RX 6000 on the one hand and has led NVIDIA to choose the GDDR6X on the other.
AMD will eventually adopt GDDR6X memory
At the moment GDDR6X memory is only found in the RTX 3080 and RTX 3090, but this will eventually be extended to the rest of the NVIDIA range, at more modest clock speeds and possibly in the RTX 3000 Super. As for AMD, something similar will happen and it is precisely the amount of GDDR6X chips that they can obtain with a performance similar to a GDDR6 at 16 Gbps is higher, which would end the stock problems.
The only impediment? AMD has opted for higher densities, so until the GDDR6X with 2 GB of density per chip does not appear, they will not make the jump, but it is possible that we will see a review of the RX 6000 with this type of memory.
GDDR6 issues affect consoles
The consoles make use of GDDR6 at 14 Gbps, a configuration that is widely used and has a high demand and manufacturing capacity.
It is said that it is the consoles that have limited the amount of GDDR6 memory for graphics cards, given the imminent launch of these, but in the case of the AMD RX 6000 on the market, the lack of stock cannot be due to that because the consoles don't use GDDR6 memory at the same speed.
GDDR6 is going to be the VRAM for the low-mid ranges
The conclusion we draw from all this, although it will seem hasty to many, is that GDDR6 is going to have a much shorter life than GDDR5 on PC or at least we will see it sharing space with GDDR6X for a long time, but being used in more modest GPUs.
As for the evolution of the GDDR6X, nothing will happen to it as long as the bandwidth does not force to place this memory at clock speeds that place it under the same problems that GDDR6 faces.
This means that you have to find other ways to scale the bandwidth of the VRAM since the traditional one is reaching the limit.
0 Comments