Graphics cards currently have a bottleneck that prevents them from progressing in terms of performance. A point where adding more power does not bring anything, because said element does not allow to move forward. Which is it? Although GDDR6 and GDDR6X seem very fast to us, they really aren’t. This is why the whole industry is already working with the GDDR7 and we will see it sooner than you think.
If there’s one thing we’re very clear on, it’s that the current RTX 4090 has a huge limitation and that’s the fact that its VRAM is the same as the RTX 3090 Ti, having 50% cores in addition and a higher clock speed. In other words, the memory system is not fast enough. Either way, it’s a GPU that even today’s most powerful CPU struggles to keep up with. We don’t even want to imagine a version with next-gen memory. In any case, we will see GDDR7 in the future RX 8000 from AMD and RTX 50 from NVIDIA if there are no last minute changes.
Current VRAM is insufficient, GDDR7 is required
It’s no secret that we’ll eventually get GDDR7, and even faster memories, however, we’re crossing the threshold where increasing the signal clock speed is counterproductive and there’s no point to increase the number of pins. In a world where GPUs are becoming more and more expensive to manufacture per mm², what is sought is to reduce their size more and more. We saw this in the RX 7900 and its disaggregated configuration and in the RTX 40 when using narrower buses.
All of this results in memories that transmit more information, but it reaches a point where the power consumption is so high that it is necessary to use PAM encoding or pulse amplitude modulation. This allows sending not only 0s and 1s on a single pin, but also combinations of multiple 0s and 1s. This allows the clock speed to be reduced, in exchange for converting the memory interface from digital to analog, but still with few pulses.
The use of this type of encoding will not be new in GDDR7, we have already seen it in GDDR6X where PAM2 is used to achieve higher bandwidths than GDDR6. The next memory standard will not only be from Micron and NVIDIA, but will be universal, will support PAM3 and will reach 36 Gbps, doubling the bandwidth of GDDR6 without increasing the number of communication pins between the chip and the memory.
The first designs of the future VRAM are now available
Note that we’re talking about designs, not chips, in this case it’s Cadence that has created a series of tools for memory chip makers and those who interface different chips with them. From graphics cards to game console chips, they can start adapting their designs to the new graphics memory standard and get products ready for launch in the second half of 2024.
Let’s not forget that the vast majority of these products only enter the production phase months before their launch, but that does not mean that there is no preliminary work in which the various parts that interact in the hardware should run like clockwork. And we don’t say it, just to say that the communication between a memory and a processor must be millimetric in terms of time.