New NVIDIA and AMD GPUs haven’t been released, but if the leaks are correct (and every indication they will be), memory bandwidth is back in check again due to the ramp-up, and with it performs. The two companies will therefore choose to improve their architectures while waiting for Samsung to finish its GDDR7 VRAM and above all, for JEDEC to approve the standard. When will it hit the market for use by NVIDIA and AMD?
The history of GDDR memory is almost written in stone, where all manufacturers try to be the fastest to implement improvements to it, as the JEDEC takes its time between release and release. The last one that remains to come represents even more risks for electronics, since as VRAM it must break with current standards and take a step forward.
GDDR7 VRAM, delays and very little information available
The JEDEC has hardly shown the changes it intends to introduce and those that arrive come from third-party companies that are working to create the first products under this type of memory. Samsung was the first to show brief hints of its Technical Day 2021 and the little they commented on was a speed of 32 Gbps and a new real-time error protection feature that looks more like part of the JEDEC standard than an implementation of it on your terms.
The size per chip was not discussed, leaving serious doubts as to whether the current size will be maintained or if it can scale for said speed. That said, we’ve found what we understand to be part of the final draft for this type of memory, although we can’t confirm that 100% either, but it’s a series of concepts that could be implemented by more manufacturers.
The first draft for GDDR7, what data is there of interest?
We won’t paste the whole list for obvious reasons, but we will cover some key points that we consider interesting and that Samsung and Micron will surely end up adding to their VRAM modules:
- Supports 4 separate independent channels with a point-to-point interface for data, address and command.
- Supports double data rate (DDR)
- Support mode operation pseudochannel.
- Supports programmable read/write latency.
- Supports bank aggregation and 16 internal banks per channel.
- Support order address, WCK2CKreading and writing training mode.
To these novelties, on which we insist, could be implemented by Samsung, Micro or SK Hynix if this were the case, it would be necessary to add others which have been confirmed in the standard and which all must respect, such as:
- Constant monitoring of the behavior of GDDR7 as VRAM.
- Protocol verification.
- Faster validation of drafts for the JEDEC specification.
There is no more relevant data at the moment, but what we do know is that it is slightly delayed and rumors of the implementation of this type of VRAM for the RTX 40 and RX 7000 n won’t happen, so we’ll have to wait at least another year, and hopefully, to see it in the hypothetical RTX 4090 Ti or maybe the next TITAN.