It’s barely been a couple of days and a well-known industry insider is seemingly backtracking on rumors about NVIDIA’s next GPU lineup. Industry insider Kopite7kimi has a fairly solid track record regarding early info and leaks when it comes to unreleased NVIDIA products but it appears their sources for the RTX 50 series are a little inconsistent these days. They had only just posted on Friday that NVIDIA is expected to stick with the 384-bit bus design currently used with its flagship RTX 4090 but are now backtracking to earlier speculation that it will indeed be switching over to a 512-bit memory interface.
I think my persistence is correct.
— kopite7kimi (@kopite7kimi) March 11, 2024
As mentioned in our previous post it is expected that rumors will begin to ramp up in the coming months as we near a possible launch of the RTX 50 series GB202 and GB203 processors. It is expected that the first models could launch as soon as Q4 2024 or Q1 2025. Kopite also responded to a reply from HardwareLuxx’s Andreas Schilling commenting that a 384-bit bus would be sufficient but then also asking what speed could be offered for the initial GDDR7 memory modules. Kopite responded by keeping it short with “28 Gbps” to which Schilling maintained that a 384-bit bus could still suffice.
So a 384 bit interface would give us 1,344 MB/s which should be in a good range of scaling. I don’t think NVIDIA does need a 512 bit interface.
— Andreas Schilling 🇺🇦 🇮🇱 (@aschilling) March 11, 2024
Roots of Memory Specification Rumors
The speculation about potentially different memory interfaces and memory speeds is the result of previous statements from memory manufacturers regarding initial GDDR7 adoption in consumer products. While the next iteration of GPU memory has been said to be capable of up to ~36 Gbps it has also been said the first batches will likely be implemented at slower speeds thus giving credence to multiple memory interfaces for it. VideoCardz has noted this in its coverage of the ongoing speculation.
Per VideoCardz:
“NVIDIA will still have a lot of headroom for future upgrades, as GDDR7 memory is typically advertised at 32+Gbps speeds.”
Another detail is that, at some point, it is also expected that NVIDIA will use 3 GB memory modules vs the 2 GB modules currently in production but that too is up in the air as to when it will happen. NVIDIA has repeatedly made similar changes in the past with its generational launches so the question is when, not if. All of these options do begin to paint a picture of potential release strategies as the GPU manufacturer is also known for expanding product lines with its SUPER, Ti, x90, and Titan graphics cards along with other refresh models.
With multiple versions of GDDR7 and memory interfaces for it, NVIDIA could easily plan different graphics card designs that utilize different memory configurations. Such strategies would be consistent in adjusting to market demands as competitors launch their next generation of graphics cards.