It is suggested that the Navi 31 GPU with AMD RDNA technology is a kind of 15,360 basic beasts. This is at least what it seems like Twitter leakers are unified in pointing out the next-generation Radeon gaming card.
Forget the fact that the AMD Radeon RX 6600 XT and its little brother are expected to launch soon, the rumor mill is erased into something much bigger and potentially much more significant in the GPU war AMD v. Nvidia.
Graphics card manufacturers will fight for anything like this five-digit core to count on a single chip and make it workable in the long run, so a multi-chip module (MCM) or chiplet design is being suggested as the best way to advance GPUs.
This is what is proposed for the new Radeon architecture, which includes the third generation of Navi GPUs, and is expected to land by the end of next year.
We have already analyzed the details about why AMD would like to create a chiplet GPU and the viability of this effort, but if the rumors come together consistently, the designs for the new graphics chips are over.
Simply put, we are reaching lattice limits for the size of individual chips that current manufacturing equipment can create. At the moment it has 858 mm2, with AMD's largest GPU, the computing-focused GPU, at the heart of its MI100 Instinct card, seems to measure approximately 709 mm2.
Therefore, chips are the best way to increase the kernel count, without being difficult and incredibly expensive.
The latest rumors collected by 3DCenter (via Videocardz), which has one of its own forum members, certainly seems to match the suggestions earlier this month from a Twitter leaker, Kopite7kimi, approximately a figure of 15,360 for the basic count of the Navi 31 GPU. We should probably point out immediately that, in this context, we refer to “cores” as flow processors or 32-bit floating point units, because it is most of the work that a GPU does in the game.
But that 15,360 figure is still pretty surprising, because if you compare it to the larger AMD RDNA chip, the Navi 21, you’re talking about an MCM GPU with a triple the basic count of a Radeon RX 6900 XT.
The total count comes from an apparent duplication of the number of flow processors within an RDNA 3 workgroup. This "workgroup" is the name of the structure of dual computing units introduced with the first RDNA, and it is suggested that AMD remove all mentions of computational units (CUs) from this generation.
According to reports, there will be 30 workgroups per GPU chiplet, each of which contains 7,680 cores (30×256), and the Navi 31 dice will contain two of these graphics chips for a combined total of 15,360 cores.
Herein lies the potential of a multi-chip graphics processor; Like its Ryzen CPU cousins, it gives you the option to seamlessly connect discrete chips together to offer more cores than could fit into a monolithic die. And for much less money too.
But it needs a lot more cache.
The noise is that the Navi 31 GPU will remain tied to a 256-bit memory bus – like its Navi 21 ancestors – but will have a much larger Infinity Cache (IC) component. There is some speculation that it could be 256 MB or 512 MB, which is two to four times the size of the infinite cache connected to the Navi 21 GPU.
However, unlike second-generation Navi, it is rumored that the IC will not be buried within the actual GPU, with a chiplet allocation, but will come as separate “blobs” of high-performance cache connected to the chiplets. themselves. It will probably be activated in a similar way to the 3D V-Cache that AMD uses for the upcoming Ryzen CPUs.
Potentially, this Silicon Infinity Cache could be the magic sauce for AMD’s RDNA 3 chiplet GPUs, and this could be what allows the entire package to be seen simply as a single graphics chip by any software that requires it.
This is the sacred grail of the arrays of several GPUs, and would mean the end of any possible CrossFire mess (or in the case of Nvidia, SLI).
With CrossFire and SLI, even pairing GPUs together will never get a linear increase in performance twice, as there are many logistics overheads when splitting game frame creation into two discrete GPUs. But with an effectively invisible multi-chip design, you’re much more likely to see a 15,360-core Navi 31 card that offers more than three times the gross power of an RX 6900 XT.
How effective will this be, will we only know for sure when we get our hands on it in the labs? But there are many parallels to be drawn between what AMD did with Ryzen against Intel and what Navi 31 could do for the battle between AMD and Nvidia, and this is potentially very exciting for the future of graphics cards for PC.