AMD’s RDNA 3-powered Navi 31 GPU is suggested to be some kind of 15,360 core beast. That’s at least what the Twitter leakers are seemingly unified in pointing out for the next-gen Radeon gaming card.
Forget the fact the AMD Radeon RX 6600 XT, and its little sibling are expected to launch soon, the rumour mill is grinding away on something a lot bigger, and potentially a lot more significant in the AMD vs Nvidia GPU war.
The graphics card makers are going to struggle to get anything like that five-figure core count into a single chip, and make it workable long-term, and so a multi-chip module (MCM), or chiplet design is being suggested as the best way to push GPUs forward.
That is what’s being proposed for the new Radeon architecture, featuring the third generation of Navi GPUs, and expected to land around the tail-end of next year.
We’ve already looked into the specifics about why AMD would want to create a chiplet GPU, and the feasibility of such an endeavour, but if the rumours really are coalescing into a coherent form then the designs for the new graphics chips have been finalised.
In a nutshell, we’re hitting the reticle limits of the size of single chips today’s manufacturing equipment is able to create. At the moment that sits at 858mm2, with AMD’s biggest, the compute-focused GPU at the heart of its Instinct MI100 card reportedly measuring in at around 709mm2.
So chiplets are the best way to increase core counts, without making them unwieldy and unbelievably expensive.
The latest rumours collated by 3DCenter (via Videocardz), featuring one of its own forum members admittedly, appear to tally with suggestions earlier this month from a Twitter leaker, Kopite7kimi, about a figure of 15,360 for the Navi 31 GPU’s core count. We should probably immediately point out that in this context, we’re referring to ‘cores’ as stream processors or 32-bit floating-point units, because that’s the bulk of work that a GPU does in-game.
But that 15,360 figure is still rather astounding, because if you compare that with the biggest AMD RDNA chip, the Navi 21, you’re talking about an MCM GPU with three times the core count of a Radeon RX 6900 XT.
The total count comes from an apparent doubling of the number of stream processors inside an RDNA 3 workgroup. This ‘workgroup’ is the name for the dual-compute unit structure introduced with the first RDNA, and there is a suggestion that AMD will do away with all mentions of compute units (CUs) from this generation.
There will reportedly be 30 workgroups per GPU chiplet, with each of those containing 7,680 cores (30×256), and the Navi 31 die will contain two of those graphics chiplets for a combined total of 15,360 cores.
Therein lies the potential of a multi-chip graphics processor; like its Ryzen CPU cousins, it gives you the option to seamlessly connect discrete chiplets together to deliver more cores than you could conceivably fit into a monolithic die. And for a lot less cash too.
But it needs a lot more cache.
The noise is that the Navi 31 GPU will still remain strapped to a 256-bit memory bus—like its Navi 21 forebears—but will feature a much bigger Infinity Cache (IC) component. There is some speculation that could be either 256MB or 512MB, which is either twice or four times the size of the Infinity Cache attached to the Navi 21 GPU.
Unlike with second-gen Navi, however, the rumours are that the IC won’t be buried inside the actual GPU, with a per-chiplet allocation, but will come as separate ‘blobs’ of high-performance cache memory attached to the chiplets themselves. That’s likely going to be enabled in a similar way to the 3D V-Cache AMD is using for upcoming Ryzen CPUs.
It’s potentially this Infinity Cache silicon that could be the magic sauce for AMD’s RDNA 3 chiplet GPUs, and that could be what enables the whole package to be seen simply as one single graphics chip by whatever software is making demands of it.
That’s the holy grail for multi-GPU arrays, and would mean an end to any potential CrossFire (or in Nvidia’s case, SLI) shenanigans.
With CrossFire and SLI, despite pairing GPUs together you never get a linear two-times performance boost as there are a lot of logistical overheads when splitting the creation of gaming frames across two discrete GPUs. But with an effectively invisible multi-chip design you are far more likely to see a 15,360 core Navi 31 card offering closer to three times the raw power of an RX 6900 XT.
How effective this ends up being we’ll only know for sure once we get our hands on one in the labs ourselves. But there are a lot of parallels to be drawn between what AMD did with Ryzen against Intel, and what Navi 31 could do for the AMD vs. Nvidia battle, and that’s potentially very exciting for the future of PC graphics cards.