AMD suggests a Ryzen-like design for RDNA 3 chiplets would be a ‘reasonable inference’
AMD’s Product Technology Architect Sam Naffziger suggests looking forward to upcoming chiplet-based RDNA 3 graphics cards (opens in new tab) Using a Ryzen-like design would be “a reasonable corollary.”Given that he knows Exactly What RDNA 3 is designed to look like, and not allowing explicit details to be given, feels like a direct “yes, they will”.
But what does this actually mean?Even before AMD admitted that its next-generation GPUs would use a chiplet design, we were expecting it to bring the first multi-chip module (MCM) graphics processors (opens in new tab) go to the market.
There’s been a lot of talk in the past about the cross-pollination of the Ryzen and Radeon design teams, and that’s actually where the massive clock speed boost for the RDNA 2 Gen GPUs comes from. Much of the focus in past Ryzen CPU designs has been on increasing the clock speeds of its previously lagging processors, and the knowledge gleaned from the enterprise has been carried over to the latest Radeon graphics cards.
Now, RDNA 3 goes a step further, pushing chiplet designs that could have a huge impact on how AMD competes with Nvidia this generation.
In a conversation with our sister site, Tom’s Hardware (opens in new tab), Naffziger confirmed that the touted chiplet design with RDNA 3 is more than just having individual memory chips in one package (so not some Vega-like HBM memory combo). There will indeed be separate chiplets in the design, he said, but would not elaborate on how it would be configured.
When asked if it could possibly be made of a design like AMD’s Alderbaran – which is essentially a pair of large chips with interconnections between them – or more like a small chip with discrete compute and I/O The chip’s Ryzen design, he responded that the latter design would be a “reasonable inference.”
The chiplet approach has the potential to allow AMD to offer both higher-spec, higher-performance GPUs in a more cost-effective and efficient manner. That’s what this approach offers for Ryzen CPUs, letting them add more cores through smaller chips with high yields, while still being cheaper than monolithic chips with much lower yields.
The currently accepted idea for RDNA 3 is that it will include a Graphics Compute Die (GCD) and a Multi-Cache Die (MCD) as well as an I/O Die (IOD), but how much of each is still a thing back and forth is up to you listen to who. The problem is that if you’re creating a gaming GPU whose main focus is to spit out full frames of the game onto your monitor, having multiple GCDs in one package is a bit of a problem.
You’re actually talking about having multiple GPUs in one package, which is basically a crossfire. Pairing a GPU with direct computing tasks is easy, but pairing it with gaming is difficult. So hard that it’s not done anymore. AMD will need to make the multi-GCD design completely invisible to the OS and ensure that if the rendering workload is split between chiplets at any point, there is zero significant latency.
That’s why one would expect the Navi 33 design to be basically the same as the current Navi 21 setup, just a monolithic chip with RX 6900 XT-level performance. I mean, that sounds great on its own. But the idea behind the Navi 32 and high-end Navi 31 designs is that they use chiplets to have a single 5nm GCD along with stacked MCDs (and possibly a memory interface) and a single 6nm IOD that houses all the display-specific technologies.
The benefit of this is that the compute chips will be purely core graphics stuff and are easier to mass produce and therefore cheaper. It’s also possible that AMD could fill it with more shaders, ROPS, etc., since there won’t be all the I/O clutter taking up space.
However, that doesn’t mean there won’t be multi-GCD designs for gaming in the future. The latest rumor is that RDNA 3 has been updated with a game-focused monster with 16,384 shaders.
It is a game card. I have inquired about it. According to the information of RGT, it is speculated that the 16384SP is 16384SP. I don’t know if this specification belongs to this new plan.June 21, 2022
Apparently it’s not some Radeon Pro Duo because it will reported to have different interconnects (opens in new tab) Computing between chiplets. That would be a potential multi-GCD RDNA 3 design, but it’s probably just a dev tool to see if such a specification actually makes a difference in terms of gaming.
If it can, then this could be AMD’s way to undisputedly lead the GPU leaderboard next year. If not, there will be an unreleased RDNA 3 multi-GCD design prototype that may surface within a decade for quick nostalgia.