Skip to main content

AMD’s next generation of graphics could use a Ryzen-inspired chiplet design

AMD’s next-gen graphics cards could adopt a new architecture leveraging a chiplet design, using a strategy that the company had implemented with its Ryzen processors. According to a patent filing with the United States Patent and Trademark Office in December 2020, AMD is looking at ways to use chiplets to give future Radeon GPUs an extra boost.

In the past, GPUs were kept on a single monolithic die because of implementation problems, like cross-firing, AMD stated in its filing.

Get your weekly teardown of the tech behind PC gaming
Check your inbox!

“Accordingly, a GPU programming model that spreads sections of work on different threads is often inefficient because the parallelism is difficult to distribute across multiple different working groups and chiplets as it is difficult and expensive computational to synchronize the memory contents of shared resources throughout the entire system to provide a coherent view of the memory to applications,” AMD lamented. “Additionally, from a logical point of view, applications are written with the view that the system only has a single GPU. That is, even though a conventional GPU includes many GPU cores, applications are programmed as addressing a single device. So, it has been historically challenging to bring chiplet design methodology to GPU architectures.”

Image used with permission by copyright holder

To solve that challenge, AMD is turning to high bandwidth passive crosslinks to couple GPU chiplets, which would allow the GPU to be communicably coupled to the processor. Processor coupling with the chiplets could allow AMD to expand its Smart Access Memory feature for Team Red gamers who also use a Ryzen processor. Here, the GPU chiplets would be arranged in an array such that each chiplet would be coupled with the first GPU through passive crosslink, Videocardz reported.

Image used with permission by copyright holder

AMD claimed that its chiplet design breaks up the SoC into smaller functional groupings, called chiplets, that performs the functions of the cores of the SoC — or a GPU, in this case. Each chiplet would have its own last-level cache, or LLC, in a way that the caches are communicably coupled to other resources so that communication remains unified and coherent across the chiplets.

Potentially coming soon to RNDA 3

AMD Radeon RX 6000
Image used with permission by copyright holder

After having launched its RDNA 2 architecture — colloquially referred to as Big Navi or Navi 2x — on its most recent Radeon RX 6000 series graphics card, the new chiplet layout could be headed to AMD’s forthcoming RDNA 3 design. Though the company had mentioned RDNA 3 in previous roadmaps, it had not detailed specifics about the architecture or the performance boost we can expect nor has AMD claimed that RDNA 3 will move to a chiplet design.

For reference, when RDNA 2 debuted, AMD CEO Lisa Su stated improvements to the new architecture meant that the company’s top-of-the-line Radeon RX 6900 XT card got a 65% performance boost compared to the prior generation. The RX 6000 series in general performed 50% better than the prior first-gen RDNA design, which already delivered a 50% boost from the old GCN architecture. If AMD’s history of performance improvements is to be followed, we can expect similar performance uplifts as the company moves from RDNA 2 to RDNA 3.

And even though it’s unconfirmed, one way for AMD to boost performance is to move to a smaller 5nm process with fab partner TSMC. The company had referred to RDNA 3 as its “Advanced Node” when detailing its graphics roadmap in the past.

It should be noted that while the use of chiplets will give AMD a unique way to boost the performance of its GPUs, the company won’t be alone in pursuing this design. Tweaktown noted that rivals Intel and Nvidia are both working on bringing chiplet designs to their GPUs in the years ahead. Nvidia is expected to debut a similar architecture when it unveils its Hopper design.

So while we can look toward a boost in performance with the arrival of chiplets on RNDA 3, it will also be interesting to see what AMD does in the future to drive even greater gains. Unfortunately, for now, insiders speculate you’ll have a bit longer to wait — RDNA 3 isn’t expected to go into a Radeon card until late 2021 or 2022. This timeline could mean that RDNA 3-powered cards will go up against Nvidia’s next-gen Ada Lovelace GPUs.

Editors' Recommendations

Chuong Nguyen
Silicon Valley-based technology reporter and Giants baseball fan who splits his time between Northern California and Southern…
GPUs just broke a 25-year-old record
Two RTX 4070 graphics cards sitting side by side.

The PC graphics card market witnessed notable growth in the fourth quarter of 2023, according to Jon Peddie Research. With shipments climbing by 6% to reach 76.2 million units, this surge marks a significant 24% increase year over year, representing the most substantial gain in over 25 years.

Projections indicate a continued upward trend, with an expected 3.6% annual growth rate from 2024 to 2026, potentially culminating in a total installed base of 5 billion units by the end of 2026, with discrete GPUs comprising 30% of the market.

Read more
Nvidia is the ‘GPU cartel,’ says former AMD Radeon manager
A hand holding the RTX 4090 GPU.

AMD's former senior vice president and general manager of Radeon has come out with some strong words against Nvidia. Scott Herkelman called Nvidia "the GPU cartel" in response to a story from the Wall Street Journal in which Nvidia's customers claim that it delays GPU shipments in retaliation for those customers shopping with other suppliers.

The accusation in question comes from Jonathan Ross, CEO of AI chip startup Groq, who said, "a lot of people that we meet with say that if Nvidia were to hear that we were meeting, they would disavow it. The problem is you have to pay Nvidia a year in advance, and you may get your hardware in a year, or it may take longer, and it's, 'Aw shucks, you're buying from someone else, and I guess it's going to take a little longer.'"

Read more
How Intel could win the GPU war this year
Intel Arc A580 graphics card on a pink background.

Intel faced an uphill climb with Arc Alchemist, and it looks like it might have another fight ahead with its next-gen Battlemage GPUs. The competition is always fierce, and AMD and Nvidia have big plans for the year ahead.

Despite the clouds that loom on the horizon, Intel might still surprise us with Battlemage -- in a good way. Here's where Intel Arc Battlemage is currently at, and why it might have a shot at being one of the best GPUs of the year.
Déjà vu
Prior to the release of Intel Arc Alchemist, one of the main complaints was that the general public was kept in the dark a lot of the time. The release date was pushed back more than once, and the information about the GPUs was fairly scarce compared to the constant hype we've all grown used to with Nvidia and AMD leaks.

Read more