Among the many slew of bulletins from AMD right this moment round their 2022 Monetary Analyst Day, the corporate providing an replace to their consumer GPU (RDNA) roadmap. Like the corporate’s Zen CPU structure roadmap, AMD has been conserving a 2 12 months horizon right here, primarily exhibiting what’s out, what’s about to return out, and what’s going to be popping out in a 12 months or two. Which means that right this moment’s replace offers us our first glace at what is going to observe RDNA 3, which itself was introduced again in 2020.
With AMD using a wave of success with their present RDNA 2 structure merchandise (the Radeon RX 6000 household), the corporate is seeking to sustain that momentum as they shift in the direction of the launch of merchandise based mostly on their forthcoming RDNA 3 structure. And whereas right this moment’s roadmap replace from AMD is a high-level one, it none the much less provides us probably the most detailed look but into what AMD has in retailer for his or her Radeon merchandise later this 12 months.
RDNA 3: 5nm with Subsequent-Gen Infinity Cache & Chiplets
Before everything, AMD is focusing on a greater-than 50% performance-per-watt uplift versus RDNA 2. This can be a related uplift as they noticed shifting from RDNA (1) to RDNA 2, and whereas such a declare from AMD would have appeared ostentatious two years in the past, RDNA 2 has given AMD’s GPU groups a major quantity of renewed credibility.
Fortunately for AMD, in contrast to the 1-to-2 transition, they don’t should discover a approach to give you a 50% uplift based mostly on structure and DVFS optimizations alone. RDNA 3 can be constructed on a 5nm course of (TSMC’s, little question), which is a full node enchancment from the TSMC N7/N6 based mostly Navi 2x GPU household. Because of this, AMD will see a major effectivity enchancment from that alone.
However with that stated, today a single node leap by itself can’t ship a 50% perf-per-watt enchancment (RIP Dennard scaling). So there are a number of structure enhancements deliberate for RDNA 3. This consists of the following era of AMD’s on-die Infinity Cache, and what AMD is terming an optimized graphics pipeline. Based on the corporate, the GPU compute unit (CU) can be being rearchitected, although to what diploma stays to be seen.
However the largest information of all on this entrance is that, confirming a 12 months’s price of rumors and several other patent functions, AMD can be utilizing chiplets with RDNA 3. To what diploma, AMD isn’t saying, however the implication is that not less than one GPU tier (as we all know it) is shifting from a monolithic GPU to a chiplet-style design, utilizing a number of smaller chips.
Chiplets are in some respects the holy grail of GPU development, as a result of they provide GPU designers choices for scaling up GPUs previous right this moment’s die dimension (reticle) and yield limits. That stated, it’s additionally a holy grail as a result of the immense quantity of knowledge that have to be handed between completely different components of a GPU (on the order of terabytes per second) could be very arduous to do – and really essential to do if you’d like a multi-chip GPU to have the ability to current itself as a single system. We’ve seen Apple deal with the duty by primarily bridging two M1 SoCs collectively, however it’s by no means been accomplished with a high-performance GPU earlier than.
Notably, AMD calls this an “superior” chiplet design. That moniker tends to get thrown round when a chip is being packaged utilizing some sort of superior, high-density interconnect resembling EMIB, which differentiates it from easier designs resembling Zen 2/3 chiplets, which merely route their alerts by the natural packaging with none enhanced applied sciences. So whereas we’re eagerly awaiting additional particulars of what AMD is doing right here, it wouldn’t in any respect be stunning to search out out that AMD is utilizing a type of Native Si Interconnect (LSI) expertise (such because the Elevated Fanout Bridge used for the MI200 household of accelerators) to straight and carefully bridge two RNDA 3 chiplets.
RDNA 4: Furthering AMD’s Efficiency & Effectivity in 2024
And whereas AMD prepares to convey RDNA 3-based GPUs to the market, the corporate is already arduous at work at its successor.
RDNA 4, because it’s being aptly named, can be AMD’s next-generation GPU structure for 2024. Not like right this moment’s Zen 5 reveal, we’re getting virtually no particulars right here – although that was the case for the RDNA 3 reveal in 2020 as nicely. Because of this, there’s not a complete lot to dissect at this second in regards to the structure aside from the title.
The one factor we do know is that RDNA 4 GPUs can be manufactured on what AMD is terming an “superior node”, which might put it past the 5nm node getting used for RDNA 3. AMD made a equally obfuscated disclosure in 2020 for RDNA 3, and as was the case again then, AMD is seemingly conserving the door open to creating a ultimate choice later, when the state of fabs for the 2024 timeframe is healthier established. Considered one of TMSC’s 3nm nodes could be probably the most splendid final result right here, nonetheless a 4nm node is just not off of the desk – particularly if AMD has to battle for capability. (As cool as client GPUs are, different forms of merchandise are usually extra worthwhile on a mm2 foundation)
Lastly, like AMD’s Zen 5 structure, RNDA 4 is predicted to land in 2024. With AMD having established a reasonably constant two-year GPU cadence in recent times, a launch within the latter half of 2024 is just not an unreasonable guess. Although there’s nonetheless loads of time to go till we attain 2024.