When Sony revealed the PlayStation 5 Pro, it did so in a very different manner to the reveal of PlayStation 4 Pro and PlayStation 5. While lead system architect Mark Cerny was ‘master of ceremonies’ for the Pro reveal, the debut has a shortlived nine-minute affair – and a good proportion of that run-time celebrated the achievements of the standard PlayStation 5 console. It was a far cry from the detailed presentation given for the reveal of Sony’s first ‘Pro’ console and somewhat bereft of detail compared to the now legendary ‘Road to PS5’ presentation given by Cerny in March 2020. We were left hungry for more details – and now, they have finally arrived.
A good three months after the reveal, Sony has delivered a brand new deep-dive presentation into the hardware design and indeed the vision behind PlayStation 5 Pro – and it’s important stuff. The philosophy behind the console marks a sea-change in the way that consoles will be made. While the core graphics rendering tech of the PlayStation 5 Pro has been enhanced over the standard machine, Cerny sees a future where machine learning plays a key role in console design, and while rasterisation is nearing its limits, he sees vast scaling in ray tracing capabilities going forward.
It’s a vision that shares commonalities with Nvidia’s strategy over the last six years, except the impression we get is that AMD will continue to be Sony’s partner of choice in delivering this new vision. To that end, a new ‘Amethyst’ collaboration between Sony and AMD has been announced.
Digital Foundry’s Oliver Mackenzie had the chance to see the new Cerny presentation several weeks ago and to interview Mark Cerny himself, alongside Mike Fitzgerald – Insomniac’s director of core technology. However, the presentation itself is something that all Digital Foundry team members were eager to discuss, so here’s our reaction to the talk – and, of course, an overview of everything we’ve learned.
GPU overview: RDNA 2.x, 16.7TF
The most obvious evolution from the base PS5 to the PS5 Pro comes in terms of its graphics processing horsepower. Rather than increasing CPU performance, improving storage speeds or targeting higher resolution displays, as you might expect from a brand new console, the main focus with PS5 Pro is shoring up performance and image quality with a more capable GPU. Interestingly, the PS5 Pro GPU is still based on AMD’s RDNA 2 (RX 6000-series) architecture, but with some features backported from RDNA 3 (RX 7000-series) – and from RDNA 4 (RX 8000-series, expected 2025) as well.
At a typical clock speed of 2.17GHz, the PS5 Pro is capable of a 16.7TF of performance – half the 33.5TF erroneously quoted before launch. This is more directly comparable to the circa 10TF number attached to the base PS5, versus the inflated (or “flop-flated”, as Mark Cerny puts it) RDNA 3 teraflop counts. Other specs line up with early PS5 Pro leaks, including 30 WGPs, 60 compute units and a max boost of 2.35GHz – though this frequency isn’t likely to be seen under normal operating conditions, perhaps due to power constraints. The maximum clock speed of the standard PS5 is 2.23GHz, but the presentation suggests that typical operating speed is the same 2.17GHz as PS5 Pro so typical TFLOPs is circa 10TF, rarely hitting the 10.23TF quoted in the original specs. As this generation has proven, however, TFLOPs is proving to be an increasingly meaningless metric – a situation acknowledged by Mark Cerny in the new presentation.
As for why PlayStation 5 Pro isn’t fully embracing later AMD graphics architectures, the reason is very straightforward: shader code for the PS5 GPU would not function on later AMD hardware, and there’s no means by which those shaders could be re-compiled in existing games. Shader compilation on the fly? It’s not viable for PS5 and – trust us – you wouldn’t want it either. Therefore, PS5 Pro has to use the same RDNA 2.x baseline, with extensions added for new features – such as enhanced ray tracing.
In terms of other features, we also know from separate disclosures to developers that the PS5 Pro has the full RDNA 2 feature set, including mesh shaders and hardware VRS, and there’s an extra 2GB of (slower) DDR5 memory on board. Of that, there’s over one gigabyte of extra memory available to developers. The GDDR6 memory used here is also around 28 percent faster than on the base PS5 – 576GB/s versus 448GB/s – despite being connected on a similar 256-bit bus.
Ray tracing improvements
Part of the PS5 Pro’s appeal is its significant improvement in terms of RT performance, allowing developers to use these features more liberally in PS5 Pro-enhanced titles without sacrificing image quality or frame-rates to the same extent as on the base PS5. This is accomplished through the RDNA 2.x architecture of the base PS5, with new extensions added from later RDNA technology.
The PS5 repurposes the texture mapping units to do box and triangle intersection tests for ray tracing, with the PS5 Pro speeding up this part of the process by two or three times. The Pro also supports BVH8 rather than BVH4, which also offers a speed-up. PS5 Pro also adds stack management in hardware, which again helps the traversal stage in ray tracing and was previously only seen in Intel and Nvidia hardware, not AMD.
Finally, more divergent RT sees a greater performance boost than more coherent RT on PS5 Pro. This divergent/coherent spectrum essentially describes the complexity of RT calculations, with shadows and reflections on flat surfaces being more coherent, and reflections on curved or bumpy surfaces being more divergent. Again, Nvidia and Intel have come up with good ways to handle more divergent RT, such as hardware sorting units and shader execution reordering, and though we don’t see exact replicas of those ideas here, the traversal upgrades and move to BVH8 ought to mean the PS5 Pro is much better equipped in its predecessor to tackle these more computationally expensive RT calculations. That opens the door to developers more easily using a wider range of material roughness for reflections, for example, rather than only sticking to mirror-like or near-mirror-like reflections.
Mark Cerny postulates that there is a limit to how far we can take rasterisation. One avenue to exceed those limitations is through RT, which has a lot of potential to scale with more research. Seeing these improvements from Sony suggests that RDNA 4 is pushing more in that direction, which is exciting stuff.
Of course, we still don’t know for sure how RDNA 4’s RT performance will turn out – it looks like a significant step forward, but there’s still no equivalent to the RT core featured in Nvidia’s RTX 20-series architecture onwards, which allows for more concurrent workloads. AMD has discussed in presentations how it’s using greater amounts of cache on its PC GPUs to fit more into the BVH, so it’ll be fascinating to see how the desktop RDNA 4 chips perform with that additional improvement.
Machine learning and PSSR
As well as the PS5 Pro’s baseline RDNA 2.x being upgraded with what Sony calls “future RDNA” for ray tracing, it also comes with “custom RDNA” for machine learning, specifically PSSR upscaling. Rather than using an existing upscaler, Mark Cerny told us that Sony focused on speed and minimising the amount of time per frame that their upscaler required. They had a choice between NPU and GPU hardware for this, and ended up going with the GPU.
The actual implementation is quite interesting, as Sony discovered that even processing a single layer of a 4K 16-channel image at 128MB was limited by memory bandwidth while using only three percent of the circa 300 TOPS available. The naïve solution was to take a smaller portion of the image through the chip at once, but then you end up with problems where the edges lack necessary information about the surrounding tiles, so that data needs to be discarded.
In the end, Sony opted to combine their vector registers in the WGPs, giving them a total of 15MB of memory and 200TB/s of memory bandwidth. It’s a similar approach to how AMD handled RT, where you’re still using generalised hardware to accomplish a specific goal. The repurposing is likely area-efficient in terms of the overall silicon footprint, doesn’t require massive changes to RDNA, and seems to have allowed Sony to reach its speed goals for PSSR.
However, PSSR isn’t yet a “fully-fused” convolutional neural network (CNN), something that Intel notably achieved with their first release of XeSS. Being fully-fused imparts a speed advantage, as you’re not touching system memory partway through upscaling a single frame. Based on DLSS taking less than 1ms and PSSR taking around 2ms through our rough calculations, it means that there’s the potential for games to run faster if this is achieved in the future. It could also open the door to upscaling being used at higher frame-rates (eg 120fps) where a circa 2ms upscale runtime is a significant percentage of the 8.3ms available for an entire frame to be rendered. There is the opportunity that AMD could take a similar approach with RDNA 4, making modifications to the main CUs to make it possible to run convergent neural networks locally on the GPU.
The future: Amethyst, ML exploration and next-gen
Sony has also announced a collaboration with AMD called Amethyst – itself a repurposed AMD GPU codename from 2014, with its purple colour coming from the combination of Sony (the Blue team) and AMD (the Red team). The idea is to make hardware architectures that work well for lightweight CNNs useful for real-time game graphics, as well as high-quality CNNs for game graphics. There’ll also be collaboration on the network architectures of those CNNs. Both parties will be able to access all of these architectures for their products.
It sounds like AMD is leading the way on the main GPU designs but Sony and AMD have this machine learning collaboration that should hopefully push things forward for both companies in terms of machine learning and ray tracing. Sony gets the benefit of iterating upon PSSR on a larger scale, AMD gets to minimise the R&D gap between itself and Nvidia. It’s not clear whether there will be some cross-pollination between PSSR and FSR on RDNA 4 GPUs, and it’ll be interesting to see where Microsoft’s upscaling efforts also fall given that they also have a PC focus, yet compete against PlayStation with the Xbox.
This isn’t just about upscaling either, it’s about computer graphics in general. There’s plenty of machine learning potential that Mark Cerny seemed quite excited by in our conversation, including denoising, improving texture quality and texture streaming, maybe even geometric quality, all within the limited bandwidth and memory space available on consoles. That’s the sort of thing that Sony should be evaluating for PS6 and an endeavour like Amethyst ensures that achieving some of this is doable on that kind of a time-frame. Likewise, announcing a partnership of this nature suggests that Sony is sticking with AMD for their next-gen console, as its major benefits won’t be available in the next couple of years.
As well as the PlayStation 6, there are plenty of exciting possibilities for the PS5 Pro as developers get to grips with the new consoles capabilities. There’s certainly potential for more advanced RT implementations as well as improved iterations of PSSR that accomplish more uniformly good results. There’s also the possibility for more improvements to running games under back-compat if Sony wants to use ML to enhance the existing library of games.
Overall, it’s been fascinating to learn more about the PS5 Pro, and we encourage you to read our interview with Mark Cerny and Mike Fitzgerald for more detail on the PS5 Pro, its development and potential for future improvement.
fbq('init', '560747571485047');
fbq('track', 'PageView'); window.facebookPixelsDone = true;
window.dispatchEvent(new Event('BrockmanFacebookPixelsEnabled')); }
window.addEventListener('BrockmanTargetingCookiesAllowed', appendFacebookPixels);