Investigating The PS5 & Xbox Series X GPU Architectures

The launch of next-generation consoles is always an exciting time, with the speculation of what these new beasts will be capable of, and how they will bring richer, more immersive worlds. Both Sony’s Playstation 5 and Microsoft’s Xbox Series X will be on store shelves by the run-up to Christmas, sporting a plethora of hardware upgrades against their veteran siblings.

But, for the Playstation 4 and Xbox One, Sony and Microsoft did something new – releasing a mid-generation refresh in an effort to offer consoles capable of taking better advantage of the newer TV technology such as 4K. Microsoft in particular, went to town pushing the ‘TFLOPs’ up for its Xbox One X, managing to reach an impressive 6TFLOPS with the Scorpio Engine’s GPU.

So, with rumors and reports that the Playstation 5 and Xbox Series X will ‘only’ push about 10-12 TFLOPS of GPU performance at most, this doesn’t sound much of an upgrade, does it? The Xbox One X has 6TFLOPS – and the Xbox Series X is rumored to be 12TFLOPs – so, does that mean the new generation consoles will only offer double the performance over the older Pro systems?

No, far from it – as you might have expected, there’s more to this than meets the eye – and it all comes down to optimization, GPU architecture and the components as a ‘whole’ rather than the just taking a single figure.

I want to dodge answering the GPU question for just one moment because there are a few things we need to establish regarding the specs of the system – feel free to skip the next couple of paragraphs if you’re familiar with this. The Playstation 5 and Xbox Series X both use AMD’s Zen 2 based processors, and while retaining the same core count as found inside the PS4 and Xbox One’s Jaguar processor, has a radically better architecture. Zen 2 provides a much greater amount of work ‘per clock’ (IPC), while pushing clock speeds up to about twice that of the previous generation and also offering SMT support.

I’ve put together a comparison run of a PS4 Pro running Linux here versus a Ryzen 7 3700X hitting different clock speeds. You can clearly see the performance difference yourself.

As for other elements of the system – the PS5 and Xbox Series X will also both sport many other upgrades, includes super-fast storage thanks to PCIe SSDs. Despite most websites leading with the headlines of “faster load times”, this is only a nice bi-product of the real benefit here. It will allow developers to remove a lot of limitations for world creation. Imagine creating a Superman game where your character can fly around the world at supersonic speeds, with the game worlds assets streaming in as needed with super high-quality textures. Or being able to virtually instantly teleport from one location to the other with no pauses or load screens. Just a few examples of how this technology can benefit games.

Anyways – back to the underlying topic for this discussion – GPU architectures. I’ll largely focus on the Playstation 5 up against the Playstation 4 here – simply because I have graphics cards which can almost perfectly mirror the results of the underlying architecture inside the PS4 Pro and a decent approximation of the PS5 GPU architecture – albeit with a few caveats.

The tests we have in store leverage two PC graphics cards from distinctly different generations – AMD’s Radeon RX 480 (which is a stand-in for the Playstation 4 Pro’s GPU) and AMD’s newer Radeon RX 5700 (and as you guessed, this is our surrogate PS5 GPU). Our goal here is to show that while clock frequencies and number of CU inside a machine, of course, do matter – there’s a tremendous difference in the efficiency of Navi/RDNA which will be found on the next-gen consoles, versus that of the current offerings.

According to AMD, RDNA first generation for PC should provide about a 25% improvement in performance per clock compared to GCN (Vega) in this below image. The Navi 10 silicon (highlighted here) has a variety of improvements to reduce power consumption – yes, but at the same clock speeds, Navi should readily outperform the older AMD architectures when it comes to gaming performance. Indeed, according to one of my sources – RDNA 2 will bring even more performance and efficiency to the table.

So, then how does this fit in with the current and next-generation consoles? Let’s start by examing the Playstation 4 Pro GPUs – as it was a bit of a curious beast. Its architecture was largely that of AMD’s Polaris (GCN 4), with some elements of AMD’s later Vega architecture thrown in. One of the biggest changes here was the inclusion of AMD’s Rapid Packed Math technology, which is a bit of a fancy way of saying that the GPU could run two half-precision floating-point (FP16) operations simultaneously on a shader, rather than a single FP32 (a normal floating-point operation). In some operations, this was a nice bit of ‘free’ performance for the PS4 Pro, and we’ve gone deeper into this topic around the launch of the console.

The PS4 Pro’s GPU ran at a 911 MHz, sporting 36 Compute Units – doing the math here (that’s 36 Compute Units x 64 Shaders Per CU x 2 Ops Per Clock x 911 Clock Speed) and we get to 4.2TFLOPS of performance.

As of the time I’m writing this, the specs of the Playstation 5’s GPU haven’t been officially confirmed. However, several leaks seem to put the GPU between 36 – 40 CU at around 2GHZ. More importantly, though, Sony is using AMD’s Navi architecture – although there is a lot of debate upon the generation here.

In the internal documentation I’ve seen (along with several other leaks online) it seems to be “Navi 10 Lite” which would imply that it’s based on the first generation of RDNA from AMD. However, given the release date of the console being later this year (after the release date of RDNA 2 for PC), and also given Sony’s history of going with custom designs… I suspect we’ll be looking at RDNA 1 with ‘some’ elements of RDNA2.

From the internal documentation – there’s 36 CU running at 2GHZ, with PS4 and PS4 Pro backwards compatibility handled by essentially altering the GPUs hardware. For example, the PS5’s GPU will run at 36 CU – albeit at a sedate 911MHz and disable new instructions if you’re running a PS4 Pro game. This will be reduced to only 18 CU at 800Mhz for a PS4 base game… and so on.

So now, onto our testing. We use a Ryzen 7 3700X running on an MSI MEG X570 motherboard with 32GB RAM as our test platform, and couple this with two different GPUs. Both the RX 5700 based on the Navi 10 silicon, and the older RX 480 (based on Polaris 10). If you notice our performance numbers for the RX 5700 is a little different than our review – it’s because our normal GPU test rig is an Intel I9-9900K running at 5GHZ for consistency in our results.

https://youtu.be/n7xWSD_NB_c

We then use a  piece of software called MSI AfterBurner which allows us to adjust the frequencies that the GPU can operate at. In essence, for the RX 480 we have two figures. One is stock (1266MHz), the second is 912Mhz (the same as the PS Pro’s GPU while still sitting at 36 CU). For the RX 5700, we run at 911Mhz, 1266MHz and finally – stock.

If you’re wondering why 912Mhz and not 911 (which would be an exact match for the PS4 Pro) – we couldn’t get the RX 5700 to remain at that speed. If we nudged the clock speed to 911, it would actually go to 909 or 910…. so after several minutes of fiddling around, I decided 912MHz would be it – and save my sanity. Oh – and also I overclocked the RAM of both graphics cards too. This (in theory) should reduce the chances of being memory bandwidth limited – I will assume that neither the Playstation 5 or Xbox Series X will have memory bottlenecks, and these tests are purely designed to see how performance scales on the two architectures (Navi and Polaris).

So – let’s go through the benchmarks.

Batman Arkham Knight is the oldest game in these tests, thus will certainly have not been coded at a time when developers could even think about taking advantage of AMD’s then uipcoming RDNA architecture. The results here speak for themselves – at 1440P (where the GPU load is at the highest) The AMD Radeon RX 480 at 912Mhz hovers at 55FPS on average, 19FPS behind the RX 5700 at the same clock speed.

Control is one of the most visually demanding games on PC, and here we’re running at just 1080P, where the 1% min FPS on the 912Mhz RX 480 can’t maintain 30fps.

So then – having gone over the results, and it is pretty clear that RDNA has huge advantages over a similarly specced and clocked Polaris based GPU. Or – to put it another way – if the Playstation 5 GPU maintained the same number of CU as the PS4 Pro, but also Sony kept the clock frequency of the GPU at only 911MHz – the PS5 would have a significant performance over the Playstation 4 Pro.

So – what notes and caveats here? Well, the Navi 10 based RX 5700 typically hits speeds around the 1700MHz mark for its clock speeds at ‘default’ operation (what I’m running here). With tweaking though, you can certainly nudge the speed closer to the 2GHz mark as we looked at during our softmodding guide.

In the above image, you can see the stock operation on the left, using Uniengine’s Super Position (a very punishing benchmark). With a quick overclock the GPU hits 1782MHz, although with our SoftMod in place, the frequency would typically rest over the 1900mhz mark. The translation for all of this – the clock speed of the Playstation 5 seems to be higher than ‘stock’ operation of the RX 5700. The Playstation 5 has been reported at both 1800mhz and 2GHz (and once again, 36 CU seems more widely accepted judging from internal documents and leaks, but there are reports of 40CU at 2GHz).

In this below image – you can see how the RX 5700 performance scales when overclocked against ‘stock’ operation using Superposition.

Furthermore – the Playstation 5 GPU will likely have the benefit of customization – and all games are optimized to take full advantage of the hardware given the nature of consoles. Long story short – the Playstation 5 GPU will totally decimate the Playstation 4 and Playstation 4 Pro silicon. So while the PS4 Pro might be 4.2TFLOPS and the PS5 is between 9.2 and 10TFLOPS (based on today’s numbers) – this doesn’t mean the graphics hardware inside the console is only about twice as powerful.

For those reading this and wondering where this leaves the next-generation Xbox – that’s a great question. From what I am able to piece together (and based on the documents I’ve seen), Microsoft’s GPU seems to be based on Navi 21 Lite – which is the second generation RDNA. I have been told by sources this generation of RDNA is further enhanced with efficiency (and power consumption too), and it looks like the number of CU for the Xbox Series X is between 52-56 – although Microsoft is running the clock frequency at a slower speed (high 1600mhz to low 1700mhz is likely for 56 CU) if MS really is aiming for 12 TFLOPS as the reports seem to indicate.

So – the same thing here applies as with Sony, the Xbox Series X might ‘only’ sport 12 TFLOPS of performance, but don’t mistake this as only being twice as powerful as the Xbox One X’s 6 TFLOPs.

As for the Xbox Series X versus the Playstation 5 – Microsoft has a TFLOPs advantage given what information we have so far, but a faster and narrower GPU for the PS5 might outperform the Xbox in some operations. Given we don’t fully understand what custom work Sony or Microsoft have chosen to implement into their respective console, it’s going to be interesting to see how all of this pans out in the long-run.

One for is for sure though – you cannot compare the PS5 or Xbox Series X TFLOPS against their older brothers. These next-generation machines are radically more efficient – and all of this skipping over topics such as Variable Rate Shading and Hardware Ray Tracing… which is something we’ll go into more in-depth soon.

Be sure to check out the first part of our Xbox Series X analysis too – which you can find here!

Comments are closed.