Sony’s Playstation 5, much like the Xbox Series of consoles from Microsoft, is an interesting beast, because the paradigms of performance aren’t just raw CPU and GPU performance now, or the increase in RAM, as has been typical from the last several generations of consoles, but instead there are a number of other fundamental changes which expands how games can be crafted and designed.
Currently, the games really aren’t tapping into the full potential of either Sony or Microsoft’s consoles, and this isn’t a surprise – it’s pretty typical for a new generation. Things will change rather rapidly in the coming year or so though, and indeed a few developers have told me that it’ll be late 2021 or even 2022 we see the true potential of what these systems are capable of, a fact Microsoft and Sony have echoed.
For the PS5, Sony has been a lot less talkative regarding the technical capabilities of the machine than Microsoft, and instead of speaking specifics of the hardware features, are talking more generally. I’ve previously leaked that one of the reasons behind this is apparently an internal decision from Jim Ryan after Mark Cerny’s Road to PS5 event was poorly received. Back then there’d been many discussions of how we’d see a much deeper teardown of the console (something Cerny actually hinted during the Road to PS5 event), but it never materialized.
I’ve leaked a number of key features for RDNA 2 and Zen 3 over the previous months, including almost all of the performance metrics of the Radeon RX 6800 series of graphics cards from AMD (along with a ton of photos before the cards were shown off to the public), and further to this some of the design challenges and decisions over at AMD. Add to this a small trickle of Playstation 5 leaks (including how the PS5 would eventually be revealed) and I think this, combined with a few new sources I now have a decent understanding of the inner workings of the PS5.
I would state that until Sony confirm this (or it naturally gets confirmed via GDC or another conference) you should take this information with pinches of salt, but I have confirmed several aspects of this with more than one source. I won’t, however, attribute which info I have received from which sources in an effort to protect their identity.
For the Playstation 5, the design is of course largely similar to that of the Xbox Series consoles in basic component selection, with a custom Zen 2 powered CPU, an RDNA 2 based custom GPU and a bunch of RAM and a fast SSD.
|Xbox Series X
|Custom Zen 2 8-Core 3.5GHZ with SMT
|Custom Zen 2 8-Core 3.6GHz with SMT or 3.8 SMT disabled
|Custom RDNA 2 | 2230 MHz / 36 Compute Units | 10.27 TFLOPS
|Custom RDNA 2 | 1825 MHz / 52 Compute Units 12.1 TFLOPS
|16 GB GDDR6 448GB/s
|16GB GDDR6 | 10GB 560GB/s 6GB 336GB/s
|825GB SSD 5.5GB/s
|1TB NVMe SSD 2.4GB/s Uncompressed
The difference between both consoles lies in its implementation, and Microsoft’s console focuses on heavy compute tasks and excels in this with its wider design, and should also do rather well with things such as Machine Learning as a result. The PS5 meanwhile is better at geometry and data latency, although I am speaking in very general terms here. The PS5 isn’t “bad” at machine learning, and the Xbox isn’t “bad” at geometry performance, instead, these are just the strengths of the machines.
Starting with the PS5’s CPU, it’s an 8 core, 16 thread design based on AMD’s Zen 2 architecture, largely similar to the desktop variant (in that I don’t believe it has any changes to improve performance in the actual cores) with the exception of the L3 cache. The L3 cache is unified like Zen 3, but it’s only 8MB according to what I am told. This information I am pretty confident in as two people have told me this, and it seems to reduce overall latency when accessing data from one cluster of cores to the other.
Unlike the Xbox, the PS5 CPU runs at only up to 3.5GHZ, though there is no SMT-less option, ie, the PS5 CPU can only run with SMT enabled at all times. In theory, having SMT disabled as an option would reduce power consumption and heat output of the Zen 2 cores of the Playstation 5, as we’ve seen this in both the desktop and also the Xbox, but Sony doesn’t seem to have enabled this for the PS5 CPU for PS5 titles, though naturally, things behave differently in back-compatibility mode (as the games don’t address all the threads).
For the CPU reserves, I was told that there’s 1 core dedicated to OS functionality (so one physical core, with two threads), and this, of course, leaves 7 cores (14 threads) available for games.
The PS5’s SSD has been the subject of quite a bit of discussion, as eagle-eyed people during the official Playstation 5 teardown noted a chip close to the SSD NAND which appeared to be DDR4. I was told this is indeed for SSD caching, and not something that the developer really has to worry about. It’s there for OS tasks and disk caching and isn’t something that the developer has to concern themselves over, and this is something which the OS and console deal with.
For the GPU, it is RDNA 2 based, with the primary tweaks being the hardware-based Geometry Engine (a block residing on the GPU portion of the PS5s SoC) and the Cache Scrubbers. Mark Cerny documented both of these briefly in the Road to PS5 event.
The PS5 GPU differs slightly from the Xbox Series X implementation, as we have Mesh Shaders not included on the GPU along with Sampler Feedback. Sampler Feedback has its own type of solutions on the PS5, and while Sony has detailed decompression technology, including for textures (Sampler Feedback isn’t decompression tech), studios and tools for Sony do have their own texture streaming technology, but the discussion of this is quite complicated, and just like rest of this article, the goal is to go over the basics and then delve further into each subject in the future.
The absence of the missing Sampler Feedback functionality (which again, Sony removed on purpose) is likely the feature that Rosario, the Sony employee mentioned here, with the statement below.
“RDNA 2 is a commercial acronym to simplify the market, otherwise GPUs with completely random features would come out and it would be difficult for the average user to choose,” wrote Leonardi.
“For example, support for ray tracing is not present in any AMD GPU currently on the market. (…) The PlayStation 5 GPU is unique, it is not classifiable as RDNA 1, 2, 3 or 4.”
“It is based on RDNA 2, but it has more features and, I think, one less…”
Inplace of Mesh Shaders, Sony is using the hardware-based Geometry Engine which essentially is Sony’s version of Mesh Shader, but totally accelerated via their own hardware block.
The GPU uses Primitive Shaders to address the GPU, along they’re apparently more complicated and more advanced than the ones found in RDNA 1. These primitive shaders can issue geometry instructions a little like Mesh Shaders and offer extreme precision on the GPU. This allows extreme culling of triangles super early on in the rendering pipeline of the GPU. The GE can be left to govern itself, but to really maximize the performance out of the console you’ll want to program it.
I was told now by multiple sources that the Playstation 5’s Geometry Engine is going to be even more critical for developing titles for the PSVR2, due the nature of pushing higher frame rates for VR solutions, and back when I was first leaking information on the PS5 development kit, this was also in line with those leaks too.
The Geometry Engine functionality is part and parcel of the custom elements of RDNA 2 from Sony, and VRS can run with extreme precision on the GPU too. We’ve already seen patents for Foveated Rendering (here is one) for the PS5 and which is used for the eye-tracking of PSVR2, and this seemingly leverages the GE of the console. It specifically mentions in the patents the varying resolution of the images based upon a users gaze.
The Playstation 5’s Geometry Engine though is a double-edged sword, as to my understanding squeezing the most out of the console will be smartly using the GE of the PS5. Again, the GE can govern itself, but you’re leaving a lot of optimization on the table, and developers aren’t even close to fully using the GE to its maximum potential.
Supposedly, the Geometry Engine has been so far best shown off in the Unreal Engine 5 Lumen demo, where it was highly leveraged along with “Nanite” to control geometry.
The second custom element is the Cache Scrubbers, which are there to handle the eviction of old data from the GPU’s caches. This is something else Mark Cerny discussed during the Road to PS5 event, but they’re perhaps the one area of the Playstation 5 which isn’t getting the attention it deserves.
Basically, old/invalid instructions on the GPU are automatically given the boot by the Playstation 5’s Cache Scrubbers, and this, in turn, frees up cache space as quickly as possible but also reduces the need to worry about this type of management from developers (though of course, you can program them too). I was told that this offers a rather substantial speedup on the GPU and reduces overheat significantly, though naturally how much this helps will depend on different game engines and on a frame by frame (workload by workload) basis.
These customizations were decided by Sony early on, and designed specifically around their needs in the console space. Indeed, RDNA 1 and 2 had a bit of an interesting history. I had been told that RDNA 1 and 2 had been a ‘nightmare’ to work on by AMD, and indeed features of RDNA 1 for desktop ended up not being implemented and held back for RDNA 2.
I also asked a developer about the Tempest Engine, which Mark Cerny (again, in the Road to PS5 event) described as a modified Compute Unit with DMA (Direct Memory Access). Its primary purpose is for audio processing on the console, but it can do more for assisting the CPU. What type of tasks? Well, if you’re familiar with the Cell processor and the SPU-like nature of the ‘helper’ cores, which were used to do a plethora of tasks on the PS3, including physics, audio, help with graphics rendering and even AI work?
According to a developer, the Tempest Engine can do much of this too, with some studios already leveraging it for things such as Physics, and likely will in the future. I asked how much the TE is being used for audio, and of course, this is a loaded question because different game types and engines all have different requirements. Though, in simpler games such as 2D titles, even with a ton of sound effects, on a typical frame, about 20 percent of the TE is used.
Clearly, a PSVR2 experience would likely increase this demand considerably, but there’s a lot of options for what the Tempest Engine can do, as a note, the Killzone 2 and Battlefield usages of the SPU of the Cell are just examples to a few things developers leverage the Playstation 3’s Cell for. This isn’t to say that they can push so many tasks to Tempest Engine, especially with audio running, and in the case of decompression in Killzone 2, there’s already decompression tech on the console.
As for the development process of the PS5, largely the OS and basic APIs are the same as the Playstation 4, albeit leveraging numerous changes to accommodate the new hardware-level features of the Playstation 5. The development environment is said to be really similar, and super friendly for optimization (as the tools are fairly mature at this point).
This seems to have been backed up by one of my leaks for the Xbox Series X, where developers were apparently having issues because the GDK (Game Development Kit) wasn’t finalized until June, and that combined with the different development environment caused Microsoft’s console for the third party early titles to underperform. My report was then backed up 3 days later by the Verge and their sources, and now various games such as Assassin’s Creed Valhalla have been patched to fix console performance, particularly of the Xbox.
From what I understand, Sony’s original PS5 development kits were also in the hands of developers quite early too, although they were far from complete and feature-packed at this point.
First Party studios were being briefed and prepped for what the Playstation 5 could do in 2017 (at this point, the design of the console had largely been locked in), but early development kits were starting to ship to first-party studios in around April 2019, about the same time as Cerny gave his interview with Wired. Most third party studios received their Dev kits in about Q3 of last year (so several months after first-party). These development kits were a little more advanced and fixed many of the missing / non-functional features.
It was Feb this year that the final hardware prototypes of the Playstation 5 development kits went out. I have already detailed the PS5 APU bring up in a separate video, so you can check that out.
Compute based functionality – 4 and 8-bit operations on the GPU, much like the Xbox Series X for machine learning. This makes sense as AMD already has confirmed the SuperSampling technology works on both the Xbox and Playstation, along with Desktop GPUs. I actually leaked that AMD RDNA 2 class GPUs had the ability to upsample for desktop, and had been told that it was a similar technology to Sony and Microsoft’s next-generation consoles, so now it’s confirmed by AMD, which is fantastic.
Upsampling is incredibly important for both next-generation consoles as well as PC (as I recently discussed in my Native Resolution is Dead) video and article, and Microsoft with DirectML is doing some pretty damn fantastic work here, with even technology to upsample low-resolution textures to 4K.
Clock frequencies tend to be really good for the PS5, with the GPU almost always staying at peak clock frequencies unless there are very heavy loads which hit the GPU. So far it seems to be about 95 percent of the time (on average, though can be more) at max speeds (so, 2.23GHz for the GPU and 3.5GHz for the CPU), though if it does dip in frequency it’s typically less than 1 frame, generally a few milliseconds. As always, different workloads will affect things differently though.
For the basic Compute Unit architecture of the PS5, it is largely identical to that of the desktop or Xbox implementation of RDNA 2, so for example, one Ray Tracing Accelerator per Compute Unit and there are the same 4 TMU per CU (like desktop or Xbox) and given it has 36 CU active, the PS5 has a peak of about 321,000G Ray Box Peak thanks to the TMU being responsible for helping the RT cores perform their work (again, just like the Xbox or desktop RDNA 2).
One developer told me like the Xbox Series X, there was some customization on the PS5 RT cores, but he doesn’t have an Xbox development kit to compare and without the RDNA 2 desktop white paper released, it makes exact comparisons difficult, Microsoft confirmed they tweaked their implementation (but still didn’t give super amounts of details as to how in public), so stay with me and I’ll follow up as soon as possible for both consoles.
The Playstation 5 has 64 ROPS, the same number as the Xbox Series X hot chips slide “GPU Evolution” (see above) where they provide 116Gpixel figure (you can calculate this by taking Xbox Series X clock frequency of 1825 and multiply it by 64).
I know these types of topics can be pretty sensitive to cover, so I will say that one of the developers who gave me this information is the very same developer who told me the Xbox Series X performance problems, which has now proven to be true as Ubisoft are fixing AC Valhalla (for example).
For those who don’t know about my content too, I will stress that I own both a PS5 and Xbox, but really neither console I intend to use as much as my PC which I primarily game on. I am fortunate to have an RTX 3080 (a review sample provided by Nvidia as full disclosure) and in general, I prefer to game on PC, but I respect those who have a console preference as I know PC is definitely not a perfect platform for many (or even myself).
None of this is also a commentary as to which machine is best, as my opinion there really hasn’t changed in a long time. Both consoles excel in different areas, and both machines are ultra-capable. If the Xbox ecosystem or the Playstation Ecosystem and games are what you prefer, then what does it matter?