The eighth console generation was a radical departure to what we saw in the ones which proceeded it as both Microsoft and Sony launched their upgraded, mid-generation refreshes to satisfy gamers who demanded consoles capable of running titles at 4K (or at least close to it) with improved graphical fidelity.
Sony’s Playstation 4 Pro launched in November 2016, but Microsoft would wait about a year to launch Xbox One X. Both machines sported vastly increased technical specifications over their ‘vanilla’ counterparts, with the Xbox One X easily the most powerful of these upgraded systems. The on-paper specs were impressive, with AMD tasked to create a massive APU, sporting 40 Polaris Comput Units outputting 6TFLOPs of raw performance. All of this computing power was tied together with 12GB of GDDR5 memory.
The only negative remained the Jaguar CPU, which was tweaked and enhanced, running at a nippier 2.3GHz, a 550MHz boost over the original Xbox One. Lots of performance on tap indeed, although console was arguably even more unbalanced and skewed towards GPU performance than the other consoles. Either way though, TSMC’s 16NM process used to great effect, with a die area of 359mm2 and coming in at a hefty 7 billion transistors.
If you’ve wondered why I’ve spent so long prattling on about the previous generation systems, when their history is somewhat established, it’s because these mid-gen refreshes created a bit of a problem. The previous generation architecture was still ingrained deeply within them, and limitations such as their slower CPU, lower IO (such as with the 5400RPM disk) and the like were starting to hold developers back.
The 1.31 TFLOPS of the Xbox One looked paltry compared to the Xbox One X, but developers still had to code for it. This provided two problems, one – developers had to worry about performance limitations of the base consoles, while still having some hardware limitations because of their roots. While simultaneously, marketing a console on say, raw TFLOPS was now much more difficult.
|Playstation 4||Playstation 4 Pro||Xbox One||Xbox One X|
|CPU||AMD Jaguar 8 Cores 1.6GHZ||AMD Jaguar 8 Cores 2.1GHZ||AMD Jaguar 8 Cores 1.75GHZ||AMD Jaguar 8 Cores 2.3GHZ|
|GPU Architecture||Based On GCN 1.1||Based on GCN 4 with some GCN 5||Based On GCN 1.0||Based On GCN 4 (Polaris)|
|GPU Specs||18 Compute Units @ 800MHz||36 Compute Units @ 911MHz||12 Compute Units @ 853MHz||40 Compute Units @ 1172MHz|
|RAM||8GB GDDR5||8GB GDDR5 & 1GB DDR3 For OS||8GB DDR3 & 32MB eSRAM||12GB GDDR5|
|Bandwidth||176GB/s||217.6GB/s||DDR3 68GB/s eSRAM 204GB/s||326GB/s|
In the so-called “bit wars” in the early days of consoles, marketing could often be boiled down to 8-bit is worse than 16-bit, and so on. Of course, this wasn’t strictly true, and anyone who saw the Jaguar 64 wheeze running code compared to any of its competitors (such as the PS1, Saturn or N64) would tell you that bits aren’t the be-all and end-all. But just like that era, it was now all about TFLOPS – with Sony advertising heavily of its TFLOP lead over the original Xbox One, and then Microsoft happily turning the tables on Sony for the Xbox One X, trumpeting its 6TFLOPS performance over the PS4 Pro’s 4.2TFLOPs.
But now, efficiency in performance and features are the name of the game, as is almost certainly increased power consumption and a shift in marketing messaging. Things are still developing for the next-generation Xbox, and we’ll be doing a deeper dive versus the Playstation 5 soon (I’m currently working on that too), but for now, I felt that any such content would be lacking context and so this piece was born.
The Xbox Series X is said to be (in terms of case size) about double the physical dimensions of the Xbox One X, a drastically increased volume which caused polarizing opinions over its design. It’s not the small, discrete box which can be tucked into your media center, instead, it looks much more like a small form factor PC build, designed to house high-end performance parts, but also without them catching on fire. While I’ve not had chance to inspect the inner mechanisms of the console, it doesn’t take much to assume that we’ll see a return of a robust cooling system like the Xbox One X (which used a custom Vapor cooling solution), particularly given Phil Spencer is pledging that the console will be ‘as quiet as the Xbox One X’.
Funnily enough, and circling back to my Jaguar comment earlier – Microsoft has actually asked users to “do the math”. Ironically, this was a key marketing message from Atari in the Jaguar 64 too. This isn’t to say for one moment I am equating the two consoles, it’s obvious Xbox will be a resounding success, but still, I find it personally amusing.
Microsoft is keen to point to a 4-fold increase in raw performance from the CPU alone, and with a 2x increase in GPU performance, which would pin the console down to be about the 12 TFLOPS mark. There is some debate whether this is ‘really’ 12 TFLOPS, or closer to 9TFLOPS with RDNA’s efficiency making up the remainder (given Navi/RDNA is way more efficient than Polaris, we’ll get to that more in a moment), but I am fairly confident this isn’t the case.
Firstly – subtle nuance like how efficient a GPU architecture is will be lost on the majority of individuals, and Phil Spencer and his team would get called out at failing to live up to the promise of 2x the performance if the GPU was only 9TFLOPS, no matter what the actual capabilities of the machine were. The second is several leaks and sources from websites such as EuroGamer and Windows Central who have stated 12TFLOPS is accurate.
I, myself have seen an alleged internal document which is apparently from the early concept of the machine and here 10TFLOPS+ was the minimum target (I cannot share this image as I was asked to not do so by the individual in question, and of course I will respect their wishes). I was told however that the team have since blown past that figure and for Xbox Series X (which was previously known only as of the codename Anaconda) is hitting 12 TFLOPS.
So, before we get into speculation for a moment – let’s go with things we do know. The RX 5700 XT is currently the highest-end Navi SKU we know (though it’s using the first-gen RDNA, once again we’ll talk about that more in a moment). This particular piece of silicon isn’t small – weighing in at 251mm2, and housing 40 Compute Units. These chips are capable of hitting 2GHz, albeit with higher power consumption and heat output than would be ideal.
For the Playstation 5, we’ve seen several APU and GPU benchmarks pegging the GPU at between 1800 to 2GHz, and assuming the same holds true for the Xbox Series X, 40CU is not going to cut it. You can do the maths yourself if you’d like. 40 (Compute Units) x 64 (number of ALU per CU) x Clock Speed x 2 (2 operations per clock). Even at 2GHz, MS would only hit 10.2 TFLOPS for the Xbox Series X.
So, we’re likely looking at more than this – 48 would be the minimum number to achieve this, but then if one of these CU is defective, the whole APU would need to be ditched (or used for LockHart, the reported 4 TFLOP lower end SKU in the next-gen Xbox Family). Realistically though, this isn’t the case – and Microsoft would instead have a margin of error built into the APU.
For both the Xbox One and PlayStation 4 for example, there were an additional 2 CU present on-die which were disabled, and existed purely for yields. The Playstation 4 has 18 CU enabled for example, running at 800MHZ to give the 1.843 TFLOPS of GPU performance. But, if one of those CU were damaged, the APU would be useless – so, Sony’s contingency plan was simple. Manufacturer an APU with 20CU, and if 1 or 2 of those CU were damaged, not a problem, there was still 18 ‘good ones’. Of course, it didn’t matter which of the 20 CU was damaged, or if none were damaged, two would still be disabled.
It’s hard to believe that Microsoft (and Sony) wouldn’t follow suit for the next-generation consoles.
However, given the verbal and written confirmation of Microsoft that we’ll see hardware-accelerated Ray Tracing in the new Xbox, and several comments of “next-gen” RDNA, it’s almost 100 percent Microsoft are opting for RDNA 2.0 for their Xbox Series X console. I was the first to break that RDNA 2 (known as Navi 2x) would have hardware Ray-Tracing, and also leaked the existence of Navi 21 and 23, the latter of which is known as “the Nvidia Killer”.
Hyperbole aside, I have recently been told that RDNA 2.0 actually is even more efficient than RDNA 1.0, which itself put out about 25 percent more work than Vega per shader running at the same clock speed (Instructions Per Clock). Details are scarce on AMD or Microsoft’s implementation of Ray Tracing, but it looks like its a hybrid approach, and I’ll go deeper into this in the future. Although a few patents and Microsoft’s recent tweaks to their DXR (DirectX RayTracing) API likely holds some clues, as they’ve recently nudged it to Tier 1.1 with numerous enhancements and improvements.
I will delve further into what this underlying technology (such as Ray Tracing and Variable Rate Shading) will mean for games in the not to distance future, but I want to keep things somewhat organized, less risk this becoming a sprawling article trying to tackle too many subjects.
So what about the CPU then? As I mentioned near the start of this, the CPU inside both Microsoft’s and Sony’s current-gen efforts is easily their weakest link, with the PS4 base console using a 1.6GHz Jaguar CPU, running with 8 CPU cores, cranking this up to a 2.1GHz speed for the PS4 Pro. Not all of these cores are allocated to games, of course. Sony (and Microsoft) did release an update unlocking some of the 7th CPU cores time to game code should developers wish it.
It’s not ‘all’ of the 7th core, and is about 50 percent of the 7th core for both machines, so 6 “full” AMD Jaguar cores and 50% of core 7. The remaining being allocated to system tasks.
The Playstation 5 and Xbox Series X are confirmed to use Zen 2, in an 8 core configuration with SMT (Simultaneous Multi Threading). Clock speeds aren’t official yet, but the leaked APU for the Playstation 5 seems to show a 3.2GHz clock speed for the CPU, while the Xbox Series X is reported to run at 3.5GHz.
So, how does this compare to the Playstation 4 and Xbox One’s Jaguar processor? It doesn’t compare – it trashes them. There’s a rather interesting series of benchmarks from a user who’d gotten GeekBench 4 running on a PS4 Pro, thanks to an exploit to run Linux. Here, his geekbench 4 score was 1400 for a single core, and 7684 for multi-core.
I did a few runs using a Ryzen 7 3700X, which is a good stand-in for the 8 core CPU inside the next-gen consoles (likely the consoles will tweak things, but we’ll discuss that another time). These runs were on Windows versus Linux, which means scores aren’t a direct apples-to-apples comparison, but still, we have a ballpark figure.
Even disabling SMT of the 3700X and running at the same clock speed as the PS4 Pro’s Jaguar processor, you can see scores are about doubled for Zen 2. GeekBench 4 isn’t indicative of all workloads by any stretch of the imagination, and clearly as developers optimize their code for consoles things will likely differ, but it does illustrate a rather simple performance comparison between the CPUs inside the current console versus their upcoming replacements.
We’ve seen recent demos of AMD’s FEMFX which is code running on CPUs with the ability to better mimic how materials deform, bend and react in the natural world. Realistic graphics are of course important, but we’ve all played a game and run into things which totally break realism because material such as a piece of cloth acts more like a piece of plywood, or wondered why you use a flamethrower on a tin can and notice how indestructible it is.
Of course, some physics and workloads will likely still run on the GPU as compute commands, but with vastly more CPU performance, developers are given access to create richer worlds, and also CPU performance will no longer be the limiting factor in higher frame rates too. High refresh rate gaming was often held back on console because of the lackluster processors inside the machine, but now for competitive FPS titles, 120FPS could now be targetted, dialing back other details as necessarily and focusing on high response and frame rates above all else.
We’ve very little information right now on the APU of the consoles, although perhaps we saw a sneak peak of the APU at E3 2019, with Microsoft’s first reveal of the console. This APU looks monstrous, and given the angles it’s very difficult to get an accurate measurement of the beast, but it appears to about on par with the size of the Scorpio Engine found inside the Xbox One X (so about 350 – 370mm2 by my guess). Given this APU is using TSMC’s 7nm+ process, this gives a great insight into just how much compute power is being stuffed into the APU.
In terms of CU, a workgroup processor inside the original RDNA architecture came in at about 4.5 to 4.6mm2, and each WGP is 2 CU. Liabe Brave over at NeoGaf did a nice breakdown on the Navi 10 silicon which is featured inside the RX 5700 series. He also tried estimating the size of the WGP inside the next-gen RDNA, however we don’t much how the architecture will evolve, but realistically it won’t get much bigger, and given we’re likely gonna see the console manufactured on TSMCs 7nm+ process, this is probably all offset anyway.
In terms of raw size then, the low 50 CU active, running at a nippy clock speed and backed up by 8 Zen 2 cores seems probable. I speculate to save a decent chunk of die space, the Zen 2 CPUs will have their L3 cache snipped by 50 percent. This isn’t confirmed, but of the 74nm chiplets of the Ryzen 3000 series (for example), just over half of the space is taken up by L3 cache, so halving it to just 16MB might be a nice way to claw back a bit of die space, if push came to shove.
Microsoft would likely be eager at this point I discuss one element of the next-generation Xbox which is a huge upgrade compared to the current Xbox One consoles, and that’s the storage. One of the worst kept secrets of the next-generation is the virtual elimination of load times in games thanks to the massive increase in IO available to these SSDs.
While on the surface this might seem like a nice quality-of-life improvement, and not necessarily something that’ll impact your enjoyment of games other than not having to watch a loading screen tick by for a minute or two, it does actually drastically change how a game can be constructed, and also provide a lot of additional options.
Do you remember the first time you loaded up CastleVania Symphony Of The Night and the ‘areas’ of the game were disguised as a corridor which you had to walk through while the PS1’s CD drive stuffed the data of the next location into the PS1’s meagre RAM? Or, how about Mass Effect and the elevators, which seemed to last for just a bit too long because the game was doing much the same thing, albeit with shinier graphics?
This isn’t a criticism of those games, merely pointing out how a seamless world could potentially be constructed with the ability of the system to stream vast amounts of data. You could instantly teleport (or fast travel I suppose) with the game barely missing a beat between one location and another, and this alone could lead to some very interesting experiences.
Unfortunately, information on what exactly the Xbox Series X (and we can presume the Lockhart will be the Xbox Series S) will be outfitted with, but it’s not difficult to imagine that it’ll be similar to the faster NVMe SSD’s available now, with read speeds capable of hitting several thousand MB/s, and with no seek times either. Seek times are a necessary evil with a spinning mechanical platter of the current generation and means that if the read head is the other side of the disk it takes time for it to physically travel and locate that missing piece of data.
Some will point to the fact you can use SSD’s on say the PS4 right now, and it helps loading times but doesn’t really provide a huge performance benefit. This is true, but by design. The PlayStation 4 launched with a paltry SATA 2 interface, which was subsequently upgraded to SATA III later on (with the arrival of the PS4), but compared to an NVMe SSD, SATA is still rather slow.
This is further compounded by developers simply not coding experiences designed to take advantage of an SSD, when the console simply doesn’t have one by default.
The next generation will be a very interesting one – and while I don’t believe graphical upgrades will be slim on the ground, they might not be what we saw with say the original Xbox to the Xbox 360. Instead, these new boxes will provide graphical and resolution upgrades, but also finally provide developers the ability to fulfil a promise they’ve struggled with for decades – deeper immersion.
Seamless game worlds, better lighting, more accurate physics, and hopefully AI which acts in a realistic way.
Hopefully you’ve enjoyed part one of this analysis and with luck you’ll stick around for the part two, where we’ll start going deeper into the technical aspects of the machine and explore Ray Tracing and other technologies.