Yesterday, Microsoft announced the full specs of the Xbox Series X, and with it, a lot of speculation ended but some new speculation began.
My goal here today is to break down the spec sheet to satisfy my own nerdy desire to compare it to the current gen hardware, while also looking at the weird abnormalities (the RAM configuration, mostly).
So, let’s start!
The CPU: It’s Zen 2 as expected, meaning for folks aware of the performance gains AMD has made on desktop in the Ryzen 3000 series parts, that can also be expected of the Series X hardware. Interestingly, a couple of weird points here – Microsoft is remaining at 8 cores (the Xbox One had much lower power Jaguar CPU cores but 8 of them all the same!) – which isn’t necessarily weird or unexpected, but I thought about it as an interesting point until I realized that Microsoft has kept a high thread count since the 360 – which had 3 physical cores and SMT to enable 6 total threads. The real oddity here is that Microsoft is allowing SMT usage on Series X – with a set lower clockspeed of 3.6 GHz in this mode compared to 3.8 GHz in the normal 8-core configuration.
This is actually a fairly elegant way to increase threading for more demanding games while keeping higher single-core performance in less demanding titles. If the Series X is expected to last for 7+ years as the flagship of the Xbox brand, the ability to flex to 16 threads will be vital, as developers are ramping up core count utilization in response to the changing conditions on the desktop.
Performance-wise, a direct comparison between Jaguar cores and Zen 2 is nearly impossible, but if we look at the difference between FX-series CPUs from AMD in the same timeframe as Jaguar and Zen 2 now, we can infer a huge IPC increase – probably well past double the Jaguar performance per clock, if not even more. Couple that with a nearly doubled clock speed, and it is quite a bit faster than the Xbox One.
Lastly, this is the best case for AMD, as using a direct Zen part in a console with game developers optimizing for it specifically means that Windows gaming on Zen-derived parts is going to have numerous tweaks and improvements over the next handful of years, which I personally care a lot about!
The GPU: This was the most known and analyzed spec prior to the full spec sheet detail, and so while it has been talked to a lot already, let’s dive in. RDNA2 is AMD’s next-gen architecture, a progression from the Navi-based RX5xxx family AMD is currently selling to desktop gamers. The performance is improved a fair amount over the last generation GCN-based AMD parts due to improvements in scheduling work and the amount of work that can be crammed in to single clock cycles. That means that a direct comparison is not 1:1 – you can’t simply take stream processors, compute units, and clock speeds of both consoles and math out which is better. While the Xbox Series X boasts more overall hardware than the Xbox One X, the amount of power offered is greater than the simple increase in teraflops rating portrays, mostly because it is more efficient in a given unit than the old GCN architecture was.
On top of this, the Xbox Series X GPU has access to more memory and faster memory, although we’ll hold off discussing that in detail for later. Overall, Microsoft promises the hardware is capable of 4k gaming at 60 frames per second, although that itself comes with assumptions about geometry, texture detail levels, effects and shaders that aren’t publicized as a part of that messaging. Overall, this means that between the CPU and GPU, the system is roughly as strong as a PC with a Ryzen 7 3700X and a hypothetical future Radeon RX5xxx or even 6xxx card, which is great because with reduced system overhead, the console will win the comparison, and it puts Microsoft in a good spot.
Finally, hardware raytracing is supported. While AMD hasn’t gone super in-depth on how RDNA2 handles this, they did recently confirm that the chips on PC would also support this, so it tells me that the feature is not going to be unique to the consoles. This is actually something that will bleed more into the PC market and help AMD there – Nvidia’s RTX, while being a first-mover and having that advantage, also suffers from a major downfall – it is a proprietary standard for how to raytrace, and while it is built on DXR, it has very specific modalities for hardware utilization. AMD’s solution, being in both the PS5 and XSX, has the advantage of having a hardware-level familiarity that developers can use to then bring the feature forward to PC as well. If they build it on DXR, it would work on both Nvidia and AMD GPUs, but AMD building specifically for the DXR codepaths should, in theory at least, mean that they will benefit tremendously from this. It will be interesting to see what the next-generation RTX cards bring to this fight, because it seems like AMD is ready to counter-hit after the software-light RTX launch.
The Memory System: So RAM is the Xbox Series X is…weird. Bizarre, even! The system boasts a total memory amount of 16GB, which is a fair enough generational improvement over the Xbox One X’s 12 GB, and uses GDDR6 on a 320bit bus, which makes it very high-bandwidth and takes the common game console tradeoff of using shared memory that is more advantageous to raw throughput for the GPU rather than the latency-sensitive desires of a fast CPU. However, what I immediately noticed in the photos of the SOC powering the system is that the memory layout shows…10 chips? Given this, there isn’t a clean way to break down the memory to 16 GB with identical chips. The spec sheet makes this somewhat clearer, indicating that 10 GB of the system RAM runs at 560 GB per second of bandwidth, while the remaining 6 GB runs at 336 GB per second. With that clearer, it still leaves me somewhat confused – if I assume the 6 modules along the sides of the SOC are each 1 GB, I can comfortably assume the 6 GB capacity comes from that, but a 2.5 GB GDDR6 module doesn’t exist that I am aware of, leaving the remaining 4 a mystery. Likewise, if I assume the 4 centered modules are 1.5 GB (another odd config!), then the 6 on the side would have to be uneven in a different way! It is very possible, then, that Microsoft is using custom memory capacities sourced from a memory manufacturer – which seems cost-ineffective, but could very well be the case! (Correction here: read Digital Foundry’s piece on the hardware, and it splits as 6x2GB modules and 4x1GB, which means that the bandwidth divide is uneven…which raises more questions!)
In terms of the split bandwidth decision, that seems very bizarre to me as well. Flagging 10 GB for graphics usage and leaving the remaining 6 GB slower for “main” memory could be a good split, but it also negates the advantage of a shared memory pool in the first place – namely, that as graphics grow more intense, that 6 GB pool could be a liability. Assuming it is fully shared, then it is probably fine, but having different bandwidths between the two pools means that if a workload spans both, it could be an issue. Given the bandwidth is still fairly large for both, I’m probably just overreacting to a scenario that isn’t there, but nonetheless, it is an interesting and unexpected choice of design.
Storage: A 1 TB NVMe drive rounds the system out, offering a similar capacity to today’s systems while having much faster raw speeds. Games today aren’t anywhere near limited by transfer speeds from storage to RAM on PC, but on consoles, loading times are obnoxious off the platter-based hard drives of those systems. Final Fantasy XV on my PS4 takes around 2 minutes to load at each loading screen, where the Steam version on a platter drive in my PC loads in less than 15 seconds. While a regular SATA SSD that could have been bought in huge volume right out of existing inventory would have worked functionally similarly, I am glad Microsoft is pushing the angle of NVMe. Their discussion around it makes clear that Microsoft is going to be using caching technology of some sort to deliver game assets to main system memory faster than would otherwise be possible using the NVMe drive as a pagefile of sorts. That tells me a few things – one, the full capacity of the drive is unlikely to be available for game installation; two, that the caching structure is going to be unique to Xbox and may be a strategic advantage for them; and three, they must have a sweet deal to mass-buy what is currently still a $90+ commodity PC part. That assumes it is a standard NVMe drive with a DRAM cache and an on-PCB controller – the possibility exists that Microsoft has custom-built a controller for it that would be on the main Xbox Series X hardware instead of on the actual drive hardware, which would cut the costs of the actual drive slightly, but might also be worse because it would then be a custom piece of hardware with design and R&D costs baked in.
Lastly, Microsoft, never content to let old fads die, is bringing back proprietary storage formats, using a custom slot for a 1 TB “expansion card” which looks like an old memory card for a console but in theory should be able to deliver similar performance to the internal NVMe drive, especially if Microsoft uses protocols like Thunderbolt to bring PCIE lanes to the connector used for the expansion card.
The exploded view of the system shows much what I expected – a sandwich PCB at the center of the system and upright, with a large heatsink bolted onto the SOC and a chimney-style cooling layout with a fan at the top pulling air in through bottom ventilation and through the hardware and heatsink on the way to the fan, which then exhausts it out of the top.
Overall, I think this is a solid performance from Microsoft. While the CPU sounds, on the surface, like an unimpressive upgrade, the massive clockspeed jump and the IPC improvements push a lot of extra power. The GPU is very exciting and this upgrade in the consoles means that the next generation of PC games can be even better. For AMD, having every console developer optimizing engines and game support for RDNA2 and Zen architectures will have a knock-on effect for PC versions of games, similar in many ways to how FX CPUs actually gained a bit of performance as time moved on in the last console generation, with the use of more cores and threads.
As for the competitive landscape, well, we know now that the PS5 is shipping with a similar base set of hardware, but with key differences (a weaker GPU, a more adjustable CPU, an NVMe SSD with a very strange capacity but a fair amount more performance compared to the XSX SSD, and some non-graphical raytracing uses!), but it is hard to say fully what to expect until pricing and availability are announced. Rumors have also swirled that Microsoft would ship two different variants of their next-gen hardware, with the announced spec listing being the high-end, and a lower-end system made for 4k60 at the top end might also be available. While there aren’t signs of that hardware just yet, the current kit is fairly good – Microsoft has a solid system on their hands that will just exceed most gaming PCs at its launch late this year, and for a console, that is exactly the kind of target you need.