This is it. After months of teasers and messages, with occasional leaks, we can finally reveal concrete data about the Xbox Series X. We visited Microsoft headquarters in Redmond in the first week of March, saw the console, played on it and even built it from the its components. We saw the expandable NVMe storage, we had a first sample of hardware accelerated ray-tracing and we saw how one of the talented Microsoft studios wants to improve one of the most impressive games currently available on the new Xbox. We had a taste of the excellent backwards compatibility features and yes, we can finally reveal all the official specifications for the Xbox Series X.
It all starts with three main aspects on which the new Xbox console was built: power, speed and compatibility. Undoubtedly, Microsoft has its message to share based on these pillars, but they also serve as a solid foundation for our history.
How powerful is the Xbox Series X?
With power, it all starts with Project Scarlett’s SoC – system on chip. The processor is manufactured with an improved version of the TSMC 7nm process and we know that it includes several improvements in technology, but does not include what is achieved with 7nm + EUV. The chip itself is a 360mm2 piece of silicon (much smaller than expected), combined with custom versions of an AMD Zen 2 CPU with a GPU with 12,155 teraflops of computing power.
As expected, we have a CPU with 8 cores and 16 threads, delivered via 2 quad-core units in silicon, with a core CPU (or two threads) reserved for the operating system and the front ‘housing’. Microsoft promises a 4x improvement for each core and overall on the Xbox One X and the CPU speeds are impressive, with a maximum frequency of 3.8GHz. This is when SMT or hyper-threading is disabled. Interestingly, programmers can choose to run with 8 physical colors at higher speeds or all colors and threads activated at 3.6GHz. These frequencies are completely fixed and do not adjust according to load or temperature, a detail highlighted several times by Microsoft.
In our PC tests, having SMT enabled can provide 30% or more additional performance when used well. However, at launch games, Microsoft expects studios to opt for 3.8GHZ without SMT. “From a game creator’s perspective, we expect many to use 8 colors because their current games are running with the distribution set to 7 colors and 7 work threads,” explains Andrew Goossen, systems architect at Xbox. “And to expand, to go to 14 hardware threads, it means that they have a system that does that, but then, you would have loads that are more effectively divided by them. And so, we are finding that most studios, in conversations about their choices for launch, the vast majority will choose to disable SMT and higher speed. “
The core CPU has been customized, especially in terms of security, power and performance and with 76MB of SRAM throughout the SoC, we can assume that the giant L3 cache present in Zen 2 desktop chips has been reduced. The exact Series X processor is used on Project Scarlett servers in the cloud that will replace the currently used Xbox One S-based xCloud models. To achieve this, AMD introduced EEC error correction for GDDR6 with no performance penalty (there is no EEC-compatible G6, so Microsoft and AMD created their own solution) and virtualization features are also included. This brings us to the jaw-dropping first moment: the Series X processor can run four Xbox One S games at the same time, on the same chip, and contains an internal video encoder 6 times faster than the external encoder used today on xCloud servers.
But so far, the focus has been on the GPU, where Microsoft delivers 12 teraflops of computational performance through 3328 shaders distributed across 52 computational units (there are 56 in silicon, 4 disabled to increase production gains) that run at a consistent 1825MHz. Again, Microsoft stresses that frequencies are consistent across machines, across environments. There are no higher speeds on the Xbox Series X.
“12 TFLOPs were our main objective from the start. We wanted to at least double the performance of Xbox One X to support our 4K60 and 120 targets. And we wanted twice that performance to apply evenly to all games,” explains Goossen. “To achieve this, we set a target of 2x the performance TFLOPs, knowing that the improvements in architecture would make the usual performance much greater than 2x. Our goal of doubling that number of TFLOPs in performance was established before considering the improvements in architecture , for several reasons. Especially because it set a bold target for energy consumption and defined the entire architecture of the system. “
“In addition, in the initial design phase, it is difficult to predict the benefits of architectural improvements in all the worst possible outcomes. Our aim was to double everything, not just the average. Therefore, the most practical way to ensure a basic improvement 2x in all the worst possible scenarios recorded in all games was to set an objective to double the gross performance of the TFLOPs. Therefore, we concentrated our efforts on making the effective performance even greater with improvements in architecture and new features.
We will have another article to talk about the format of the Xbox Series X, but for now, here’s an image of the console horizontally
|Xbox Series X||Xbox One X||Xbox One S|
|CPU||8x Zen 2 Colors at 3.8GHz (3.6GHz with SMT)||8x Jaguar Custom colors at 2.13GHz||8x Jaguar Custom colors at 1.75GHz|
|GPU||12 TFLOPs, 52 computational units at 1.825GHz, customized RDNA 2||6 TFLOPs, 40 computational units at 1,172GHz, custom GCN + Polaris features||1.4 TFLOPS, 12 computational units at 914MHz, custom GCN GPU|
|Process||TSMC 7nm Enhanced||TSMC 16nmFF +||TSMC 16nmFF|
|Memory||16GB GDDR6||12GB GDDR5||8GB DDR3, 32MB ESRAM|
|Bandwidth||10GB to 560GB / s, 6GB to 336GB / s||326GB / s||68GB / s, ESRAM at 219GB / s|
|Internal storage||1TB NVMe Custom SSD||1TB HDD||1TB HDD|
|Entrance exit||2.4GB / s (Raw), 4.8GB / s (Compressed)||120MB / s||120MB / s|
|Expandable storage||Input for 1TB||–||–|
|External storage||Support USB 3.2 HDD||Support USB 3.2 HDD||Support USB 3.2 HDD|
|Optical drive||4K UHD Blu-ray Drive||4K UHD Blu-ray Drive||4K UHD Blu-ray Drive|
|Performance Target||4K at 60fps – up to 120fps||4K at 30fps – up to 60fps||1080p at 30fps up to 60fps|
AMD’s RDNA architecture offers much more performance per teraflop, due to the radical new design, combined with much higher speeds (the Series X GPU runs with a 56% frequency advantage over the Xbox One X), but there are multipliers that should have an effect through the use of new features inserted in the design, such as variable rate shading, which basically tries to increase and decrease the rendering accuracy according to visibility.
However, even basic ports that hardly use the new Series X features are achieving impressive results. Mike Raynor and Colin Penty of The Coalition showed us Gears 5 running on Series X, produced in just 2 weeks. In collaboration with Epic Games to get UE4 running on Xbox Series X, basically all the quality settings have been raised to the Ultra equivalent of the PC version, added contact shadows and new global lighting via ray-tracing from UE4. In addition, the Gears 5 cutscenes, which run at 30fps on Xbox One X, were set at perfect 60fps. We’ll talk about that later, but the most important thing here is that they showed us how a port made in two weeks, without optimization, already achieves a performance very similar to that of an RTX 2080.
“Regarding our position and just looking at our experience with the hardware in this game, I think it is very positive to see how it is performing, especially knowing how much performance remains to be explored there, based on what we have already done,” says Mike Raynor. “Gears 5 will be optimized so, what you saw today will be there, available at the launch of Series X. The game will have support for Smart Delivery, so if you already have the game, you will have it for free on Xbox Series X. “
It was an impressive presentation for a game that does not even access the new generation features on the new GPU. At the moment, it is difficult to quantify the type of improvement in visual quality and performance that we will have over time, but as there are obvious parallels with current generation machines, the mixture between new hardware and new APIs will allow different loads to run on the GPU. Machine learning is a feature we have already talked about, especially Nvidia’s Turing architecture and DLSS IA upscaling. The RDNA 2 architecture used in the Series X has no equivalent colors, but Microsoft and AMD have created a new solution based on standard shader colors. With 12 FP32 computing teraflops, RDNA 2 also allows you to duplicate that with FP16. However, loads on the machine often use less than that accuracy, so the RDNA 2 shaders have been further customized.
A view of all components on the Series X chip
“We knew that many algorithms need only 8-bit and 4-bit positions for mathematical operations and related weight represent the majority of the performance overhead on these algorithms,” says Goossen. “So we added special support for the harfware in this specific case. The result is that Series X offers 49 TOPS for 8-bit integer operations and 97 TOPS for 4-bit integer operations. Note that the weights are integer, so are TOPS, not TFLOPs. The end result is that the Series X offers unprecedented intelligence in machine learning. “
There are other revolutionary features. Again, as in the Turing architecture, mixing shaders were incorporated into RDNA 2, allowing for an explosive improvement in geometric detail.
“As GPUs grew and computational performance increased, geometry processing became more targeted at the fixed function vertex, triangular scheme and GPU weaving blocks.” Mixing shading allows programmers to completely overcome these fixed-function bottlenecks by providing an alternative to existing parts of the GPU cadence. In addition to performance, mesh shading offers programmers flexibility and memory savings. Mesh shading will allow you to increase detail in the shape and animations of objects and render more complex scenes without sacrificing the frame rate. “