Microsoft entered the console market between 2001 and 2002 with the first Xbox, a console that was far above its competitors in the technical section and that inherited PC hardware. Almost twenty years later, every desktop console on the market follows the same philosophy as the original Xbox, but each has its own unique design personality.
The Xbox Series X SoC
The SoC of the Xbox 360 is a chip created to measure by AMD for Microsoft, which has been created to measure and is therefore exclusive to the Microsoft console. It is a monolithic 360 mm2 processor, which includes the following components inside:
- CPU based on AMD Zen 2 architecture.
- Custom GPU based on AMD’s RDNA 2 architecture
- Accelerators specialized in certain specific hardware tasks
We will go through the different parts of the Xbox Series X SoC point by point throughout this article. since it is the main unit that includes all the hardware except the RAM of the system and its SSD storage.
The Xbox Series X CPU
The Xbox Series X CPU is exactly the same CPU that AMD integrated into the SoCs that they released for laptops and PCs in 2020 under the Ryzen 4000 trademark and codenamed Renoir. These being “trimmed” versions of the Ryzen 3000 desktop cores. The reason we put it in quotes is because except for the L3 cache that has been trimmed for space reasons, the rest of the chip and its functionality remain the same, unchanged.
Unlike its counterpart on PlayStation 5, of the same architecture, it can work in two different ways:
- In the first mode, it works with the SMT active and therefore can run 2 threads per core. In said mode it runs at 3.6 GHz, making it 0.1 GHz faster than its counterpart on the PlayStation 5.
- In the second mode, the SMT remains inactive, but in return each core operates at 3.8 GHz. This mode is specially designed for Xbox One (X) games that ran on 8 AMD Jaguar cores without multi-thread support.
The fact of using the Zen 2 architecture represents an important leap, the new CPU not only doubles the clock speed of the standard Xbox One models, but also has more than double the power per clock cycle, assuming one of the biggest leaps between console generations.
The Xbox Series X GPU
|Base architecture||AMD RDNA 2 custom|
|Base clock speed||1825 MHz|
|Boost clock speed||Does not have|
|ALUs FP32 by CU||64|
|Power in FP32||12 TFLOPS|
|Power in FP16||24 TFLOPS|
|Power in Int 8||32 TOPS|
|Texturing rate||379.6 G pixels / s|
|Fill Rate||116.8 GPpixels / s|
|Cache L2||5 MB|
|Mesh / Primitive Shaders||Yes|
|Variable Rate Shading||Yes|
|Sampler Feedback Streaming||Yes|
Although they share the same basic architecture, Microsoft has taken a completely different path from SONY when it comes to designing its console and choosing the GPU. While in PlayStation 5 a design with fewer Compute Units on the GPU has been chosen, offset by a higher clock speed, in the case of Microsoft a lower clock speed has been chosen but a greater number of Compute Units. From our knowledge, Microsoft’s choice is better, since consumption increases exponentially with clock speed.
Being a unique design, its configuration in Compute Units does not correspond to any model in PC, since it has 52 active Compute Units running at 1825 MHz, which gives it about 12 TFLOPS of power continuously and without fluctuations. Like PlayStation 5, it is a design based on RDNA 2 but it has more of this architecture than its direct rival, which only has RDNA 2, the unit in charge of calculating intersections in Ray Tracing.
The technology that AMD and Microsoft have not integrated is the Infinity Cache, which is normal given the large size of the SoC. We cannot forget that we are looking at AMD’s largest SoC under TSMC’s 7nm node, since the PlayStation 5 is smaller than the Xbox Series X.
Render Backend + and VRS
ROPS units, which have the function of writing the image buffer, are called Render Backends. In RDNA 2 for PC and Xbox they have earned a + at the end of the name due to the fact that together with the rasterization unit they have been modified to perform Variable Rate Shading. Which consists of grouping the repeated Pixel Shader lines into one to calculate them as one instead of repeating their calculation.
The difference between the PC RB + and the Xbox Series X is that the console is a somewhat trimmed unit, since while the PC can write up to 128 32-bit pixels per clock cycle, 128 ROPS, in the console version the thing stays at 64. Otherwise the functionality of the Variable Rate Shading is intact and is one of the differential elements with which Microsoft takes its chest when it talks about its new console.
RAM on Xbox Series X
Xbox consoles have always used a UMA system, so we have all the memory in the same well, this time made up of 10 GDDR6 memory chips with asymmetric capacities, since although its bus from 320 bits to 320 bits at 14 Gbps, 560 GB / s of bandwidth we do not find a configuration of 10 GB or 20 GB, which would be the usual if we were talking about PC hardware, but 16 GB, since it combines different 1 and 2 GB chips.
Although we are outside the GPU area, we have to clarify that in the case of RDNA 2 architectures the number of partitions of the L2 cache is directly proportional to the number of GDDR6 channels in the system. Since GDDR6 is 2 channels per memory chip and we have 10 chips then we are talking about 20 channels and cache partitions on the GPU.
In order to avoid contentions, or conflicts in access to RAM between the CPU and its clients and the GPU and its clients, the 16 GB capacity is divided into two different blocks,
The 16 GB of memory is divided into two different blocks, one of 6 GB and the other of 10 GB, which are accessed under the following conditions the following conditions:
|Channel 1||Channel 2||Total|
|Area 1: 10 GB||280 GB / s GPU||280 GB / s GPU||560 GB / s|
|Area 2: 6 GB||56 GB / s CPU||280 GB / s GPU||336 GB / s|
The reason why the CPU accesses only 56 GB / s has to do with the way in which within Zen architectures the unit called UMC accesses RAM, it always does so at the speed of memclk * 32 bytes / cycle. Since the memclk of the GDDR6 is 1/8 of the Gbps, this means that it works at 1.75 GHz, which gives a bandwidth of 56 GB / s when the CPU through the UMC makes use of that channel.
Accelerators on the Xbox Series X SoC
An accelerator is a unit designed to do a specific job and free a more complex processor from that task, there are two kinds of accelerators:
- Fixed function: they are those that are not programmable, they always perform the function in the same way. Due to their simplicity, they are the ones that occupy the least space and consume the least, but their lack of versatility means that their performance is static.
- Specialized purpose processors: These are small CPUs designed to run a microprogram, they are usually microcontrollers inside the processor with their own instruction memory and therefore with a Harvard-like design, some may have access to the memory.
In Xbox Series X we have a series of accelerators, which can access one of the two RAM memory wells described above or both at the same time.
|DMA # 0||Both of them|
|DMA # 1 + LZ Encode||Both of them|
|DMA # 2 ´LZ Decode||Both of them|
|DMA # 3||Ambox|
|SSD codec||Both of them|
The organization of these is the same as that of Xbox One, but the most important of them is in charge of managing access to the SSD, colloquially called by Microsoft as Velocity Engine, which implements Microsoft’s DirectStorage technology.
SSD and Velocity Engine
Microsoft has taken a different path for the SSD, instead of designing its own drive, it has decided to install a drive of the M.2 2230 type, which uses a 2-lane PCI Express 4.0 bus for access. Which gives you 2.4 GB / s bandwidth, but using the data compressor and decompressor in the SoC you can reach 4.8 GB / s of bandwidth. The unit, as you can see in the photograph, unlike the PlayStation 5 SSD, is not soldered to the motherboard and despite being slower, it has a greater storage capacity by bringing 1 TB.
To expand storage, Microsoft has not opted for M.2 drives but for modified CFExpress memory cards, which contain a 1 TB capacity M.2 disk. These cards use the other two PCI Express lanes assigned to the SSD and therefore their access capabilities are the same as the main unit. The counterpart of it? As it is a single unit, Microsoft will control the price very strongly and since storage is necessary, we believe that Microsoft could capitalize on its console in that sense.
Velocity Architecture and Sampler Feedback
The unit in charge of managing access to the SSD is the Velocity Engine, since without it it would be necessary to use several CPU cores to decompress the data. If on your PC the CPU suffocates by decompressing at a rate of a few hundred megabytes, imagine with rates of gigs. That is why all new generation consoles have special units dedicated to this task within the main SoC of each of them.
In the case of Velocity Engine, it is compatible with the DirectStorage API and is designed to apply Sampler Feedback Streaming, which consists of developers when loading a texture which versions of it will be loaded and which ones will not. , in order to avoid loading into the RAM of superfluous and therefore unnecessary data.