Microsoft Reveals Xbox Series X Specifications and Features

[Guest_Jim_*]

Though it is a console and will not be releasing until the end of the year, the Xbox Series X with its RDNA2-based GPU may give us insight of what can be expected from future desktop GPUs from AMD. Today Microsoft shared the full specifications for the console, confirming some information we already know such as it using an 8 core/16 thread Zen 2 CPU, which will naturally have some customizations to it for the Xbox Series X. This CPU will be capable of running all the cores at 3.8 GHz with SMT off and at 3.66 GHz with SMT enabled. Microsoft wants game developers to be able to decide whether they want to utilize the additional logical threads or stick to higher-clocked physical threads.

For the GPU we are again seeing the 12 TFLOPS number for compute performance (at FP32) of the custom RDNA 2 design, but now we also have a compute unit count of 52 and a speed of 1825 MHz. For comparison, the RX 5700 XT, AMD's currently most powerful RDNA1-based graphics card, has only 40 CUs. The RX 5700 XT has a base clock of 1605 MHz and a boost clock of 1905 MHz, but the game clock you can expect to see while gaming is 1755 MHz, which is lower than that 1825 MHz which is a fixed speed; the Xbox Series X will see no boosting so there is no variability for developers or gamers to deal with.

It might also be worth noting the Xbox Series X is using a SOC, so both the CPU and GPU portions are on a single chip. This chip will be paired with 16 GB of GDDR6 on a 320 bit bus, but the memory has an asymmetrical layout to it with GPU optimal memory and Standard memory. The GPU Optimal memory is 10 GB running at 560 GB/s of bandwidth while the remaining 6 GB of Standard memory is running at 336 GB/s. A game will be given 13.5 GB of memory, which is all of the Game Optimal and 3.5 GB of the Standard, while the remaining 2.5 GB will serve the operating system and front-end shell. To the console, it is still a unified memory system.

To demonstrate what the hardware is capable of, Microsoft shares some images turning ray tracing on and off in Minecraft as well as the results of a two-week port of Gears 5. For this port, the developers at The Coalition worked with Epic Games as the game uses Unreal Engine 4 and increased the internal quality presets to what equates to the PC version's Ultra configuration, adding improved contact shadows, and the engine's new, software-based, ray traced Scree-Space Global Illumination (SSGI). Even with this increase in graphics work, the Xbox Series X was able to achieve a flawless 60 FPS at 4K in the opening cinematic. As Eurogamer/Digital Foundry points out, this is performance similar to what the NVIDIA GeForce RTX 2080 can offer from an unoptimized port. It was also shared the game is able to run over 100 FPS, so the team is looking into enabling 120 FPS for multiplayer modes. The Xbox Series X Optimized version of Gears 5 is to be available at the console's launch, and under Microsoft's Smart Delivery system, all those who own the Xbox One version will receive this version for free.

Something else we learned already is that RDNA2 GPUs will fully support the DXR Tier 1.1 standard with its hardware acceleration component, but now we have a little more information, at least for this custom design. By utilizing the dedicated hardware to offload the work from the shaders, the console is able to reach an equivalent of 25 TFLOPS of performance while ray tracing. The fixed design of the console hardware also has the potential to enable new kinds of optimizations, thanks to Microsoft allowing low-level access to the hardware. One example of this is offline construction and optimization of the BVH structures that are used to map the traversal and intersections of a ray. Such low-level capabilities might or might not translate to a desktop GPU, but the 380 billion intersections per second number might be informative.

Other technologies the GPU will support include DirectML and mesh shading. With mesh shading the graphics pipeline can be better manipulated by developers to open up performance. The pipeline has long had a fixed design and order to it, which allows bottlenecks to exist and slow the entire process, but with mesh shading the pipeline becomes more software defined. This enables developers to work around points in the pipeline that bottleneck some operations in some scenarios, allowing greater performance and greater geometric detail. Mesh shaders are something Microsoft has previously announced will be coming to DirectX 12 so that is a feature Windows desktops should eventually be able to enjoy.

The DirectML support for machine learning is something that could also prove interesting for developers and ultimately gamers. The 12 TFLOPS of FP32 compute can become 24 TFLOPS of FP16 thanks to the Rapid-Pack Math feature AMD launched with its Vega-based GPUs, but machine learning often works with even lower precision than that. With some extra work on the shaders in the RDNA2 GPU, it is able to achieve 49 TOPS at I8, 8-bit integer precision, and 97 TOPS of I4 precision. Microsoft states applications of DirectML can include making NPCs more intelligent, making animations more lifelike, and improving visual quality.

There was a lot more information revealed for the console expected to launch Holiday 2020 that is potentially not going to influence desktops, including the interesting Xbox Velocity Architecture, Sampler Feedback Streaming, DirectStorage, and hardware decompression. All of these technologies relate to accessing game assets on the console's 1 TB SSD or 1 TB of expandable storage. For raw I/O throughput, the console will be capable of 2.4 GB/s with this SSD, but by using compression it will reach 4.8 GB/s, and to keep it from slowing the CPU, a custom decompression block has been added to the CPU. Compressing data on the storage drive will naturally mean it will consume less space, and just like what is seen with GPU compression, the effective bandwidth can increase when assets are compressed as less bandwidth is needed to move the asset, but the information still needs to be decompressed and that takes some processing. To get around the cost, a special decompression component has been added so all eight of the Zen2 cores will remain accessible and free to use by the game and console. DirectStorage reduces the overhead for the I/O operations of background asset streaming, bringing the work from multiple cores to a fraction of a single core, and Sampler Feedback Streaming allows just the specifically required parts of a texture to be loaded into memory, again improving performance. The Xbox Velocity Architecture brings these technologies together to improve the performance of streaming in assets, enabling 100 GB of assets to be instantly accessible by the developer, effectively multiplying the amount of physical memory available.

It definitely looks like the future of game development could prove very interesting, as so often it is the capabilities of the consoles that dictates what will be possible on all platforms.

Source: Microsoft and Eurogamer/Digital Foundry

Back to original news post