Tecnologia

Custom AMD GPU to Power Google’s Stadia Gaming Service (Updated)

Today at GDC, Google announced its Stadia game streaming service, which promises to deliver high performance gaming to PCs, tablets, and phones that would otherwise not be able to game. In a great piece of news for Team Red and AMD CEO Lisa Su, who was in the audience, Google announced that it decided to go with a custom Vega-based AMD GPU.Currently, GeForce Now, the other major game streaming service, uses Nvidia GPUs, so Stadia marks the first time AMD GPUs are being used for a competing service.

AMD shared some basic info on the architecture of the new custom AMD high-performance Radeon data center GPUs for Google Stadia:

  • Second-generation High-Bandwidth Memory (HBM2) to provide power savings in a compact footprint.
  • Critical data center features such as Error Correcting Code (ECC) protection to help ensure data integrity.
  • Fast, predictable performance with security features for cloud-based gaming, via the industry’s first hardware-based GPU virtualization solution built on industry standard SR-IOV (Single-Root I/O Virtualization) technology.
  • AMD also is supporting Google with its software development tools and Linux-based, open-source Vulkan driver to help game developers optimize future titles to run on the new GPU-powered platform.

According to Google, the Vega GPU will have 56 compute units that provide 10.7 TFLOPs of performance, which is similar to a Vega 56 desktop GPU. That’s more performance than the PS4 Pro and Xbox One X, combined. Google also revealed the GPU provides almost 500 GB/s of memory bandwidth, which is equal to the bandwidth of the Vega 64, and notably faster than the Vega 56.

But these are definitely data center-class GPUs, as evidenced by support for ECC HBM2 memory. That means we’re likely looking at some variant of AMD’s Pro V series, like the V340. Lisa Su presented the above graphic during AMD’s Radeon VII reveal earlier this year. The Radeon Pro graphics card next to the Google Project Stream branding, which was essentially Google’s dry run for the Stadia game streaming service, is telling.

SR-IOV support, which allows the virtualization of PCIe resources, is also a clear sign that these cards are geared specifically for data center use. If Google’s implementation follows a standard hyperscale deployment model, these cards will be employed en masse in a virtualized (or containerized) environment (think along the lines of Nvidia’s GRID). The service (game streaming) would then be provisioned to users on an as-needed basis. Google also disclosed that the systems support Linux and Vulkan, along with the ever-popular Unreal, Unity, and Havok game engines. Google covered the bases here.

Google also says its Stadia instances come with 16 GB of memory, though it’s not entirely known if that’s just system memory, just VRAM, or both. It’s likely both. Vega 56 only comes with 8 GB of HBM2, and Google also says the memory provides “up to” 500 GB/s, so HBM2 makes sense because DDR4 doesn’t provide that level of throughput.

Google also revealed that the Stadia instances feature a custom x86 CPU that operates at 2.7 GHz and comes armed with 9.5MB of combined L2 and L3 cache, hyperthreading, and support for AVX2 instructions. The “hyperthreading” moniker implies this is a custom Intel CPU, as AMD uses the actual “SMT” (Simultaneous Multi-Threading) term for its threading implementation, but there’s also a distinct possibility the custom silicon is a souped-up AMD CPU.

The CPU’s combined L2 and L3 cache capacity also imply this is an Intel processor, but notably, Intel is missing from the list of Google’s partners. AMD’s processors, at least of the class we would expect for this type of implementation, typically come with much more L3 cache than 9.5MB (and more than Intel’s models). It’s also possible the processor could also come with the new Zen 2 microarchitecture that will debut with the Ryzen 3000-series processors.

The Vega 56-like GPU should be able to meet Google’s promises of 4K gaming at 60 FPS, as long as you’re fine with lowered settings. For now, we don’t know much about the Stadia hardware beyond the intentionally-vague specifications given, but we’ll update this post as we learn more.

Image credits: Google