[Hardware] Microsoft Xbox Series X's AMD Architecture Deep Dive at Hot Chips 2020

[.-AdiiLo-.]

AMD's next-gen console APUs come out of hiding.

Today at Hot Chips 2020, Microsoft will detail more of the internals of the Xbox Series X system architecture. Much of the information in the pre-published presentation covers details we already knew, but with the new slides in hand, there are some additional juicy morsels inside, including a diagram of the processor die. We have the full slide deck at the end of this article, but let's quickly hit a few of the highlights.

Starting off with a bang, while we've known Xbox Series X would have an 8-core/16-thread Zen 2 CPU and a 52 CU GPU, this is the first time we've seen the full die shot. That GPU section is, not surprisingly, massive. The full chip is 360.4mm square, with 15.3 billion transistors. Doing some quick image analysis, the GPU takes up roughly half of the die (47.5% if you want a more precise estimate).

 

 

Considering everything stuffed into the silicon, the die size is actually pretty manageable.

A Zen 2 CPU chiplet measures 74mm square (with four times the L3 cache compared to the Xbox Series X APU), and then tack on a GPU that has more features and shader cores than Navi 10 (RX 5700 XT), which measures 251mm square. That's 325mm square without the enhanced Navi 2x cores and 12 additional CUs. Alternatively, it's over twice as big as AMD's Renoir APU, which measures 156mm square.

Taking a closer gander at the die shot, it looks as though the full chip includes up to 56 CUs. (Feel free to correct us, as counting these things can be a bit tricky.) That's because with a relatively large die, yields of functional chips can be greatly improved by building in some redundancy. If our counting is accurate, it suggests AMD and Microsoft have only disabled 4 of the dual-CU clusters, and of course all eight of the CPU clusters have to work in order to get a functioning chip.

 

 

The presentation also takes some times to discuss the decreased difficulties in chip scaling relative to Moore's Law. While the chip size of the Xbox Series X is in line with previous console hardware (375mm square for the Xbox One in 2013, 367mm square for the Xbox One X in 2017), and transistor counts have more than doubled relative to the Xbox One X (6.6 billion to 15.4 billion), the die cost is higher. Microsoft doesn't specify how much higher, but lists "$" as the cost on the Xbox One and Xbox One S, "$+" for the Xbox One X, and "$++" for the Xbox Series X. As we've noted elsewhere, while TSMC's 7nm lithography is proving potent, the cost per wafer is substantially higher than at 12nm.

 

Microsoft also provided additional details on the Velocity architecture and its motivation.

The Velocity architecture deals with using lower cost SSD storage in place of increasing system memory. In the past, the price for DRAM decreased by 30% each year, but for the past 8.5 years it has been dropping at 5% per year. Flash memory in contrast has continued to decline in price by 23% each year over the same period.

Because SSDs are so much faster than the spinning platters used in HDDs for the previous consoles, there's a huge benefit in game load times. Combined with the sampler feedback streaming, the hardware is able to leverage the flash memory for improved overall performance without the need to massively increase the amount of DRAM.

 

 

One of the more interesting new aspects of the next-generation consoles is ray tracing support.

We've known both the Xbox Series X and PlayStation 5 would support hardware ray tracing, but we didn't have any clear indication of what level of performance to expect until now. Microsoft at least appears to be downplaying the importance of using ray tracing in games, or at least that's our interpretation. It emphasizes the hybrid rendering approach, much as Nvidia's RTX cards have done, noting it's "not a complete replacement" for traditional rendering methods.

As far as ray tracing performance goes. Microsoft only gives peak performance of 380G/sec ray-box calculations, and 95G/sec ray-triangle calculations. How does that compare with Nvidia's RTX 20-series hardware, where it claimed the RTX 2080 Ti could do 11 Gigarays/sec? We don't really know, as Nvidia didn't give peak performance and Microsoft didn't give typical performance.

 

The slide does note that it was only a "minor area cost for 3-10x acceleration." That seems relatively similar to what we've seen from Nvidia's hardware. In games that combine ray tracing with traditional rendering, running with ray tracing enabled on Nvidia's RTX 2060 is often two or three times as fast as the previous generation GTX 1080 Ti.

 

We still don't have concrete details on whether the Xbox Series X will have any additional hardware related to machine learning, for example like Nvidia's Tensor cores that are used for DLSS 2.0. The slides do mention Machine Learning Acceleration, and note an "ML inference acceleration" of 3-10x for a small die area cost, but that could be via FP16 or INT8 calculations run on the CU clusters.

 

Wrapping things up (for now), the Xbox Series X will support HDMI 2.1 output, allowing for 120Hz 4K output, or even 8K60. There's no DisplayPort output, not that it was expected, but that does limit the capabilities if you were hoping to hook the console up to a traditional PC gaming monitor.

The video encoder/decoder also has support for 4K and 8K AVC and HEVC/VP9 decoding, with AVC (H.264) and HEVC (H.265) encoding support. That should allow for improved streaming capabilities for anyone that wants to broadcast their latest gaming exploits.

 

The presentation occurs at 7pm PT tonight, at which time we will update the article with more information from the live presentation.