Hot Chips 32, the (virtual) conference that delves into silicon designs from across the injury, is in full swing this week and AMD is one of the companies teeing up to talk about its hardware. The Xbox Series X has been the subject of much speculation — now, thanks to Microsoft’s presentation, we know more about how the functional blocks of the SoC are arranged.
AMD/Microsoft opted for two quad-core CPU clusters rather than a unified eight-core chip. It would be very interesting to know what the penalty is for a cross-cluster L3 search, but we didn’t get a ton of new information off this slide.
The total size of the Xbox SoC is 360mm2, down slightly from the 375mm2 of the Xbox One and the 367mm2 of the Xbox One X. In the past, process node shrinks resulted in a net cost reduction. Today, that’s no longer the case. The density leap from 16nm to 7nm is equivalent to a full node reduction, per AMD’s comment.
This is the high-level overall feature chart for the SoC. Blue blocks refer to hardware processing blocks, and the last elements referenced are security hardware blocks. The reference to machine learning acceleration is interesting, but there’s no sign of a dedicated AI processing block on the die shot Microsoft released. This may refer to GPU-based processing — or Microsoft could have another surprise or two, I suppose.
As for the GPU:
The remarks on “Ray-box” and “ray-tri” are likely references to ray-box intersection tests and ray-triangle intersection tests. Both of these are crucial steps in ray tracing, though simply knowing these two numbers doesn’t tell us anything about what the solution will be capable of or how it compares with Nvidia hardware. The talk of a minor area cost for a 3x – 10x acceleration is interesting. One thing we don’t know yet about RDNA2 is whether or not it has any specialized hardware blocks for these types of calculations. Spending more die size to improve performance is not a bad tradeoff, though obviously it’s more impressive when the speedup is towards the 10x end of the range.
The RDNA2 diagram Microsoft provided looks like RDNA with the addition of the ray acceleration blocks. Each CU is capable of performing either four texture operations per clock or four ray “ops”, but not both. We don’t know anything about what a ray “op” consists of, but the fact that it’s an either/or situation may explain why Microsoft’s example specifically states ray tracing will be an “economical upgrade” rather than a complete replacement.
Overall, Microsoft’s description of its ray tracing sounds similar to what the PC has achieved to-date with Nvidia’s Turing, though obviously we don’t yet know the specifics of AMD’s technology. It’s entirely possible that RDNA2 will be faster than Turing at ray tracing — but what that might allow AMD to do is pack a more PC-like performance level into a console as opposed to deploying some kind of RT-redefining performance beast.
Still lots we don’t know, but interesting details here. The Xbox Series X is going to pack PC-like performance and PC-like capabilities, from backwards cross-generational capability to at least some games delivered in a “purchase once, play on both” fashion. The hardware inside this system wouldn’t be out of place in a midrange gaming PC. The one big thing we don’t know? The price. Microsoft is keeping mum on that so far.
- Xbox Series X Launches in November, Halo Infinite Delayed
- Microsoft’s Cheaper Next-Gen Xbox Series S Console Confirmed
- Bethesda Pledges Next-Gen Versions of Doom Eternal, Elder Scrolls Online Will be Free to All PS4, Xbox One Owners