Microsoft’s unveil of Project Scorpio at E3 this week was both expected and exciting. We knew Microsoft had to have something in the works once Sony announced the PlayStation 4 Neo; there’s no way Redmond could fail to answer that challenge. What Microsoft actually announced, however, could leapfrog the PlayStation 4 Neo in some significant ways.
First, while there was some initial speculation over whether or not AMD would provide the silicon for the next-generation Xbox, the company ended that with a tweet yesterday.
Here’s what we know about the console: It features an eight-core CPU, over 320GB/s of memory bandwidth, a new GPU, and 6TFLOPS of processing power — far more than the 4.2 TFLOPS rumored to be inside the PlayStation 4 Neo. The question is, how will Microsoft hit these figures, and will the company take an evolutionary path or attempt something more revolutionary?
The evolutionary path is simple and has already been sketched out by the PlayStation 4 Neo’s rumored specs. Microsoft could have paid AMD to do a 14nm or even a 10nm die shrink on the existing Xbox One SoC, keeping the Jaguar CPU architecture but increasing its clock speed. The 6 TFLOPS figure isn’t hard to figure: AMD’s RX 480 is already expected to deliver over 5 TFLOPS in a 150W power envelope. Push the clock on the RX 480 up to 1275MHz and add in the GFLOPS from a 2.1GHz Jaguar, and the Xbox One Scorpio would edge across the 6TFLOPS line. Power consumption in this architecture would likely be ~200W, but while that’s considerably higher than the Xbox One or PS4, it’s roughly in line with the first-generation PlayStation 3, which drew roughly 200W at the wall while gaming.
The big question here is whether or not Microsoft ditched the Xbox One’s ESRAM cache or not in the new version of the chip. It’s difficult to imagine them ditching the cache if they want to retain seamless backwards compatibility, since the programming guides for the Xbox One explicitly recommends that game developers utilize this cache to avoid putting pressure on the Xbox One’s limited memory bandwidth. As we’ve covered elsewhere, DDR4-3200 would give the Xbox One Scorpio roughly 102GB/s of main memory bandwidth. The rest of the memory bandwidth could come courtesy of a larger ESRAM cache with a higher clock.
The Xbox One’s ESRAM cache must be clocked around 850MHz in order for its 109GB/s uni-directional bandwidth to make sense (850MHz * 1024 bits / 8 = 108.8GB/s). Were Microsoft to increase the clock to 1.7GHz, that would give it 218GB/s of cache bandwidth, backed up by 102GB/s of main memory, or just barely over 320GB/s in total. Microsoft could expand the ESRAM cache size at the same time to enable higher resolutions; a 64MB cache is well within the capabilities of a 14nm process.
This evolved hardware might not be quite capable of 4K gaming natively, but could probably handle 1080p 60Hz no problem, or 1440p with a 4K upscaler.
Over at Eurogamer, they’ve got a different hypothesis. They predict that Microsoft will actually leap for a GPU based on Vega running at a relatively low clock speed but with a heck of a lot of compute capability — 3584 – 3840 cores, clocked at 800-850MHz, backed up by 12GB of GDDR5 with a 384-bit memory bus. They think Microsoft will dump the ESRAM cache option altogether and walk away from its DDR-based memory subsystem, in favor of a Sony-style unified memory pool.
I respect the guys at Eurogamer a great deal, and for all I know they’ve got a well-placed mole in Redmond feeding them specs the rest of us aren’t privy to, but I think this is comparatively unlikely. That’s speculation on my part, but it’s based on what we know of AMD’s plans for its next-generation architecture and the path the company took with both the Xbox One and PlayStation 4. Both consoles are clear adaptations of technology AMD already had in development — both use separate four-core SoC blocks, for example, rather than a unified eight-core SoC. While both chips are customized to varying degrees, both also rely on a common CPU and GPU architecture.
Vega, AMD’s next-generation high-end architecture, is based on HBM2, not GDDR5(X). There’s an enormous difference between a GDDR5-style memory architecture and HBM2, and neither AMD nor Nvidia has given any sign that they intend to implement support for both standards in the same chip. This means AMD would have to do a completely new memory controller redesign Microsoft’s GPU. It would mean Microsoft would have to do some fancy footwork to ensure games that were designed for the Xbox One and its 32MB ESRAM cache still ran well on the Xbox One Scorpio, particularly given GDDR5’s different latency timings.
Is this impossible? No — which is why I’m not going to say that Eurogamer is wrong. One could argue that by fixing some less-than stellar design decisions of the Xbox One now, Microsoft gives itself more room to launch future consoles along similar lines at some point in the future.
One point I do want to make, however, is that FLOPS remains an utterly terrible metric for evaluating gaming performance. Eurogamer states that “Based on existing AMD Radeon technology, the bottom line is that 6TF of GPU power isn’t enough to power a convincing 4K experience. AMD’s R9 390X offers around 5.9TF and struggles to push 4K resolution at anything like 30fps on modern PC titles.”
FLOPS can be used as a rough stand-in for gaming performance only within the same architecture. Eurogamer thinks that the next-generation Scorpius uses AMD’s Vega, and Vega is a complete departure from previous the previous GCN uarch. For proof of this, consider the GTX 580 (1.58TFLOPS of single-precision floating point) compared to the GTX 650 Ti Boost (~1.5TFLOPs of single-precision floating point). As Anandtech’s Bench figures show, the GTX 580 is regularly ~30% faster than the 650 Ti Boost, even though they have roughly equivalent GFLOPS metrics. Unless the underlying architectures are extremely similar, GFLOPS simply isn’t a good metric to rely on.
If the Xbox One Scorpio is based on an all-new CPU and GPU architecture, as Eurogamer posits, than the chance that it can hit 4K gaming goes up, especially since the console isn’t expected for another 18 months and TSMC’s 10nm process should be ready to ship by then (even if the gains are expected to be small in comparison to 16nm FF). If Microsoft opted for the more restrained design, we’ll probably see something much more similar to the PS4 Neo, albeit with enough extra oomph to (Microsoft hopes) give Redmond a decisive edge.