For months, Sony and Microsoft fanboys have lined up to hurl insults at each other over which console would pack more hardware, hit higher performance targets, or prove a better design for the next generation. With the two consoles launched, the game-to-game comparisons have mostly come out a wash, with a slight edge for the PS4. But there’ve still been questions about the underlying chip design — which architecture is more efficient, and what unique sauce went into each console?
The fine folks at Chipworks have completed their teardown of the Xbox One and given us an answer to that question — and a few puzzles to go with it.
The Xbox One die is 363 square millimeters, up from the PS4′s 348 sq mm. The 5% additional space, despite having the smaller GPU core, is mostly due to RAM. The Xbox One contains a whopping 47MB of on-die RAM, and that pushes the die size up considerably. It’s also why Microsoft didn’t have room on the APU for a larger GPU.
There are some interesting differences to explore. First, consider the Xbox One’s Jaguar CPU blocks. Like the PS4, it has two quad-core chips — but the Xbox One has a bit of circuitry hanging off the CPU that the PS4 lacks. Here’s a comparison of the Xbox One and PS4 CPU islands. We had to rotate the blocks to line them up identically, which is why the label is reversed.
See the block in red? The PS4 doesn’t seem to have an equivalent. What it actually does is unclear. It’s a bit large to be the built-in audio or the IOMMU that HSA theoretically requires. There’s nothing analogous on any of the Kabini floor plans we’ve ever seen.
(It’s also possible that this is a Photoshop artifact or deliberate obfuscation. Companies often mask details on die shots. )
Now, over to the GPU. Like the Sony PS4, the Xbox One contains more Compute Units than are actually active on the console. The chip has 14 CUs, 12 of which are turned on, while the PS4 has 18 active CUs out of a 20 on-die. These are disabled to improve yield. Whether Sony or Microsoft might one day choose to enable the CUs in future console versions is an unknown — typically console manufacturers don’t update core specs post-launch, but consoles have been trending towards greater upgradeability over the past two generations. It’s not impossible that this could change.
The other mystery? The Xbox One GPU cores are physically shorter than the PS4′s equivalents. I don’t mean the GPU block, which is obviously smaller — one GPU Compute Unit on the PS4 diagram, is 50 pixels wide, 395 pixels tall. On the Xbox One, each Compute Unit is 42 pixels wide, 347 pixels tall. It looks as though Microsoft may have picked a tighter arrangement for its GPU core, again possibly to save the maximum amount of space and make room for as much SRAM on die as possible.
Speaking of SRAM, the arrangement of the Xbox One’s was a considerable puzzle when Microsoft unveiled the console architecture. According to the company, the Xbox One doesn’t really have a 32MB contiguous cache, but four 8MB cache blocks instead. There are two blocks of cache to the right of the GPU and a smaller block to the left. This smaller block is possibly used for cross-CPU communication.
It’s hard to tell exactly how the Xbox One’s 47MB of claimed SRAM fit into the floor plan, however. We know that the CPUs in question contain a total of 512K of L1 and 4MB of L2. If the two blocks to the right are ESRAM, each block should be 16MB, for a total of 32MB of cache there. The GPU should contain 512K to 1.5MB of L2 (512K being standard for a GCN chip of this size, with more L2 if Microsoft choose to boost that capability), and about 224K of L1 in total.
That leaves about 10MB of cache missing. If the SRAM block between the two CPUs is that large, it’s far more dense than the SRAM to the right of the GPU.
Chipworks also tore into the Xbox One controller, but it’s not that interesting. It has an ultra-low power Freescale microcrontroller and a Cortex-M0+ core. A custom Microsoft WiFi chip handles communication with the mother ship. The chip count here is kept minimal to speed manufacturing and lower cost. A teardown of Kinect should be up and available in the not-too-distant future.
Different designs lead to similar places
After looking at both the Xbox One and PS4, I think we see companies arriving at the same point through rather different approaches. Both manufacturers chose architecture they felt would allow them to work most effectively. Microsoft invested more silicon in large, low latency caches, while Sony sank more money into raw bandwidth. As far as performance is concerned, this could well end up a tie; as the Xbox One should be able to access data more quickly, while the PS4 can stream sustained data far more effectively. Since game developers can leverage both of those features, the final result could be a wash.
Both companies also picked designs that should be relatively easy to migrate to new process nodes. As 20nm technology comes online, we’ll probably see refreshes in 12-18 months. It won’t surprise me if the first SSD designs start to pop up then, too — there’s too much potential upside in a premium SKU with solid state storage for either company to ignore the possibility.