Of all the rumors that swirled around Kaveri
before the APU debuted last week
, one of the more interesting bits was that AMD might debut GDDR5 as a desktop option. GDDR5 isn't bonded in sticks for easy motherboard socketing, and motherboard OEMs were unlikely to be interested in paying to solder 4-8GB of RAM directly on to their own motherboards to support AMD's chip. Such a move would shift the RMA responsibilities for RAM failures back to the board manufacturer.
Given the limited upselling capabilities and low market share AMD currently enjoys, it seemed unlikely that Sunnyvale would consider such an option. A deep dive into Kaveri's technical documentation
, however, shows that AMD did indeed consider a quad-channel GDDR5 interface. It must be noted that there's no chance of this capability being functional on current chips -- Kaveri has two memory controllers, not four, and they're designed for DDR3, not GDDR5. Nonetheless, plans were clearly far enough along that AMD considered the option.
The BIOS Guide makes reference to a "GDDR5 Mode:"
It's also possible that this GDDR5 mode could be implemented alongside a DDR3 main memory pool. It's possible that future versions of the Kaveri APU might implement 2x 64-bit DDR3 channels alongside 2x 32-bit GDDR5 channels, with the latter serving as a framebuffer for graphics operations. (AMD references 32-bit controllers at one point in the documentation, which is why this seems to be the most likely configuration, given that four memory controllers are mentioned in Kaveri's spec).
A 64-bit path to a 1GB GDDR5 frame buffer might not seem like a great deal of bandwidth, but GDDR5 clocked at 1500MHz would provide 48GB/s of memory bandwidth. That's an increase of 40% over DDR3-2133. Given that the Kaveri IGP is already bandwidth-limited, a 40% bandwidth improvement could offer a 25-30% IGP performance improvement in many titles. Alternately, AMD
could be eyeing this technology for a mobile variant. If OEMs want to use low-power versions of DDR3, a higher-powered GDDR5 GPU frame buffer that can be power-gated and shut off when not in use could be quite valuable.
Is an eight-core Steamroller in the wings?
The other document making the rounds is AMD's software optimization guide for Family 15h processors. This guide specifically shows an eight-core Kaveri-based variant attached to a multi-socket system. In fact, the guide goes so far as to say that these chips in particular contain five links for connection to I/O and other processors, whereas the older Family 15h chips (Bulldozer
) only offer four Hypertransport links. This has been taken by many to mean that AMD is secretly planning such processors, but the data sheet
for the specific family of products named in the PDF (Family 15h Models 30h - 3 Fh) is specified as having only 2 or 4 cores. In other words, it's possible that the eight-core reference is an old one.
There's no reason AMD can't
build an eight-core Kaveri, but it's not clear if such a part could drop into the current Opteron
sockets. Even if it can, the update would be marginal. Assuming the same patterns of performance hold, Kaveri would make a much better update for AMD's low-power Opteron products than its high-end family. Multi-core scaling would be significantly improved, and clock rates in the low power parts might be higher than we've seen from that family of parts. Still, for all that the A8-7600 improved AMD's low watt game, it didn't provide the company with performance parity against Intel. Without the GPU
hardware to leverage, how strong an option is the new Steamroller
For AMD, the advantage of ramping all-new server or desktop parts has to be balanced against the question of whether or not those parts are likely to extend its advantages in the mainstream market. The GPU core is easily the strongest part of the chip, which means it may not make much sense to push out an eight-core CPU variant if the gains are only 5-10%. AMD's share of the 4S market where a fifth CPU link might be most beneficial is rather small.
Between the two, a GDDR5-equipped mobile variant or even a future GDDR5 chip based on 2015's Carrizo update seems to make the most sense. I'm not sure the server gains will be great enough to justify a comprehensive update to the family, though we won't know for certain until the upcoming quad-core / dual module Opterons are established in the channel.