AMD Considered GDDR5 For Kaveri, Might Release Eight-Core Variant

MojoKid writes "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. Such a move would shift the RMA responsibilities for RAM failures back to the board manufacturer. It seemed unlikely that Sunnyvale would consider such an option but a deep dive into Kaveri's technical documentation shows that AMD did indeed consider a quad-channel GDDR5 interface. Future versions of the Kaveri APU could potentially also implement 2x 64-bit DDR3 channels alongside 2x 32-bit GDDR5 channels, with the latter serving as a framebuffer for graphics operations. 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 and Piledriver) only offer four Hypertransport links."

Latency vs bandwidth

[evilsofa]

DDR3 is low latency, low bandwidth. GDDR5 is high latency, high bandwidth. Low latency is critical for CPU performance while bandwidth doesn't matter as much. On video cards, GPUs need high bandwidth but the latency doesn't matter as much. This is why gaming PCs use DDR3 for system RAM and GDDR5 on their video cards. Video cards that cut costs by using DDR3 instead of GDDR5 take a massive hit in performance. The XBox One and PS4 use GDDR5 shared between the CPU and GPU, and as a result have the rough equivalent of a very low-end CPU paired with a mid-range GPU.

Re:Latency vs bandwidth

[Anonymous Coward]

Somewhat false. Latency is approximately the same for DDR3 vs GDDR5, at least in terms of nanoseconds from request to response. GDDR5 is impractical for use in CPUs due to the need to solder to a board and high power consumption (enough to need cooling). That CPUs don't need so much bandwidth just makes it useless in addition.

Also, the Xbox One uses DDR3, and makes up for the lack of bandwidth with four channels and 32MB of ESRAM for graphics use.

A Better Explaination At Anandtech

[Anonymous Coward]

Anandtech's writeup [anandtech.com] (which Hothardware seems to be ripping off) has a much better explanation of what's going on and why it matters.

Let me be very clear here: there's no chance that the recently launched Kaveri will be capable of GDDR5 or 4 x 64-bit memory operation (Socket-FM2+ pin-out alone would be an obvious limitation), but it's very possible that there were plans for one (or both) of those things in an AMD APU. Memory bandwidth can be a huge limit to scaling processor graphics performance, especially since the GPU has to share its limited bandwidth to main memory with a handful of CPU cores. Intel's workaround with Haswell was to pair it with 128MB of on-package eDRAM. AMD has typically shied away from more exotic solutions, leaving the launched Kaveri looking pretty normal on the memory bandwidth front.

It's also worth noting that the Anandtech article implies that AMD is still on the fence on Kaveri APUs with more memory bandwidth, and that it may be something they do if there's enough interest/feedback about it.

Re:news for nerds

[SuricouRaven]

Because it's electrically so delicate that you can't keep bit sync when shoving such high frequencies through a slot connector. The price of higher bandwidth, in both the analog and digital senses.

Re:AMD could do a 24 core desktop chip right now

[symbolset]

Fine. You do it your way. I want this technological achievement so I can do it mine.

Re:Latency vs bandwidth

[Bengie]

DDR3 and GDDR5 have nearly the same latency when measured in nanoseconds. When measured in clock cycles GDDR5 is higher latency, but it has more cycles. This has to do with the external interface, which is more serial in nature than the internal. The data path for the internal is quite wide, but the external datapath is not because traces on a motherboard are expensive. To crank up the bandwidth, you increase the frequency. If the internal frequency remains fixed and the external frequency goes up, the external "latency" seeming goes up.