GDDR5 uses a feature called Clock Data Recovery, where the memory clock speed is actually generated out of the data stream. Upon power-up, a system using GDDR5 memory actually "trains" the interface between the memories and the host device (the graphics chip, in this case). Command and address clocks are synced up, the data clock is derived from the signal on memory reads, the data/address/command clocks are de-skewed to align signals properly. This "training" process helps produce cleaner signaling and is one way the standard achieves high clock rates.
It also makes designs that use GDDR5 more tolerant of different trace wire lengths than GDDR3 or GDDR4. Currently, board designs using high-speed memories are tricky. A lot of care has to be taken to make sure the trace wires are all the same length, so they're a mess of zig-zagging routes. Combine this with the extra wires and board layers required for wider memory busses, and graphics cards get expensive, difficult to build, and prone to failure. The clock training system of GDDR5 should, in theory, help alleviate some of that.
PCB Routing
click on image for full view
We've only just scratched the surface of the tricks GDDR5 uses to provide more bandwidth without increasing pin count or voltage. For instance, the protocol features error detection (but not correction) in both read and write directions, allowing the interfaces to re-request a bad error transmission. This boosts reliability at high clock speeds, and should be fun for overclockers (you might, in some cases, boost to a higher clock speed but see lower real-world bandwidth as you introduce errors that need to be re-requested).
Address and data bus inversion lowers power consumption. Typically in a data stream, you say that positive voltage is a "1" and no voltage during a clock cycle is a "0". Well, if you have more 1s than 0s in a data transmission block you simply send along an addition bit saying "flip that so that the no-power states are 1s and the power states are 0s." This means that no data or address block will ever have more than half of its signal containing voltage, regardless of what data is in there. Continued...
One of the key issues, from a board design perspective, is how man chips, and how they'll lay out. GDDR5 will support chips with 16 DRAM banks, enabling fewer, higher capacity chips, which should simplify board designs a bit.
What we really want to know, from the end-user perspective, is where the rubber will meet the road. Exactly how much bandwidth will we get? How much will it cost? What products will use it, and when will we be able to buy them?
Bandwidth first: A system using GDDR3 memory on a 256-bit memory bus running at 1800MHz (effective DDR speed) would deliver 57.6 GB per second. Think of a GeForce 9600GT, for example. The same speed GDDR5 on the same bus would deliver 115.2 GB per second, or twice that amount. Take any GDDR3 bandwidth on a given clock rate and bus width and double it, and you get GDDR5's bandwidth. Of course, the marketing guys love big numbers and would undoubtedly not call it 1800MHz, just as 1800MHz GDDR3 is really running at 900MHz. Expect the marketing guys to call memory at that speed 3200MHz.
When will we see it in products? AMD has confirmed that at least some of their upcoming "700 series" graphics cards, surely called the Radeon HD 4000 series, will use GDDR5. It hasn't said what the clock speed will be, or which of the products will use it. There's also no word on availability yet, but expect products "soon." GDDR5 is on track to be ratified by JEDEC later this summer (products with new memory types often ship before final JEDEC ratification). Nvidia has not yet divulged their GDDR5 plans, if any, and they're unlikely to talk about it at all if their rivals at AMD will be first out of the gate.
What about cost. This stuff is going to cost a fortune, right? Well, yes and no. High-speed GDDR3 and GDDR4 memory is certainly expensive. We're told to expect GDDR5 to initially cost something like 10–20% more than the really high speed GDDR3. Of course, you don't buy memories, you buy a graphics card. Though the memory will cost more, it will be offset somewhat in other places on the product you buy. You don't need as wide a memory interface which means a smaller chip with fewer pins. The board doesn't need to contain as many layers to support wider memory busses, and the trace wire design can be a bit more simple and straightforward, reducing board design costs. As production ramps up, GDDR5 could be as cost effective overall as GDDR3. It will only be appropriate for relatively high-end cards at first, but should be affordable enough for the $80–150 card segment over the next couple years.
