GDDR5 vs DDR3 Latency & Bandwidth.
Since the Playstation 4′s and Xbox One’s specs were leaked on the internet (but of course, back then the systems were known as Orbis and Durango), there’s been much debate of the latency of GDDR5 and DDR3 memory.
GDDR5 5500 (1375MHZ actual) memory, in the case of the Playstation 4 (which uses a 256-bit bus) giving a total of 176GB/s of peak bandwidth. As a small note, according to interviews with developers, the actual real number when factoring in overheads and the like is more along the lines of 172GB/s. It’s still far higher than the Xbox One’s DDR3 2133 68GB/s.
The Xbox One’s RAM puts out a peak of about 68GB/s (developers haven’t spoken about actual bandwidth available, so I’ll avoid place a number here). But despite the massive lack of main system bandwidth, the common theory is the following:
The GDDR5 memory had lots of bandwidth, but with this bandwidth came lots of latency. DDR3 makes up the gap somewhat by being snapper in terms of latency – and so, in theory some tasks (involving the PS4′s CPU) could fall foul to such latency.
Is this even true? Well, let’s take a look.
Firstly – the CPU both machines use is the AMD Jaguar. The AMD Jaguar comes with very effective cache, and the CPU is extremely good at predicting which pieces of data that it requires next. The CPU supports Out-Of-Order-Execution and Speculative Execution. Both are large subjects themselves, but they are effective at ensuring that the right data is being processed and is ready to process. Part of the way they do this is the AMD Jaguar’s caching. the Cache in the CPU is large enough to hold instructions, and from what we’re hearing it has a good “hit rate”. Hit rate is basically the chance that relevant data is present in the caches.
CPU’s themselves aren’t that sensitive to latency it appears. A quick Google around the internet will show you that expensive ram kits with extremely tight timings make little difference, even on high end desktop CPU’s.
Benchmarks at Tomshardware for example shows a great story of how DDR3 latency means for little difference.
GPU’s and bandwidth requirements
Okay – so now we’re aware that the CPU is likely not too affected by latency, but what about the GPU?
Both machines use Radeon 7000 series technology (or cores that are pretty damn close). Developers have praised the PS4′s GDDR5 ram which feeds both the CPU and GPU. GDDR5 memory is found in mid to high level GPU’s. The Radeon 7750 is a perfect example of this. This desktop card comes in two flavors: DDR3 and GDDR5. The GDDR5 1GB version kicks the butt of the DDR3 2GB version pretty regularly. A few examples on this would be found at: here and also from AMD’s own website.
It just makes sense. The cards are being starved of data, and so the fast GDDR5 makes all the difference. From PC gamers who remember the original SDR Geforce cards, or the Geforce 2 MX cards, you’ll be familiar with the tale. Not enough data being sent to the GPU = performance suffers.
Remember that high end GPU’s such as the Radeon 7970 use 384 bit buses to ensure there’s enough data being pumped to the GPU. But, they also have almost double the compute units of event he PS4 (32 compute units, vs the PS4′s 18).
It’s possible that because the Xbox One only sports 12 GCN units, that the DDR3 won’t hurt ‘as much’. But remember, the 7750 results linked earlier don’t have as many CU’s as the Xbox One – and performance takes a dive when using DDR3. So, now it’s down to latency.
But what about the actual latency of DDR3 vs GDDR5?
It’s worth remembering that if you do a little bit of Googling around, you can easily come across the specs of various GDDR5 memory. Hynix for example place PDF’s of their products freely available on the internet.
The latency on GDDR5 and DDR3 memory does vary some, but the typical CAS latency of most sticks of DDR3 is between 7 and 9 (see the earlier Tomshardware link for more information). GDDR5 RAM meanwhile is a typical latency of about 15. Once again, all of this depends on the make and models of the RAM.
So – there’s the answer right? Let’s say 8 CAS vs 15 CAS? No, it’s not. We have to remember that the speeds are for all intents and purposes – relative. If you take the CAS of both, and then multiply it by the clock speed – you get the ns of delay. CAS of 15/1.35 = 11ns.
I’ll save you the trouble and say that it’s between 10 – 12 for both DDR3 and GDDR5 varying heavily on the clock speed of either the DDR3 or GDDR5 AND in the timings they’ve chosen.
Also, take into account what Mark Cerny has said about GDDR5 latency:
“Latency in GDDR5 isn’t particularly higher than the latency in DDR3. On the GPU side… Of course, GPUs are designed to be extraordinarily latency tolerant so I can’t imagine that being much of a factor”.
What does it all mean?
It means that the Xbox One doesn’t have a latency advantage really with the DDR3 memory. It’ll have to try and rely on its ESRAM to make up for the lack of bandwidth which the DDR3 is of course, the culprit for. The X1′s ESRAM will have its work cut out, and from what the rumors are so far, the memory requires a lot of manual labour from the developers to get the most out of. It’s currently speculated that it’s being used for both a frame buffer and for tiled resources. However, much of its workings still remain a mystery.
Of course – in this article, we are not taking into account other aspects. These include the PS4′s extra graphical power, the extra compute work of the PS4 or its volatile bit. It’s also not taking into account API’s, drivers and code from developers.