Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
DDR4 to trade latency for throughput and power consumption?
Adata has announced the first DDR4 memory officially available:
http://anandtech.com/show/8315/adata-formally-announces-ddr42133-cl15-udimms
You can't actually use it yet, as there aren't any CPUs available with a DDR4 memory controller. That will change shortly, of course, as Haswell-E is scheduled to launch sometime within the next two months.
But what's interesting to me about this announcement is the specs: 2133 MHz, yes, but latency timings of 15-15-15. Let's just take the first one, CAS 15, for simplicity.
As a quick reminder, memory latency timings are in numbers of clock cycles. To get the latency in nanoseconds, you take the latency timing and divide by the clock speed (Hz = 1/s), and then multiply by 2 to compensate for the DDR doubling. In this case, the memory has rated CAS latency of 14.0625 ns. That last digit isn't significant, of course, but it's a hair over 14 ns.
For comparison, New Egg has a lot of 2133 MHz DDR3 memory kits. The slowest of them has CAS 11, for a latency of slightly over 10.3 ns. And remember, that's the slowest 2133 MHz DDR3, not the average. Now, 2133 MHz is pretty high clocked for DDR3 (the highest bin of memory that the fabs sell), but the most common speed of memory on New Egg is 1600 MHz, and the most common latency for 1600 MHZ DDR3 is CAS 9--which comes to slightly over 11.1 ns latency. Indeed, the slowest 1600 MHZ memory there is CAS 11, which is 13.75 ns latency--still faster than the first DDR4 kit. And that's cherry-picking slow DDR3, even.
In all, over 99% of the DDR3 kits on New Egg have faster CAS latency (in nanoseconds) than A-data's just announced DDR4 kit. If you look at latency timings, it's even worse; the highest CAS latency on New Egg is CAS 13--and that's for 3200 MHz DDR3. If willing to pay a premium for low-latency memory, you can get CAS 9 2400 MHz DDR3--for latency of 7.5 ns.
Now, it has long been the case that memory latency didn't improve nearly as fast as throughput. Electrical charges pass through copper traces however fast they go, and that is a meaningful factor when considering memory access times. The most common speed of DDR2 on New Egg is 800 MHz, and the most common CAS latency for that speed is 5. That comes to 12.5 ns latency, which is fairly slow for DDR3--but still faster than the new DDR4. But 800 MHz CAS 6 DDR2 is pretty common, too, and that's 15 ns latency. 667 MHz CAS 5 is also common, which gives the same 15 ns latency.
So while being higher latency than nearly all DDR3, the new DDR4 kit is lower latency than the bulk of DDR2. But that only highlights that in the past, latency has tended to decrease: DDR3 tends to be lower latency than DDR2, which was lower latency than DDR.
Does that mean that DDR4 is worthless? No, of course not. It's still higher throughput, as the bottom bin DDR4 is the same 2133 MHz as the top bin DDR3, and it's only going to go up from there. A stock voltage of 1.2 V for DDR4 as compared to 1.5 V for DDR3 means huge reductions in power consumption, too. In desktops, that doesn't matter much, but in laptops or data centers, it sure does. But in the past, new memory standards have typically meant that things got better all around. This time, that's not the case, as latency is getting worse.
Now, system memory latency isn't as important as it used to be. Moore's Law has offered CPUs vastly more transistors as time passed, and that means room for spacious caches. Memory throughput isn't a meaningful restraint in most desktops and laptops unless you're trying to feed integrated graphics, and GPUs are very much built for throughput, not latency, so 2133 MHz DDR4 will handily beat 1600 MHz DDR3 even if it is CAS 15 versus CAS 9. But I still think it's interesting to note--and I regard it as unlikely that A-data would start by announcing their low-end cheap junk while keeping the higher end kits that will launch at the same time secret until later.
Comments
Well yes, higher clock speed means higher throughput. If the goal is to transfer as much data from memory to the CPU as you can in 1 ms, then typical DDR4 will beat typical DDR3 handily. But what if the goal is to transfer 4 bytes from system memory to the CPU--which is, after all, a very common use case? In that case, there's no guarantee that typical DDR4 will beat typical DDR3; it could easily be the other way around.
its long but well
jogos de moto
http://anandtech.com/show/8348/adata-officially-launches-xpg-z1-ddr4-memory
Apparently that was A-data's bottom of the line part. The others are:
2133 MHz CAS 13 (12.2 ns latency)
2400 MHz CAS 16 (13.3 ns latency)
2800 MHz CAS 17 (12.1 ns latency)
For comparison, the 1600 MHz CAS 9 DDR3 that I bought in 2009 because it was the cheapest available is 11.25 ns latency.
The first DDR4 memory is now listed at New Egg. It's made by Crucial, and there are two speeds:
2133 MHz CAS 16 (15 ns latency)
2400 MHz CAS 16 (13.3 ns latency)
High latency, indeed.
Interestingly, New Egg lists the memory as having a release date of August 29. Good chance that's the Haswell-E release date. Crucial doesn't have any reason to delay memory sales, but New Egg might want to delay it until there's a product that can use it, and Haswell-E will be the first.
Interestingly, it's very expensive: two 4 GB modules come to $134 for 2133 MHz or $180 for 2400 MHz. With Haswell-E having four memory channels, there's no real need for a ton of bandwidth, either. It will be interesting to see how prices drop as time passes.
Well, DDR4 will have much higher density... 16gb stick is very likely at some point. Internal Vref is important too. Bandwidth will increase for a number of reasons.
IMO it is all about replacing the SSD during operation at this point, large data transfers to caches or video cards, latency is loosing some of its importance.
Yes, it will mean lower power consumption. That's the point.
Power consumption is heat generation. That's where the power that is consumed goes. It can't vanish entirely, as conservation of energy applies here, too. Any part that consumers power is going to put out heat.
But the better question, and likely what you were getting at is, is the power difference enough to actually matter? And there, the answer is, it depends. In laptops, it will; remember that DRAM is volatile, so if it's on, it has to be consuming power all the time, even at idle. Reduce idle power consumption and you increase battery life. In data centers, it will also matter: reduce power consumption of something by 1 W, and multiply that by 10,000 of the something in a warehouse and you've made a difference of 10 kW.
The reduced power consumption would matter in tablets and cell phones, too, if DDR4 were to be used there. Which it likely won't be, as those tend to use different memory standards such as LPDDR3--which wasn't even defined until 2012 and is not at all the same as DDR3--that optimize more for reduced power consumption even at the expense of bandwidth.
As with many other computer components, there are trade-offs between performance and power consumption. There are different memory standards to handle different choices in the trade-off. If you want high performance and are wiling to burn power to get it, you go with GDDR5, as video cards meant for gaming generally do. If you need low power consumption, you take LPDDR3, or until recently, LPDDR2. If you want somewhere in the middle, you go with DDR3 today, and soon DDR4. There's also DDR3L, which is basically a lower voltage (and hence lower power) version of DDR3, for situations that need somewhere between standard DDR3 and LPDDR3.
The stock voltage of DDR2 was 1.8 V, DDR3 is 1.5 V, and DDR4 is 1.2 V. That that voltage is decreasing is not a fluke. Incidentally, DDR3L is 1.35 V.
All else equal, a computer chip's power consumption will be proportional to clock speed * voltage^2. That the contribution of voltage is squared means that lower voltage is a big deal. Now, all else is not equal, of course, and there are a lot of reasons why even two memory modules with exactly the same specs can have very different power consumption.
For any voltage, there is some maximum clock speed at which a chip can run and be stable. Well, provided that the voltage isn't so low that it won't be stable at any clock speed, or so high that heat becomes a major problem or the chip fries outright. Higher voltages tend to allow higher clock speeds. That you need higher voltages for higher clock speeds is why power consumption goes up a lot faster than clock speed when you try to overclock things.