Why I am not excited about DDR4

[kalmquist]

The following table shows the time for a memory read on a page that is already open. The first column is the memory type, the second is the frequency, the third is the CAS latency, and the fourth is the number of nanoseconds required to complete a read operation.

I've marked the DDR3 1600 CAS 9, which is about the least expensive DDR3 memory available. There is slower DDR3 memory available, but it doesn't cost any less. I've also marked DDR4 2133 CAS 12, because 2133 is the fastest DDR4 memory that Intel supports on Haswell-E, and CAS 12 is the shortest CAS latency available at a reasonable price. I have seen DDR4 2133 CAS 10, for an outrageously high price, but even that has higher latency than DDR3 1600 CAS 9.

For sequential scans of memory, the only thing that matters is the frequency, and DDR4 will generally beat DDR3. But for random memory access, latency counts, and the table below shows that DDR3 is generally superior on that measure.

DDR3 2400 10 11.67
DDR3 2400 11 12.50
DDR3 2133 10 13.13
DDR3 2400 12 13.33
DDR3 1866  9 13.93
DDR3 2133 11 14.06
DDR3 2400 13 14.17
DDR3 1600  8 15.00
DDR3 1866 10 15.00
DDR3 2133 12 15.00
DDR4 2933 14 15.00
DDR4 2933 15 15.68
DDR3 2133 13 15.94
DDR3 1866 11 16.07
DDR3 1600  9 16.25  <-- basic DDR3
DDR4 2933 16 16.36
DDR3 1333  7 16.50
DDR4 2666 14 16.50
DDR4 2133 10 16.88
DDR3 1866 12 17.14
DDR4 2666 15 17.25
DDR3 1600 10 17.50
DDR4 2133 11 17.81
DDR3 1333  8 18.00
DDR4 2666 16 18.00
DDR4 2400 14 18.33
DDR3 1066  6 18.75
DDR3 1600 11 18.75
DDR4 2133 12 18.75  <-- DDR4
DDR4 2400 15 19.17
DDR3 1333  9 19.50
DDR4 2133 13 19.69
DDR4 2400 16 20.00
DDR3 1066  7 20.63
DDR4 2133 14 20.63
DDR3 1333 10 21.00
DDR4 2133 15 21.56
DDR3 1066  8 22.50
DDR4 2133 16 22.50
DDR3 1066  9 24.38

To explain how these numbers are computed:
1) The actual frequency is half the listed frequency; for example DDR3 1600 actually operates at 800 Mhz.
2) The data transfer itself takes 4 clock cycles on DDR3 and 8 clock cycles on DDR4.
For example, for DDR3 1600 CAS 9, we have 9+4=13 clock cycles at 800 Mhz. To get the time in microseconds, we divide the number of clock cycles by the frequency in Mhz. We then multiply by 1000 to convert to nanoseconds. 13/800 * 1000 = 16.25 nanoseconds.

 

[ReefaMadness]

While I understand your point, and it is completely valid...today.

What is going to happen is the same thing that happened when we transitioned from DDR2 to DDR3, which is that in time the newer processors and motherboard will all REQUIRE the use of DDR4 so as long as you don't need to make the move to a new platform, you'll be fine...for the short term. However, eventually OS and other software that we currently use will no longer be supported and you'll be forced to decide between keeping your old platform and being limited in what you can do with it, or moving on to whatever hardware is required for the type of computing that you do.

By the way, I'm quite sure that forcing us to move on was always a part of the equation. Intel, MS, and the rest of them can't exist if we aren't buying new components on a regular basis.

 

[BonzaiDuck]

While I understand your point, and it is completely valid...today.

What is going to happen is the same thing that happened when we transitioned from DDR2 to DDR3, which is that in time the newer processors and motherboard will all REQUIRE the use of DDR4 so as long as you don't need to make the move to a new platform, you'll be fine...for the short term. However, eventually OS and other software that we currently use will no longer be supported and you'll be forced to decide between keeping your old platform and being limited in what you can do with it, or moving on to whatever hardware is required for the type of computing that you do.

By the way, I'm quite sure that forcing us to move on was always a part of the equation. Intel, MS, and the rest of them can't exist if we aren't buying new components on a regular basis.

Well, revisit the "requirements" list for -- say, Windows 7-64. Right off, you need a dual core processor with the 64-bit capability, but right there, we're going back to before 2006. All the subsequent OS's so far have supported that hardware, and whatever memory is being used wouldn't matter.

They are two different dominant-firms and "near-" monopolists working with a highly-competitive RAM brand-name with a fewer number of chip suppliers, who could only rely on product obsolescence, product design and marketing hype to create more demand for their basic product. They are in two different markets -- hardware versus software -- but they are so intertwined that new features in hardware require changes to the OS, but the OS would be backwardly compatible to some extent. M$ is not going to look out for narrowing consumer options between old and new Intel hardware; Intel is not going to attempt something that would cause problems with newer OS versions.

And basically, for a change in memory spec, we're only talking about the availability of chipset drivers for some particular OS.

 

[Xtrem]

While I understand your point, and it is completely valid...today.

What is going to happen is the same thing that happened when we transitioned from DDR2 to DDR3, which is that in time the newer processors and motherboard will all REQUIRE the use of DDR4 so as long as you don't need to make the move to a new platform, you'll be fine...for the short term. However, eventually OS and other software that we currently use will no longer be supported and you'll be forced to decide between keeping your old platform and being limited in what you can do with it, or moving on to whatever hardware is required for the type of computing that you do.

By the way, I'm quite sure that forcing us to move on was always a part of the equation. Intel, MS, and the rest of them can't exist if we aren't buying new components on a regular basis.

But, WHY are they pushing for it if it is slower?

 

[BonzaiDuck]

But, WHY are they pushing for it if it is slower?

Without having a crystal ball, or an inside understanding of memory design and production, I'll say that it's a new spec which promises greater performance in the "long run." What you get in the short run may overlap the older spec in performance.

The kits you will likely see are quad-channel -- 4x X-GB modules.

We'll see what happens with any new round of consumer-boards and Broadwell. Who knows?

 

[exar333]

DDR4 uses less power and is needed on the mobile side. Because mobile and desktop generally share the same dies, including memory controllers, it has to be done across the board.

We are hardly memory-bandwidth starved these days anyways, so really that doesn't matter. It does open-up more dense capacity options (16GB/32GB, etc.), which is good. Win-win.

 

[Blain]

I don't have any DDR4 MBs, so I join in your apathy.

 

[hhhd1]

DDR3 1600  9 16.25  <-- basic DDR3

DDR3 1600 11 18.75

that is not correct, ddr3 1600 default jedec timing is 11 and not 9, 9 is achivied by overclocking/XMP

 

[Cerb]

that is not correct, ddr3 1600 default jedec timing is 11 and not 9, 9 is achivied by overclocking/XMP

No, that's stock settings, and the standard has from 8-8-8 to 11-11-11, as of 3C (Wikipedia's link is dead, archive.org lacks it, so I'm not going to link to what I could find, but you can likely find it really easy, too ).

As for the OP's table, and results, welcome to SDRAM, DDR2, and DDR3. This is nothing new. Over time, the stuff gets faster, and approximates the old one's latencies. A slight increase in latency isn't too bad, either, since the added bandwidth makes for the bus using less time on that data.

New RAM is nothing to get excited about. It's something that will enable more performance from CPUs and GPUs over the next several years, not something that is amazing in itself. If you drop a CPU's RAM speeds to typical DDR3 performance levels in 5 years, you'll find it will perform worse than whatever speed DDR4 is common, just like DDR 3 v. DDR2 several years ago (in particular, with AMD's, where you could actually use either, sometimes even on the same board).

 

[kalmquist]

I just realized I overlooked the burst ordering when computing my table. In DDR3, a read requires 8 transfers, which can be accomplished in 4 clock cycles, In DDR4, a read requires 16 transfers, which takes 8 clock cycles. My chart gives the times to perform a complete read, which is one of the reasons that the DDR4 memory is slower than the DDR3. (The other is that DDR4 tends to have higher CAS numbers than DDR3.)

However, Wikipedia tells us that SDRAM supports burst ordering. That means that when a processor wants to read a word from memory, it can request that the value being read be included in the first transfer from memory to the CPU. The CPU can use this value in computations without waiting for the remaining data transfers to complete.

I haven't verified that current Intel and AMD processors take advantage of burst ordering, but it's a good bet that they do. If so, getting to get the DDR4 latencies equal to DDR3 latencies, all memory manufacturers have to do is to get the CAS timings the same at the same frequency, which seems like it should happen eventually.

 

[nenforcer]

It's the same as it was for the transition from DDR2 -> DDR3. i.e. DDR2 800MHz PC2-6400 (or even DDR2 1066MHz PC2-8500) to DDR3 1066MHz or DDR3 1333. The initial bandwidth was nearly identical but the CAS timing were worse for the DDR3 resulting in higher latency. Eventually the bandwidth (DDR3 1600) was easily eclipsing the older DDR2 all the while at lower voltages (1.5).

I'm not touching DDR4 until 2016 at the very earliest after AMD has their Zen chips out and the price has shrunk accordingly for DDR4 chips.

 

[ShintaiDK]

DDR4s 3 main advantages are:
Lower power consumption.
Higher densities.
Faster bandwidth for IGPs.

But latency is a disadvantage.