Why Is DDR4 RAM Still Slower Than DDR3 Overall (Mainly Due To Higher CAS Latency)?

[jerubedo]

So it's been almost 5 years since the launch of DDR4 and I'm kind of surprised that we still don't have great CAS latency on most kits. Basically, DDR3 2200MHZ CL7 is still the best overall RAM given that it's able to transfer the first word to the CPU in 7.82ns, the fourth word in 7.73ns, and the eight word in 9.55s. The only place DDR4 is able to claim a victory is with the 4600MHz CL18 kit and the 4800MHz CL 19 kit, but even then it only wins on the eight word transfer, and one can argue that the first word transfer is the most important because the words are sent in critical word first order which allows the CPU to begin work immediately using that first word. The chart below demonstrates what I'm saying. I've also included my own kit at the bottom, 3200MHz CL14. It wasn't a cheap kit, either, and it's handily beaten by both the DDR3 2200MHz CL7 kit and the 2133MHz CL7 kit in first word, fourth word, and eight word, and both of those kits were similarly priced to my kit. It seems the only real advantage to DDR4 is the lower power draw. So does anyone know WHY the CAS latency is so much higher on DDR4? Why don't we have DDR4 2133MHz CL7 kits when we had DDR3 2133MHz C7 kits? I can only guess it's because of the lower power?

 

Generation	Type	Data rate	Transfer time	Command rate	Cycle time	CAS latency	First word	Fourth word	Eighth word
DDR4 SDRAM	DDR4-4600	4600 MT/s	0.217 ns	2300 MHz	0.435 ns	18	7.82 ns	8.48 ns	9.35 ns
DDR4 SDRAM	DDR4-4800	4800 MT/s	0.208 ns	2400 MHz	0.417 ns	19	7.92 ns	8.54 ns	9.38 ns
DDR3 SDRAM	DDR3-2200	2200 MT/s	0.455 ns	1100 MHz	0.909 ns	7	6.36 ns	7.73 ns	9.55 ns
DDR4 SDRAM	DDR4-4600	4600 MT/s	0.217 ns	2300 MHz	0.435 ns	19	8.26 ns	8.91 ns	9.78 ns
DDR3 SDRAM	DDR3-2133	2133 MT/s	0.469 ns	1066 MHz	0.938 ns	7	6.56 ns	7.97 ns	9.84 ns
DDR4 SDRAM	DDR4-3200	3200 MT/s	0.313 ns	1600 MHz	0.625 ns	14	8.75 ns	9.69 ns	10.94 ns

[LogicWeasel]

Since you're comparing <2ns of difference in most cases, does this matter in the grand scheme?  Is there a noticable user case where downgrading to DDR3 makes a lot of sense?

I don't get why First word is being given importance compared to overall benchmark performance on different memory kits/speeds.

[Jurrunio]

because bandwidth is more important than latency?

[jerubedo]

48 minutes ago, LogicWeasel said:

Since you're comparing <2ns of difference in most cases, does this matter in the grand scheme?  Is there a noticable user case where downgrading to DDR3 makes a lot of sense?

I don't get why First word is being given importance compared to overall benchmark performance on different memory kits/speeds.

Well, it's not less than 2 ns in most cases. The chart I gave compared really high end RAM. Your "average" RAM is DDR4 3000 CL 16, which is 10.67 ns for first word transfer compared to similarly priced high end DDR3 which can do that in 6.36ns. That's a difference of over 4 ns. First word is given priority because it is the most true measure of throughput. 

 

As for benchmarks, I don't see any that compare faster DDR3 RAM to slower DDR4 RAM on the same CPU. But even if we could find one, DDR3 is only used on older processors where the CPU might actually be the bottleneck at this level of RAM speed, making the results between the two functionally the same (if I'm right about the CPU being the bottleneck).

 

However, I'm mostly talking about on-paper, not necessarily real world scenarios. On paper the throughput of DDR3 is generally better because of way better latency.

 

And if we are talking real world, then I can assure you there would be a difference between DDR4 2133MHz C7 vs DDR4 2133MHz C14. I really just want to know why we've seemingly fallen backwards on latency.

[jerubedo]

1 hour ago, Jurrunio said:

because bandwidth is more important than latency?

No it's not, that's what the first word transfer demonstrates (and the fourth and 8th word). That's the measurement of THROUGHPUT, which is the overall metric for RAM. Better throughput = better (faster) RAM.

 

Just to give you an idea on how much timings can affect benchmarks see here. The 3200MHz C14 SHOULD beat the 3000MHz C16 but in several benchmarks it doesn't! It turns out it was the SUB-TIMINGS, not even the primary timings:

 

https://www.techpowerup.com/reviews/GSkill/F4-3200C14Q-32GTZSW/6.html

[jerubedo]

Oh, and just for reference, it looks like the lowest latency on the market right now is 3200MHz CL13, which puts it at 8.125ns for first word transfer, functionally beating 4600MHz CL19 (but still beaten by 4600MHz CL18).

[porina]

It is important to note there are JEDEC standard timings, and then there are "overclocker" type modules offered by more gaming orientated companies. DDR3 2133 is only defined to CAS 11. Anything faster than that is non-standard. Having said that, DDR4 2133 is only defined down to CAS 14. 

 

The long standing argument over if bandwidth or latency matters more still to this day boils down to, it depends on what you do with it. My personal use cases are primarily bandwidth limited. For a quad core consumer Intel CPU, I consider 3200 ram to be inadequate, and estimate 4000 would be getting towards "practically unlimited" bandwidth. I don't actually own a 4000 kit yet although I'm kinda shopping for one when the right price is in stock as I don't want to overpay for it. In the same testing, I've tried varying latency up and down, and it makes hardly any difference compared to bandwidth. I am aware of different applications that can be more latency sensitive, but since I don't run them, it doesn't matter to me.

 

For the standard timings, the best cycle timing of each speed tends to work out to about the same elapsed time. So while faster kits don't necessarily reduce timing, you still gain from the bandwidth part.

 

If you care about latency, you can do what I did, ONCE. That is to manually optimise ALL the timings available in ram. This is extremely time consuming as one wrong step anywhere will lead to instability. You will reboot a lot. You risk corrupting the Windows install. But if you make it through this, the benchmark latency results will destroy ANY off the shelf kit. It took me about a week of evenings to dial in my B-die kit, and even then I made a mistake somewhere as it isn't 100% stable and my life isn't long enough to go back and find where. Some errors are not easily detectable and only crop up after extensive testing. Because of the pain involved, this is probably only something for competitive overclockers looking for every last bit of performance. For normal users, buy any old fast kit and it'll be good enough.

[jerubedo]

14 minutes ago, porina said:

It is important to note there are JEDEC standard timings, and then there are "overclocker" type modules offered by more gaming orientated companies. DDR3 2133 is only defined to CAS 11. Anything faster than that is non-standard. Having said that, DDR4 2133 is only defined down to CAS 14. 

 

The long standing argument over if bandwidth or latency matters more still to this day boils down to, it depends on what you do with it. My personal use cases are primarily bandwidth limited. For a quad core consumer Intel CPU, I consider 3200 ram to be inadequate, and estimate 4000 would be getting towards "practically unlimited" bandwidth. I don't actually own a 4000 kit yet although I'm kinda shopping for one when the right price is in stock as I don't want to overpay for it. In the same testing, I've tried varying latency up and down, and it makes hardly any difference compared to bandwidth. I am aware of different applications that can be more latency sensitive, but since I don't run them, it doesn't matter to me.

 

For the standard timings, the best cycle timing of each speed tends to work out to about the same elapsed time. So while faster kits don't necessarily reduce timing, you still gain from the bandwidth part.

 

If you care about latency, you can do what I did, ONCE. That is to manually optimise ALL the timings available in ram. This is extremely time consuming as one wrong step anywhere will lead to instability. You will reboot a lot. You risk corrupting the Windows install. But if you make it through this, the benchmark latency results will destroy ANY off the shelf kit. It took me about a week of evenings to dial in my B-die kit, and even then I made a mistake somewhere as it isn't 100% stable and my life isn't long enough to go back and find where. Some errors are not easily detectable and only crop up after extensive testing. Because of the pain involved, this is probably only something for competitive overclockers looking for every last bit of performance. For normal users, buy any old fast kit and it'll be good enough.

 

I agree with this 100%. Many applications prefer bandwidth to latency. My question was more oriented at the on-paper specifications. My main question was more so the below

 

50 minutes ago, jerubedo said:

So does anyone know WHY the CAS latency is so much higher on DDR4? Why don't we have DDR4 2133MHz CL7 kits when we had DDR3 2133MHz C7 kits? I can only guess it's because of the lower power?

 

Overclocked or not, DDR3 was able to achieve way lower CAS latency at the same speeds (2133 vs 2133). So why can't we have the best of both worlds with DDR4: higher bandwidth WITH lower CAS latency? What is the driving force preventing this? Just the fact that DDR3-2133MHz is defined to CAS 11 whereas DDR4-2133MHz is defined to CAS 14 confuses me.

[porina]

2 minutes ago, jerubedo said:

Overclocked or not, DDR3 was able to achieve way lower CAS latency at the same speeds (2133 vs 2133). So why can't we have the best of both worlds with DDR4: higher bandwidth WITH lower CAS latency?

I don't have the answer to this, but I have a guess: Voltage. One of the drivers of each generation of DDR is a drop in voltage and consequently power. With DDR4 we run at 1.2v at lower speeds, with 1.35v being more common for the less extreme higher speeds about around 2800. From memory, DDR3 was 1.5v standard and even low power DDR3 was 1.35v. A common trick for ram overclockers is in order to get timings down, you have to really jack up the ram voltage. Maybe they simply decided when setting the standard the tradeoff for most people to lower voltage, and therefore power consumption, was more important than slightly lower latency in some cases.

[jerubedo]

2 minutes ago, porina said:

I don't have the answer to this, but I have a guess: Voltage. One of the drivers of each generation of DDR is a drop in voltage and consequently power. With DDR4 we run at 1.2v at lower speeds, with 1.35v being more common for the less extreme higher speeds about around 2800. From memory, DDR3 was 1.5v standard and even low power DDR3 was 1.35v. A common trick for ram overclockers is in order to get timings down, you have to really jack up the ram voltage. Maybe they simply decided when setting the standard the tradeoff for most people to lower voltage, and therefore power consumption, was more important than slightly lower latency in some cases. 

Yep, that was my guess as well

[Tristerin]

Posts like these (people way more versed than I am on a topic) are why I love this forum.  I just read all of that and I will easily admit I have no idea whats going on here - However I am more educated than I was 15 minutes ago.

 

Continue please...lol

[Galaxyrunner77]

I read once(don't quote me here lol) its due to speed. CAS is based on clock cycles not real time, my 3770k was around a two thirds of the clock speed of my 9900k, so with faster clock cycles and higher latency, it should be similar real time. I was curious about that when playing with one of my windows 98 pcs... DDR 400mhz CAS 3, and that was the most reasonable answer i found. Perhaps the way ram is manufactured the speeds get faster but no real improvements in access times?