So, I was looking at all the recent computer building threads and noticed that the DDR4 RAM has noticably higher CAS Latency. Back when I was informing myself of the new DDR3s a good module had CL9 or CL11 (when clocked significantly higher) while we see up to CL16 now.

So I wondered what was going on and did some math. I noticed while throughput indeed increased (obviously), the (random) access times actually got higher (bad). But hasn't the throughput been much higher than any PC component can sustain anyways for like a decade now? And isn't answering an "unexpected" (or random) memory access faster, more important in real world applications? (See below if you want some numbers)
I'd say yes to both which would lead to the conclusion that purely from the RAM point of view it would make sense for highest end builds to stay on DDR3 for now. I haven't been keeping up with the latest in CPUs, so tell me: Are the new CPU architectures worth the increased memory latency or is this all just a hype for new tech?

A comparison of high performance modules:
DDR3: 8GB GEIL Frost White EVO POTENZA DDR3-2933, CL12
Throughput: 2933 MHz / 2 * 128 bit / 8 bit/B / 1.024^3 B/GiB = 21.85 GiB/s
Time until 1st word can be fetched from the pins: 1/2933 MHz * 2 * 12 = 8,18ns
Time until 8th word can be fetched from the pins: 1/2933 MHz * 7 + 8,18ns = 10,6ns

DDR4: 8GB CORSAIR Vengeance LPX Black DDR4-3466 CL16
Throughput: 3466 / 2 * 128 bit / 8 bit/B / 1.024^3 B/GiB = 25.82 GiB/s
Time until 1st word can be fetched from the pins: 1/3466 MHz * 2 * 16 = 9,23ns
Time until 8th word can be fetched from the pins: 1/3466 MHz * 7 + 9,23ns = 11,3ns

=> Throughput is 18.1% faster but random access times are 9.32% slower.

Nude, asking the hard questions. I like it.

If you look at this article you can discern some of the important breakpoints in frequency vs latency in DDR4. The question has been raised many times and although we'd all like to just assume the higher number bearing DDR4 is immediately better, we may not be at the point of maturity for that to be true in terms of MHz vs latency. Corsair has some data on this here although they are in the business of selling memory so take that as you will.

My usual method to get a 'good' memory kit is to get the highest frequency at the lowest latency I can find before the price gets insane. This is a rather basic and obvious idea however. My idea of a good ratio in the current market is a frequency above 3000 with a latency of 16 or below since this seems to be the pain point before A. no kits available in the selected parameters or B. price goes through the roof.

Cool chart from Crucial:


Post has been edited 1 time(s), last time on Dec 25 2015, 3:56 pm by Excalibur.

David Patterson, the guy who wrote THE book on computer architecture, has a saying that goes, "latency helps bandwidth, bandwidth hurts latency". Basically, improving latency causes your throughput to increase (if bits move faster over the line, more bits can go over the line in a set period). However, the opposite isn't true. Imagine going from a 32bit bus to a 64 bit bus. Before we can transmit, all 64 lines must be ready to go. This may take a little bit longer to do. So we increased latency slightly to allow for more total data. Your quick math experiment is a little flawed as well. You never grab a word from memory. You grab whole blocks way larger than a word (or DWORD) from main memory and then put them in cache.

http://www.ll.mit.edu/HPEC/agendas/proc04/invited/patterson_keynote.pdf