brispuss :
Thank you for your detailed response. But there is still I believe there is some misunderstanding on your part.
And I was not referring to RAMCheck LX device in my previous post.
True/effective latency is actually: 2000 x CAS Latency/memory frequency ~ constant, for that particular module with certain memory chips. The "Constant" value varies according to what memory chips have been used in the particular module.
This equation MUST define the effective latency for each specific module based on it's characteristics. Module characteristics are dependent on the memory chips used. Some memory chips allow low CAS Latency, other high CAS Latency.
Yes, there are modules @ DDR3-1600 with CAS Latencies as low as about 7, I believe, but that is entirely dependent on the grade/quality of the memory chips used.
So in the case of memory of DDR3-1600 with CAS Latency of 11, the grade of memory chips used must limit the CL to 11, so it will not be possible to run the module with lower CL (at same voltage). Yes?
The absolute lowest tCAS that can be programmed on any DDR3 SDRAM chip is 5 and the absolute highest tCAS that can be programmed on any DDR3 SDRAM chip is 14. These are design constraints that cannot be bypassed regardless of speed grade.
The DDR3 SDRAM chip requires at least one cycle to perform the combinational logic on the backend and serialize the 8 words for the read command into the IO logic. Then it requires 4 cycles (fixed) to transmit the data on the IO bus, with one transmission occurring every half cycle.
In practice, the back-end logic can only be performed in a single cycle at extremely low command clock frequencies (at or below DDR3-800, or 400Mhz), but the front-end transmission logic takes 4 cycles regardless of the configured data rate. At higher data rates, the bulk of the time is spent waiting for the backend.
There are two ways to speed this up:
1. Crank up the supply voltage
2. Hand pick the best chips and put them together
The JEDEC specifications are meant to maximize interoperability and stability, they pay little heed to performance. tCAS means little when employed in industrial and medical devices, but stability is very important.
There are only a small number of DDR3 SDRAM chips on the market.
Samsung has only two 4 gigabit DDR3 die designs that I know of (with minor stepping revisions of each), the K4B4G0446 which is a 4 gigabit chip with 4 IO pins, and the K4B4G0846 which is a 4 gigabit chip with 8 IO pins.
Hynix operates in the same way. They have 3 4 gigabit DDR3 die designs with minor stepping revisions of each. H5TC4G43 with 4 IO pins, H5TC4G83 with 8 IO pins, and H5TC4G63 with 16 IO pins.
Micron is the same as well. They have 3 4 gigabit DDR3 die designs. MT41K1G4RG with 4 IO pins, MT41K512M8RG with 8 IO pins, and MT41K256M16LY with 16 IO pins.
In general, each manufacturer makes each die available as a 1.5v DDR3 chip and a 1.35v DDR3L chip. Then, each chip is made available in 2-3 speed grades. Presently, DDR3-1600, DDR3-1866, and DDR3-2133 are available from manufacturers. Each is of course backwards compatible.
What this means is that all of the various DIMMs available on the consumer market are nearly identical underneath the heat spreader. Almost all 8 gigabyte unbuffered DIMMs will be constructed as two ranks of 4 gigabit 8-IO DDR3 SDRAM chips. This means that the selection of chips are limited to the K4B4G0846 from Samsung, the H5TC4G83 from Hynix, and the MT41K512M8RG from Micron.
Purchasing a tray of higher speed-grade chips may enable the DIMM manufacturer to squeeze a little bit of extra performance out of the resulting DIMM by virtue of the SDRAM manufacturer's testing process.
For example, the K4B4G0846Q (the latest stepping of Samsung's 4 gigabit 8-IO DDR3 die) is available in two speed grades. It is available as K4B4G0846Q-HCK0 (DDR3-1600 11-11-11), and K4B4G0846Q-HCMA (DDR3-1866 13-13-13).
An assortment of K4B4G0846Q-HCMA may, without overvolting, have an easier time running at a very tight non-standard profile such as DDR3-1600 7-8-7 than the K4B4G0846Q-HCK0. Neither Samsung, nor JEDEC, nor Intel make any guarantees when their products are used outside of the standards to which they are designed.
It is therefore very important to separate the theory, the standards, and the implementation from each other.