Samsung memory has poor bandwidth efficiency

[Khenglish]

So I found out that my old dell latitude came with the infamous samsung chips capable of 2133mhz+, however in the 2GB per stick variety. I came across bizarre results that I did not expect.

When running tests at 8-9-9-20 1866MHz, I found that the IGP was performing substantially worse than with the G.skills at 10-11-9-27 1866MHz. I ran SANDRA's memory bandwidth test and got the following:

(ignore that the samsung memory reports as revision C instead of D, which it really it. Needed to flash to an SPD of a rev C to get the memory to work in a toshiba...)

blue at 24.2GB/s is the G.Skill at 10-11-9-30 1862MHz (factoring in that BCLK is 99.77MHz not 100MHz)

green at 22.06GB/s is the G.Skill at 9-10-8-25 1660MHz (yes 1660 not 1600. BCLK overclock)

orange at 22.2GB/s is the samsung at 8-9-9-20 1862MHz

green at 23.2GB/s is all 4 modules combined at 10-11-9-28 1862MHz

unlisted run at 22.0GB/s for the samsung at 10-10-10-23 1862MHz

The samsung only has 91% of the bandwith as the G.Skill at the same frequency and roughly equal timings, and 92% of the bandwith with the timings minimized. However, the samsung can likely do 2133 at CAS 10, which will cause bandwidth to shift into the samsung's favor. My laptop kept reverting the memory to 1866 when attempting 2133, so I could not test the bandwidth advantage at 2133.

As for memory latency we have the following:

81 cycles samsung 8-9-9-20 1862MHz

91 cycles g.skill 10-11-9-30 1862MHz

94 cycles all memory combined 10-11-9-28 1862MHz.

As expected for the 25% CAS latency reduction we have a major 11% latency advantage in favor of the samsung. This is expected since the memory controller adds substantial latency regardless of memory timings.

If I could get around the 2133MHz bug I was experiencing, I would expect the samsung to slightly beat the G.skill in bandwidth while beating it substantially in latency. I would guess 4% more bandwidth and 9% lower latency. Although it is possible that the samsung could pull off CAS9 2133 and have a substantially larger latency advantage, but I doubt that and expect CAS10.

So when comparing samsung to other memory makers, they do seem to still be the winner if you can use XTU or flash the SPD to raise memory frequency and lower timings. Unfortunately due to this bandwidth inefficiency, samsung does not outperform the other memory makers by as much as expected.

[angerthosenear]

So the Samsung only would get a lead at higher speeds correct?

I was confused by your wording, the G.Skill has the controller?

Number wise the Samsung has better latency, but the G.Skill has better bandwidth?

[Khenglish]

So the Samsung only would get a lead at higher speeds correct?

I was confused by your wording, the G.Skill has the controller?

Number wise the Samsung has better latency, but the G.Skill has better bandwidth?

if you set both at the same frequency yes the G.Skill gets more bandwidth, but the samsung will be lower latency. The samsung can flat out run higher frequencies which you can do to get minor bandwidth and latency advantages.

In summary if you can have fine control over things like frequency and latency, then go with samsung memory, but if say you have a hard memory frequency cap like on an i3 or an APU, then go with something other than samsung to maximize bandwidth for the IGP.