You will be hard pressed to find a consistent answer to that. It depends on the game, and the differences are very small either way. If you are interested in tweaking and overclocking go with 3200, if not, maybe still go with 3200. Even if you get a 4600 MHz kit and a very good board you are still playing the silicon lottery in hopes of getting a CPU capable of hitting those speeds.
If you are patient, and set reasonable expectations, you can make either choice without regret.
Somewhat agree with this. Just because the RAM was tested to run at 4600mhz on a lucky silicon CPU, does not mean that it will even post at 4600mhz on your CPU.
Highest my 5960x CPU would post after increasing the system agent voltage by quite some was 3200mhz. Newer CPUs might fare better but there is no guarantee still, whereas the 3200mhz RAM will most likely run just fine.
Personally i prefer low latency over high bandwidth, which is very beneficial if you run virtualbox or vmware, where there is not much sequential processing requiring a lot of bandwidth, but rather many random tasks which make your virtual environment feel snappy and closer to running a navite system if the RAM has low latency.
But lower latency at same frequency also means more bandwidth. I am not an expert in this and i doubt many who think they got this figured all out are either. It seems to be quite complex.
While i was experimenting with overclocking my RAM on a 5960x i was using AIDA64 memory benchmarks to figure out how far i got.
The memory read bandwidth and memory copy bandwidth was increasing both with higher frequency and lower CL values. However, the memory write seemed to have been unaffected.
Memory write turned out to be affected by the cache frequency on the CPU. Unfortunately, my 5960x is quite a lemon(or maybe the motherboard is the lemon) and would not allow anything higher than 3500mhz for the cache while staying below 1.25v for the cache and +.03v offset for the system agent.
Also quite surprising, in the AIDA64 benchmarks, the L1 and L2 cache bandwidth did NOT increase by raising the cache frequency much but rather increased by raising the CPU frequency.
Also notable, the cache frequency affected the memory latency by quite a lot. Same frequency and same CL for the RAM, running the cache at default 3000mhz vs 3500mhz, the RAM latency in the AIDA64 test would drop by quite a few nanoseconds.
Also the RAM bandwidth would go up by increasing the cache frequency.
The CPU frequency also dropped the RAM latency by a few nanoseconds but to a lesser extend than the cache frequency did.
Those are my findings related specifically to the 5960x CPU. It might be very different on other architectures. What i realized is that having a CPU which can run its cache at very high frequencies at reasonable voltages is more beneficial than what some people on the web suggest as the cache frequency affected both the bandwidth and latency of the RAM.
But more important, it was the only factor in memory write bandwidth or so it appears to me. Maybe i am missing some knowledge here or some setting.
At least if other CPUs are anywhere similar to the 5960x i am using.
That 4600 has faster primary timing and other timings are mixed. Effectively, the time it takes to get data can be calculated from clock speed as well as latency timings.
3200MHz makes each clock cycle 0.31 ns, 4600MHz is 0.217 ns.
CL is in cycles, so:
- 3200 CL14-14-14-34: 4.37-4.37-4.37-10.6 ns
- 4600 CL18-22-22-42: 3.91-4.78-4.78-9.13 ns
Both are most likely based on the same dies, just tested and binned differently.
Overclocking is always hit and miss.
The formula many use is usually 2000 * CL / RAM_frequency which would in the case of the 3200mhz RAM result in 8.75ns - 8.75ns - 8.75ns
In the case of the 4600mhz RAM you would end up with 7.82..ns - 9.56..ns - 9.56..ns
Now i did not really check how people arrived at this formula, but it is definitely the formula a popular marketplace in europe called geizhals is using when listing different RAM sets.
Also the formula this ram latency calculator uses
https://notkyon.moe/ram-latency.htm
Your initial 3200mhz being 0.31ns per clock cycle and 4600mhz being 0.217ns appear to make sense as 1s/3200mhz does result in 0.31ns i suppose. Then 14*0.31ns = 4.37ns also APPEARS to make sense. 34*0.31ns = 10.6ns
But then is everyone else doing the calculations wrong?
Also, which set is better according to your calculations? Some numbers on the 3200 set are lower and vice versa. What is the expected bandwidth and latency estimate i would get in various memory benchmarks according to the numbers you computed for both sets? Which set would perform better in bandwidth and latency according to your numbers?
Also note that the 4600mhz set needs 1.45v whereas the 3200 set gets there already at 1.35v. It would appear that the cheaper set is actually better, if it can also handle 1.45v with even lower timings, which it should as both appear to have the same heatsinks and probably have the same dies as well.