Hi folks, I haven't posted since threadripper and this mobo came out. Wondering if the Gskill FlareX sets of 128 have been tested and people can recommend. I'm looking at 2933 with 14-14-14-34 or 16-16-16-36 I might even do 64gb instead.
Anyone have any real world experience with these sets? From what I'm reading, 2933 is the best we can expect out of the board.
That one and runs it at a slower speed by choice. His reasoning is he needs rock solid stability for the simulations that the system is used for and that memory size is more important these days than outright memory speed.
To help us help you - please provide as much information about your system and the problem as possible.
You may notice small speed differences in synthetic benchmarks (usually nothing more than a couple percent, if that) but not in real world software (except for a couple of very specific usage scenarios, and even there the actual difference would be minimal).
My experience has been that memory speed and the channel count can radically impact performance, but the number of sticks and processor mode are part of the equation. We can get into a ridiculous amount of detail here, and obviously your mileage may very, but here is the short version:
AMD officially supports the following speeds and arrangements:
Channels Grades DRAM DIMMs per channel The number of DIMM Speed Four channels Peer One 4 of 8 DDR4-2667 Four channels Peer Two 8 of 8 DDR4-2133 Four channels Dual rank One 4 of 8 DDR4-2400 Four channels Dual rank Two 8 of 8 DDR4-1866
Note that as you add more DIMMs you cut the maximum supported speed. Even on the RAM that is "built" for Threadripper, we see a similar fall off in supported speed, depending on the number of DIMMs:
Also note that CAS 16 supports faster RAM, but only if there are only four DIMMs.
On it's face this seams to suggest that more DIMMs are a bad thing for threadripper, but in fact in certain workloads more DIMMs, even slower ones, provide better overall performance. To understand why, you need to understand how AMD built the beast. The architecture is not a single CPU die with a bunch of cores, instead it is effectively a multiprocessor machine baked into a single package. The processor dies (chips) are connected together with the proprietary AMD Infinty Fabric.
AMD recognized that different workloads would require radically different hardware configurations, so rather than build a machine that was married to a single configuration, they built Threadripper to selectively enable features to support the various workloads and then provided us with the Ryzen Master software to configure those modes. Out of the box Ryzen Master has two default modes of operation, Creator and Gaming, these modes in turn have two memory modes Distributed and Local. In Creator Mode memory is configured as distributed memory (UMA) meaning it is shared via the infity fabric to all the dies.
UMA is meant for applications like video rendering or data analytics that benefit from a huge pool of RAM. I saw improved performance in Creator Mode when every DIMM slot was populated, but it is very difficult to even get the XMP specs to be stable. The alternate configuration is local (NUMA) mode.
In NUMA mode the memory is dedicated to the CPU dies, allowing it to be significantly faster by bypassing the infinity fabric and doing direct to the CPUs. The grey DIMM slots corespond to the primary dies, and the black DIMM slots are tied to the secondary dies. What makes this interesting is that in Gaming Mode the secondary CPU dies are disabled.
In Ryzen Master you can manually enable NUMA while simultaniously enabling all the dies and cores, allowing for fast but "segmented" RAM that is often a best of both worlds configuration, if you have lots of small applications configured to run on dedicated threads. This mode even seams to allow for faster RAM (CAS16) configurations even with all the slots populated. the trick in this configuration is that if a CPU tries to access memory that is slave to the other half of the chip, performance goes right out the window, so you need to granularly control which apps are running on which cores which can get a little challenging. (Side note: linux has better tools for this type of orchestration, but you can manually assign processor affinity in Windows.)
If you are planning on staying in Creator Mode with all the cores enabled then, in my experience, you will benefit from more CAS14 DIMMs. For gaming you will benefit from only having four sticks of faster CAS 16 RAM.
To dive deeper into these issues I would reccomend the following links: