Hmm. Early Zen and Zen1+ AM4 Ryzens were very picky about RAM, but this is significantly improved with Zen2 and was almost entirely eliminated with Zen3 and later and is mostly historical at this point. By the time Zen2 came around you could use almost any RAM you wanted without stability issues, and Zen3 and later were even more stable.
Were you trying to populate four RAM slots maybe? Or maybe trying to use dual rank RAM? Because that does run into signaling issues on both AMD and Intel platforms, and usually requires the RAM to run more slowly in order to be stable. Usually the advertised memory clocks are only fully supported when using one single rank memory module per memory channnel. It has been this way for a very long time on both Intel and AMD simply because as memory performance has increased, the signalling has become increasingly sensitive and difficult to maintain.
This is usually documented straight in the manual for the motherboard, but looking at the manual for the ProArt board you have, it seems Asus neglected to mention this. They just reference the Memory support list on their webpage, and it does not speak to differences in memory speed with different module quantities installed. (They do notably not have any 4x kits in the list)
If you were just using two RAM slots, then you may have simply had defective RAM. Sometimes unstable RAM can be made stable by slightly boosting the voltages to the memory or the IMC. (care needs to be taken here to avoid overheating or damage)
It’s also curious that they advertise 8000MT/s capability. While this is true, in most cases when used the performance will actually be lower than slower memory kits. The key to AMD memory performance is to run UCLK (uncore clock, the speed the IMC operates at) = MCLK (memory clock, the clock the memory modules operate at (a 1:1 ratio). This results in the lowest latencies in the overwhelming majority of cases, and bandwidth is directly related to latency.
The issue here is that UCLK has a max specified speed of 3000 Mhz, making DDR5-6000 the fastest officially supported RAM in a 1:1 configuration.
It is possible to go higher (up to 6200 and even 6400) but this involves overclocking the IMC and running it at higher voltages which is not guaranteed to work, will run hotter and if you set the voltage too high can damage things.
If you use higher RAM clocks (higher MCLK) than uncore clocks (UCLK) you will be in an asyncronous multiplier mode, which results in higher latencies, and will be bandwidth limited by the UCLK and even the FCLK (fabric clock, the communication link between the CPU dies and the IO Die) and this is generally not positive for performance, but there are exceptions with certain very specific workloads which have to clear the memory often, in which case the higher clocks reduce the wait time while the memory is cleared.
Absolutely.
But unfortunately what we used to call HEDT is dead.
The last were probably Intel’s x299 and AMD’s Zen2 Threadrippers (Threadripper 3xxx)
After that generation HEDT disappeared, and instead you got a choice between proper workstation or consumer product. The in-between “pro-sumer” category pretty much disappeared. And I think there are many reasons for this.
1.) Most people who bought HEDT platforms were high end enthusiasts, but as prices for consumer platforms started getting higher and higher, there was a diminishing number of people who were willing to spend more than what a high end consumer platform cost, other than the people who were already buying proper workstations already, and as volumes diminish, costs further increase making it even less viable.
2.) In the past with DDR4 some systems were able to be compatible with both UDIMM’s and RDIMM’s. The flexibbility allowed HEDT platforms to flourish, with high end prosumers installing cheaper (and faster) UDIMM’s and light workstation users able to install RDIMM’s. With DDR5 this is no longer possible which forces a delineation. Most workstation products require RDIMM’s and consumer products require UDIMM’s. This helped kill off the tweener HEDT platforms as the flexible crossover between the two that allowed for greater volumes justifying the platform was now gone.
3.) Increased internal cannibalization of higher margin products by lower margin ones. If a consumer platform gets more than 2 memory lanes, or gets more than 28 PCIe lanes, then some professional users may start buying cheaper consumer platforms instead of buying higher margin professional platforms.
Sadly, I think that era is now thoroughly dead. The market is now thoroughly delineated. There are not consumer platforms with high clocked CPU’s with high performance per thread and low latency RAM which are great at consumer applications and games, but suffer from having rather few PCIe lanes and few memory channels, and there are workstations which now are more like their server counterparts, highly capable with many cores, lots of memory channels and lots of PCIe lanes but designed with lower core clocks for lower performance in lightly threaded tasks, and higher latency buffered RAM.
I don’t think this will change. The costs are too high, and the benefits to both the board makers and chip makers are too small and the risk of cannibalization is too great.
What this means is that even though AMD has been making 16C/32T consumer CPU’s for some time now, for many workloads there just isn’t enough memory bandwidth to keep up when all cores are busy, forcing them to sit idle a lot waiting for RAM, and that harms performance.
The jump from DDR4 to DDR5 did help things though by roughly doubling the available RAM bandwidth, but then again, in that time the CPU cores also got faster and gobbled up a lot of that increase.
I’m with you. I have been lamenting this loss for years now. It is the main reason why I am still using my Threadripper 3960x as my main workstation, simply because modern proper workstation platforms are crazy expensive, and modern consumer platforms just don’t have enough PCIe lanes for my needs.
My Threadripper 3960x like my x79 system before it and my x58 system before that all served as “one system to rule them all” builds with no compromises. Able to both be top tier platforms for work and expansion and for the rare game I play at the same time. Now we have to choose.
When the Threadripper 3960x started to become a little slow for some game titles I wanted to play I made the difficult decision to move to a two system setup, keeping the Threadripper for work and expansion, and building a dedicated AM5 system for when I want to run a game. I hate that I have to do this, but that’s the world we now live in.