For more than two years, I've enjoyed my 5960X build. But the time came where I needed to re-purpose it for something else. And with Zen coming out, I took the opportunity to build myself a new main rig for coding, gaming, and Anime. Being a drop-in replacement for my existing setup, most of the parts were inherited and are therefore priced at zero.
I knew coming in that there was a high likelihood that this would be a bit of a downgrade from my 8-core Haswell-E. And indeed this was the case. After factoring in overclocks, my main use-case (compiling code) ended up about 10-15% slower on the 1800X than the 5960X. But this wasn't exactly the time for me to splurge since I have my eyes on bigger hardware later this year.
So I jumped on the Zen train on launch day (March 2). But what a mess...
This was perhaps the ugliest build I've ever done since 2008. I went out on launch day completely expecting to have a fully working system before midnight. In the end, it took 2 motherboards, 3 sets of memory, 3 weeks, and $1000 over budget did I finally complete the build to my satisfaction.
For this reason, you will notice that the pictures contain parts that are not in the final list here. And it does remind me why I always do open-air testing on all new builds before I put them in a case. And in this case, the build had been sitting open-air for 2 weeks while I sorted out the issues.
The DDR4 shortage along with the Zen demand meant that I had trouble getting memory. I ordered most of the parts a week before the Zen launch hoping to short-circuit the demand. This included a cheap set of 4 x 16GB G.Skill TridentZ @ 3200 MHz, but Newegg oversold and canceled it - some two days later. In other words, one week before launch was already too late. By then, all the cheap memory was sold out. And all that's left were some 8 x 16GB TridentZ kits.
Being a true fanatic, I grabbed a set of those (which sold out 4 hours later), and shoved them in that 5960X build that I was going to repurpose. That displaced a set of 8 x 8GB Corsair which I could temporarily use in this Ryzen build.
It turns out that the Corsair memory wasn't stable in the Ryzen build - despite being on the QVL. It took me a long time to identify this memory instability because there was also a driver instability which kept crashing the system before Windows updates could be applied.
To make matters worse, while I was sorting out the memory and driver instabilities, I ran into the FMA3 hard-freeze which was later confirmed by AMD as a processor errata.
In other words, I was debugging 3 sources of instability without knowing there were that many. And if you're wondering, I'm the "Mysticial" who discovered the FMA3 errata. So the bug was unknown at the time, and I was unfortunate enough to have had to debug that among other things at the same time...
In the end, I was able to isolate all 3 sources of instability once I noticed that the system was completely stable under these conditions.
- I swapped out the 8GB Corsair sticks for the 16GB TridentZ sticks that I ordered for my Haswell-E box.
- Windows was fully updated, or I was running Linux.
- Other people had responded to my HWBOT thread confirming that the FMA3 crash was not unique to my system.
From there I was able to work backwards:
- Putting the 8GB Corsair sticks back in reproduced one set of freezes and BSODs.
- Reinstalling the OS and attempting updates reproduced the other set of freezes and BSODs.
- Avoiding the FMA3 freeze was as simple as, "Doctor it hurts when I do this. So don't do it!"
But I wasn't done yet. While the system was stable with the 16GB TridentZ sticks, I couldn't actually run it that way since the memory belonged to my other machine. So I ended up ordering a set of 4 x 16GB Ripjaws - which fortunately was stable in the Ryzen rig.
The next set of problems came with overclocking and configurability. The Asus motherboard has no base clock control or the ability to disable SMT. Furthermore, the BIOS overclocking options are so poorly documented that it's unusable. Lastly, the southbridge chipset was very prone to overheating as it idled at 70C. Since it's directly under the video card, it's hard to get airflow over it.
The Asus mATX was actually my second choice motherboard. My first choice was the Gigabyte mATX. But it wasn't in stock on launch day. But so much time had passed that it was almost available. So I played the refreshing game on Newegg. And the moment it stocked, I dropped my order - which then sold out minutes later...
The Gigabyte mATX is much better in that the overclocking options were sane. But it has other issues:
- The VRM mosfets reach 130C under load and require active cooling. This is difficult to do for an mATX build.
- The latest Ubuntu doesn't boot on it. The Ubuntu developers are aware of this so I assume this will be fixed at some point.
- It still has no base clock control. Apparently only the highest-end motherboards do.
By the time the build was complete (March 22nd), I was around $1000 over budget from the second motherboard, the additional 128GB of ram that I didn't expect to get, and a few other small things.
- Despite the fact that the 1800X can turbo to 4.1 GHz, it's not stable with all cores at even 4.0 GHz when leaving the VCore alone or even a VCore boost of 0.100v. I'm not entirely sure why or if it's something LLC related. In all my previous builds, it was always possible to clock all cores to the maximum single-core boost without touching the voltages. Not for Ryzen though...
- I was never able to get the memory to run above 2400 MHz. In fact, it wouldn't even post at 2666 MHz. The memory is rated for 3200 MHz.
- Water cooling in an mATX is precarious. mATX builds are already tricky to do since there's limited space and little airflow. Water cooling makes this worse since there's no airflow at all around the CPU/water-block. With normal air cooling, the CPU fan is close enough to the VRM mosfets to give them some airflow. But in the case of water cooling, it's all stagnant air. To solve this problem I had to mount a 120mm fan on the side-panel that was powerful enough to get airflow all the way down to the VRMs.
- There is no clearance between the radiator and the video card. This was something I foresaw while planning the build, so I was ready to mount the radiator in the front. But it did end up fitting just perfectly. This isn't a problem on the Asus mATX since the video card is one slot lower.
Both CPU and memory frequencies are maxed out with no headroom at all. The system won't post with the memory at 2666 MHz. And the CPU overclock is largely unresponsive to VCore and VSOC increases. At 4.0 GHz, upping the VCore from 1.350 to 1.400 lets it survive y-cruncher for 5 minutes instead of 1 minute. But increases beyond that had no effect. Even at 1.450 VCore it still failed y-cruncher within minutes.
So far overclocking has been quite disappointing. And I hold on to some hope that a suitable BIOS update will let me hit 4.0 GHz stable and to get the memory beyond 2400 MHz.
Current 24/7 Overclock Settings:
- Base Clock: 99.80 (cannot change)
- Multiplier: 39.25 x 99.80 = 3.917 GHz
- Memory: 24 x 99.80 = 2,395 MHz
- Timings: 16-18-18-38
- VCore: Stock - 1.350v under load
- VSOC: Stock - 0.800v under load