Prebuilt [SFFn] ASRock's DeskMini A300 - Finally!

[ShamedGod]

Funny. I just bought that Noctua yesterday. Mainly because I was sick of tying to shove the AMD cooler into that tiny case (I like to swap the 2.5 drives a lot) more than for heat reduction reasons, also have other SFF builds in mind that I could use it for and it seems to be one of a kind. As far as that G Skill RAM I found some okay timings, I can get it down around 15-16-16-16 and am not too upset but I really wanted C14 @3200 as that tends to be Ryzens best friend. I am still planning on returning my Kingston 2933 memory but if I can find someone selling their known B die 3200 sodimms I'll probably buy them just to have an extra set for said other projects.

Yeah, that is REALLY annoying, especially when the optional USB front port cords keep getting in the way. Hopefully I won't be swapping out parts to much. If I was swapping out the 2.5" drives all the time I'd 100% spend the money on the NH-L9A. The fact it's quieter is a nice bonus.

How in the world did you get your timings down to 15-16-16-16? I tried 17 -17-17-17 and the A300 refused to post. Did you up the ram voltage to 1.35? Are you on BIOS 3.20 like @Danlopez1222

 

[ShamedGod]

Opps, I was wrong, I couldn't get them lower then 18-17-17-17. Just checked my saved profile. Can't remember why I changed it back to 18-18-18-18.

 

[Primerib]

Yeah, that is REALLY annoying, especially when the optional USB front port cords keep getting in the way. Hopefully I won't be swapping out parts to much. If I was swapping out the 2.5" drives all the time I'd 100% spend the money on the NH-L9A. The fact it's quieter is a nice bonus.

How in the world did you get your timings down to 15-16-16-16? I tried 17 -17-17-17 and the A300 refused to post. Did you up the ram voltage to 1.35? Are you on BIOS 3.20 like @Danlopez1222

I bumped up the voltage to 1.35, yes. I have yet to try out 3.20, I've been running 3.40. Maybe I'm booting but I'm not stable. I need to run a proper RAM stability test. I've just been using the CPU-Z stress test.

Update: used Memtest64 and was not stable at previous timings (15-16-16-16) but would still post but I am stable at 16-17-17-17-28 with Hynix die. I can only imagine how awesome timings could be with Samsung B die. I returned my garbage Kingston Sticks with Micron dies. I'll never abide by a QVL ever again. Never have in the past and will keep it that that way.

 

[samfisher5986]

I finally got my A300W! Took a while to put together.

I found that my system was unstable, randomly BSODing when I tried to get into a game.

Setting my 3200mhz ram to 3000mhz seems to have fixed it which makes sense as I read it wasn't that stable above 300mhz. However now I'm wondering how some of you guys have ram at 3200mhz etc?

I'm running Bios 1.20 which is what it came with, I didn't read anywhere that updating past this would improve stability.

 

[Karmeck]

I finally got my A300W! Took a while to put together.

I found that my system was unstable, randomly BSODing when I tried to get into a game.

Setting my 3200mhz ram to 3000mhz seems to have fixed it which makes sense as I read it wasn't that stable above 300mhz. However now I'm wondering how some of you guys have ram at 3200mhz etc?

I'm running Bios 1.20 which is what it came with, I didn't read anywhere that updating past this would improve stability.

Patch note for bios says they improve compatibility for ram.

 

[samfisher5986]

Patch note for bios says they improve compatibility for ram.

Only for 1.20 though, which mine came with.

Maybe I should upgrade anyway.

 

[ConsolidatedResults]

Only for 1.20 though, which mine came with.

Maybe I should upgrade anyway.

I went 1.20 -> 3.20 -> 3.40 -> 3.20 -> 3.40 -> 1.20. I could not find any significant differences in the clock or timings I could get for my particular DIMMs. The box even POSTs and boots to Windows at 3600 on auto-timing (no, that is not stable ).

I am however reasonably sure that the two 3.X BIOSes with AGESA 0.0.7.2 perform worse, in terms of memory bandwidth and latency, given same timings, than the 1.20 BIOS with AGESA 1.0.0.5. It's as much as e.g. 5% difference in Unigine Valley Basic for me.

Furthermore, I am still trying to figure out whether I am seeing things, or whether the A300M-STX generally shows significantly lower memory bandwidth for a given processor and memory speed / timings than other platforms. I have posted about that at reddit

Spoiler: reddit thread

In the review, John noted that the platform performed worse than an ITX platform running to the same spec and I wonder whether RAM might be causing it. Would anyone be willing to share his or her bandwidth and latency benchmarks?

Edit: Spoilered the reddit wall of text

 

[Quango]

Maybe DIMM vs SO-DIMM? Any other AM4 SO-DIMM platform to compare with?

 

[ShamedGod]

@samfisher5986

I believe I've found the best balance for my particular set of components. No mater what I tried DDR4-3400 or above would not stay stable OR provide better performance in benchmarks. DDR4-3333 had the best combination of added speed, low timings, and was (most importantly) 100% stable.

100% stable was defined as 0 errors in MemTest, 0 crashes while bench marking, and able to run sustained stress testing for hours.

After several days of normal use there has been no issues with these settings.

For testing I used primarily:
3DMark
CPUZ
FurMark
UserBenchMark
MemTest

The ram chips being used and researched was verified with Thaiphoon Burner:

The BIOS Settings where verified by Ryzen Timing Checker:

I had to reset the BIOS more times then I care to admit finding these settings but it was fun experimenting. Real world gains over the standard XMP profile probably isn't significant enough to justify the time spent tweaking but, again, it was for fun. It also really got me curious on the G.Skill 3800 and COSAIR 4000 speed memory chips. If money was no object I probably would have bought both to see what the performance gain, if any, would be.

Provided below are screenshot of my saved DDR4-3333 settings strait from the BIOS:

Hope that's helpful. Your mileage will very and I am using a sample size of one with my personal silicone lottery results.

Edit: UserBenchMark from today:

Please click the green human to continue - UserBenchmark

CPU: usually 98th-100th percentile
Graphics Card:With the 3333 ram settings is usually 22.1% or 22.2% benchmark (high 80th to low 90th percentile)
Drives: the Intel 660p is all over the place. Sometimes it's in the 25th percentile, other times it's in the 90th percentile. The Drive is a budget NVMe which is perfectly suited for this build.
Memory Kit: The G.Skill kit is usually between 102-106% (95th percentile or grater)

This is not on a clean windows install but after weeks of use and between 20-30% of the system drive filled with installed applications. Benchmark run at IDLE after boot

 

[ConsolidatedResults]

Maybe DIMM vs SO-DIMM? Any other AM4 SO-DIMM platform to compare with?

I believe I've found the best balance for my particular set of components. No mater what I tried DDR4-3400 or above would not stay stable OR provide better performance in benchmarks. DDR4-3333 had the best combination of added speed, low timings, and was (most importantly) 100% stable.
[...]

Likely not SO-DIMM in itself. At the moment I suspect the issue may be that my particular SO-DIMMs are single rank with 4 DRAM ICs, x16 wide. While there is not much literature on this, apart from a Micron whitepaper that is focused on DRAM for network applications (e.g. hitting a lookup table millions of times per second), the fact that x16 DRAM ICs have (by JEDEC spec) half the banks (8) and half the bank groups (2) vs. x8 DRAMs, may cause performance penalties. I have updated the previously linked reddit thread with a bit more detail.

@ShamedGod for example has got his results with x8 DRAM SO-DIMMs (16 banks). He hits 33.2 GB/s Single Core combined RAM speed in his UserBenchmark results at DDR4-3333. I need DDR-3466 (with somewhat wonky timings tbh) to hit almost exactly the same score (33.3 GB/s).

I won't complain, I got a stable 3466MT/s out of a system pull rated 2666 (OEM kit for new Mac Minis) and as of now both the fastest kit of that particular RAM module and the fastest A300/2200G system on UserBenchmark

I'll look out for a bargain on a kit with x8 DRAMs and will do a direct compare at some point Luckily DDR4 prices a still falling.

 

[Valantar]

Likely not SO-DIMM in itself. At the moment I suspect the issue may be that my particular SO-DIMMs are single rank with 4 DRAM ICs, x16 wide. While there is not much literature on this, apart from a Micron whitepaper that is focused on DRAM for network applications (e.g. hitting a lookup table millions of times per second), the fact that x16 DRAM ICs have (by JEDEC spec) half the banks (8) and half the bank groups (2) vs. x8 DRAMs, may cause performance penalties. I have updated the previously linked reddit thread with a bit more detail.

@ShamedGod for example has got his results with x8 DRAM SO-DIMMs (16 banks). He hits 33.2 GB/s Single Core combined RAM speed in his UserBenchmark results at DDR4-3333. I need DDR-3466 (with somewhat wonky timings tbh) to hit almost exactly the same score (33.3 GB/s).

I won't complain, I got a stable 3466MT/s out of a system pull rated 2666 (OEM kit for new Mac Minis) and as of now both the fastest kit of that particular RAM module and the fastest A300/2200G system on UserBenchmark

I'll look out for a bargain on a kit with x8 DRAMs and will do a direct compare at some point Luckily DDR4 prices a still falling.

Hm, that's interesting. Maybe someone should get a SODIMM-to-DIMM adapter (like this) and test if the performance degradation carries over to a regular desktop platform?

 

[Primerib]

Hm, that's interesting. Maybe someone should get a SODIMM-to-DIMM adapter (like this) and test if the performance degradation carries over to a regular desktop platform?

I really wish I could find that product but reversed so its DIMM-to-SODIMM. I can't find any talk of one online because of very few STX platforms being commonly run out there that would actually have the ability to use such a adapter.

 

[josephclemente]

I just built a system at work with a DeskMini A300W. I am very pleased with it. It will be used as a basic office computer, so I went with a 2200G and had 16GB of DDR4 2400 memory from a dead laptop. Used the 2200G's stock heatsink with the outer shroud removed.

 

[Valantar]

I really wish I could find that product but reversed so its DIMM-to-SODIMM. I can't find any talk of one online because of very few STX platforms being commonly run out there that would actually have the ability to use such a adapter.

I have my doubts as to whether this would be possible at all - to make such an adapter you'd need quite a large PCB (much larger than a DIMM-to-SODIMM adapter), that it would significantly lengthen trace distances for the RAM, and due to the shape (narrow socket fanning out into a wide socket) some traces might end up much longer than others. That would likely have dramatic effects on signal integrity and timings, both of which are things RAM are highly sensitive to. Then there's the issue on a lot of boards of (potentially) fitting two of these in stacked, tightly spaced, flat-style SODIMM sockets, which likely wouldn't be possible, shrinking the potential market for such a product by quite a lot.

 

[glow]

DDR4 DIMMs also have a few more pins than DDR4 SO-DIMMs. I do not know exactly what is the difference between the pinouts (a WAG would be extra power pins for reference purposes).

 

[Quango]

A thin mini ITX board - which always use SO-DIMMs - with AM4 could be compared to the Deskmini A300. And/or Intel platforms with H310 in mini-STX and and formats with DIMMs to check for differences.

 

[ShamedGod]

Buildzoid did a thing on this. Ram is Ram is Ram. It's built by Micron, Samsung, or Hynix.


You have to take all your components into consideration.

 

[ShamedGod]

Also, I just saw this video in my YouTube feed.

 

[Primerib]

It's weird to say ram is ram is ram when we all clearly know that Ryzen performance is determined by the manufacture of the die of your ram and the speed and latency it is capable of, as a Threadripper user I know this to be very much dependent on what ram you choose. I had to spend $500+ on quad channel ram just to make the X399 platform run correctly. Sad to hear Samsung B dies will be discontinued. Now the world must turn to Hynix which I suppose I can live with but everyone said the silicon shortage was just a myth. I guess all these inflated ram prices were not just money grabs all along. I just hope we can get the next APUs on our deskminis. If I have to buy another one than fine I guess they got $150 out of me for a cool product and I'll never regret it. I"m happy that they allowed us to enjoy such a SFF Ryzen build at a reasonable price.

 

[NateDawg72]

I don't think the "ram is ram is ram" thing is saying ram doesn't matter. It's saying no ram module vendor has special sauce that is going to make their modules better than everything else, so you can dig straight into what it is made with. There are really only 3 large manufacturers for the chips, so what you need to know is which chips from which vendor and then you know what those modules can do.