Supermicro H12DSi-NT6 VS Gigabyte MZ72-HB0 rev. 3.0

[Poleronburst]

I've used both vendors EPYC boards. I'd like to add:

H11/H12 DSI are fine in desktops for VRM temp with 240W processors. I am using 4U Supermicro coolers and a fair amount of 140mm intake and exhaust fans in the 1000 RPM range (so fairly quiet), with no shroud.

the Supermicro fan control SUCKS. needing to use IMPI scripts or constant speed fans. You can get rid of the alarm on low RPM for fans, but many fans run at like 200 RPM because the default 4 pin PWM control speed is like 20%. it has no ability to control 3 pin fans and 4 pin frequently run too slow. I instead prefer to use only 4 pin on the CPU with good speed fans (like the SMC 4U coolers are 1000 - 3500 RPM range) and 3 pin (which run 100% all the time) for all the rest. I just tune my fans by buying the right speeds for my application (1000-1500 RPM @ 140mm typically).

I keep buying SMC, but in truth I kind of liked my gigabyte experience a little more.

Great! thank you for your post, very usefull!
I'm kind of nudging towards the Gigabyte at the moment (not 100% sure why) just wondering about the difference in BMC chip.
the AST2600 on the SMC seems like a big upgrade when compared to the AST2500 on the Gigabyte

https://www.servethehome.com/aspeed-ast2600-launched-as-your-next-gen-bmc/

I must admit that I'm not 100% sure if I would even notice the difference, though what is this extra power used for?

 

[Spartus]

Yeah I have no idea, I have both an H11 board and H12 board sitting next to me (AST2500 and 2600), and other than the visual difference of the webpage I have no idea whats different. Does it feel like triple core when I'm interfacing with it... I couldn't say I've noticed.

Actually on that note I'd like to make another complaint. I can't use the H12 IPMI (AST2600) in chrome, I need to switch to firefox or edge or basically anything else.

 

[Poleronburst]

I ordered the Gigabyte and will see how it goes, can always switch to the Supermicro, this build will probably break the bank anyways

Something else that I'm wondering about, these boards have 8 ram slots per cpu and epyc 7003 cpus have 8 memory channels of 1 or 2 dimms per channel so for the Gigabyte and Supermicro boards 1 dimm per channel.
Would this mean that populating all 8 dims per cpu would yield faster/higher memory bandwidth? So 4x64 would be slower than 8x32? (Assuming the same spec memory ofcourse, so cas en freq etc)

 

[alex_stief]

Yeah, you need to populate at least one DIMM per channel in order to have close to full bandwidth available. Just like every other platform. Whether that affects your particular workload is a different question.

 

[ari2asem]

I ordered the Gigabyte and will see how it goes, can always switch to the Supermicro, this build will probably break the bank anyways

Something else that I'm wondering about, these boards have 8 ram slots per cpu and epyc 7003 cpus have 8 memory channels of 1 or 2 dimms per channel so for the Gigabyte and Supermicro boards 1 dimm per channel.
Would this mean that populating all 8 dims per cpu would yield faster/higher memory bandwidth? So 4x64 would be slower than 8x32? (Assuming the same spec memory ofcourse, so cas en freq etc)

yes, 8 dimms per cpu is faster than 4 dimms. no matter 64gb or 32gb stick.

 

[Poleronburst]

Great thanks! Makes sense, unfortunately on a small bord like this (1 slot per channel) it would mean that it would be best to install all memory at once making it financially a bit more challenging in order to avoid replacing Sims. I assume registered ecc vs regular ecc would only be interesting for large Sims?
Does single sided vs double sided simms make any difference?

 

[alex_stief]

I assume registered ecc vs regular ecc would only be interesting for large Sims?

You can't use unbuffered memory aka regular ECC with this hardware. It has to be reg ECC, or one of the other supported types.

Does single sided vs double sided simms make any difference?

To my knowledge, this does not have any effects.
What does matter on the other hand is memory ranks, especially with one DIMM slot per channel. A single rank per memory channel is not quite enough to get the maximum memory bandwidth in sequential benchmarks. To an extent where it has a measurable effect on bandwidth-intensive applications. Use memory modules with at least two ranks to avoid this. Yes, this still applies with reg ECC memory.

 

[Poleronburst]

Great thanks! Saves me a potential headache...
Registered ECC Dual or quad rank it is then

 

[Poleronburst]

Thought I'd post a short update, I've just received the MZ72-HB0 rev 4.0 (didn't know there was a 4.0, assumed the latest was the 3.0) and my first impression is that the packaging could be better. I'm not talking about the lack of a fancy colored box (this one was just plain brown) but the lack of protection the packaging provides. The VRM heatsink (the one between the sp3 cpu sockets) ripped the antistatic bag and even left marks in the lid of the box. The box also had a slight bulge. So I hope that non of the components underneath the heatsink got crushed...

I'll post another update this weekend when I will assemble the system and give it a try.

 

[alsaadib]

Thought I'd post a short update, I've just received the MZ72-HB0 rev 4.0 (didn't know there was a 4.0, assumed the latest was the 3.0) and my first impression is that the packaging could be better. I'm not talking about the lack of a fancy colored box (this one was just plain brown) but the lack of protection the packaging provides. The VRM heatsink (the one between the sp3 cpu sockets) ripped the antistatic bag and even left marks in the lid of the box. The box also had a slight bulge. So I hope that non of the components underneath the heatsink got crushed...

I'll post another update this weekend when I will assemble the system and give it a try.

Thanks for the update. I'd like to follow your build. Can you give any info as to the rest of the build? How will you populate the PCIe slots? What type of power supply / chassis will you be using?

Best of luck