Minisforum MS-01 / RTX A4000 SFF Benchmarking

[playlogic]

Hey all,

Just for anyone curious about the MS-01 Geekbench numbers under a non-windows OS
CPU 13900H / 64GB (DDR5 5200) - Which puts it at about an top M2 PRO Max performance

Micro Computer (HK) Tech Limited Venus Series - Geekbench

Benchmark results for a Micro Computer (HK) Tech Limited Venus Series with an Intel Core i9-13900H processor.

browser.geekbench.com

GPU RTX A4000 SFF / 20GB VRAM

OPENCL Micro Computer (HK) Tech Limited Venus Series - Geekbench

VULKAN Micro Computer (HK) Tech Limited Venus Series - Geekbench

Comment a photo from a previous post I did to show the current setup:
After a small bit of modding to the case…was able to get the 4000 SFF in without modifying the card itself. Couple of things:

had to remove the bracket at the back of the card…even the half height bracket would make it a tight fit.

Remove a secure clip from the top of the front panel and the PCie card screw bracket…it prevent most cards from going in.

shaved the rubber pads down with a razor do to the card having a solid backplate unlike the a2000 mobile

The bottom plate unscrews from the top, creating two halves…I am planning on modding the since if I want to replace it down
the line…it would be much easier the void the GPU warranty. Looking at doing a 3D print top

Will update as I do more testing

 

[playlogic]

Update 1:
AI Generation - Fooocus
Generations at about 2.35 it/s on avg; temps when as high as 73c on a 4 batch image set
CPU usage never exceed 5%, (however, I think btop is reading the temps wrong for the package)

 

[oharag]

Cool to see you can shove the RTX A4000 into the MS-01. Do you mind me asking how you obtained the card - and what you spent? Looking online it's a $1500 card. I think the A2000 goes for around $850.

I'm interested in Intel Arc cards and what they will produce for half-height systems.

I'm an Engineer and interested in whether I can use the MS-01 to push heavy Engineering SW.

I don't know how you benchmarked but from this website an A4000 can achieve 108000ish for Vulcan:

Vulkan Benchmarks - Geekbench

I'm sure it's the 8X limit on the PCIE sot reducing the numbers. But it's more than 1/2

 

[playlogic]

Cool to see you can shove the RTX A4000 into the MS-01. Do you mind me asking how you obtained the card - and what you spent? Looking online it's a $1500 card. I think the A2000 goes for around $850.

I'm interested in Intel Arc cards and what they will produce for half-height systems.

I'm an Engineer and interested in whether I can use the MS-01 to push heavy Engineering SW.

I don't know how you benchmarked but from this website an A4000 can achieve 108000ish for Vulcan:

Vulkan Benchmarks - Geekbench

I'm sure it's the 8X limit on the PCIE sot reducing the numbers. But it's more than 1/2

The card retails for $1250; What you are seeing is markup from the demand...I was able to get it at retail before the price hikes
The Vulkan score an limitation has nothing to do with the x8 slot...but rather the TDP of the card being only 70 watts. The scores you see on geekbench are for the RTX A4000 and RTX A4000 Laptop GPU and not the RTX A4000 ADA SFF.

Geekbench is just one synthetic benchmark and sometimes it results can be skewed...that is why I am running test with software I will be using on it.

IF you look at their testing...this card scored even or higher with the a4000 (non ADA) at half the power

Review: Nvidia RTX 4000 SFF Ada Generation

Nvidia RTX 4000 SFF Ada Generation workstation GPU delivers in all types of workflows, but this level of performance doesn't come cheap

develop3d.com

You should only look at this card if you were previously looking at the RTX a2000, it is the same size, shape, and has the same power draw...just more VRAM and is faster. This card is very niche and only solves a hand full of use cases.

I blame Nvidia's purposely misleading naming conventions on the confusion of the cards. There are cards with the same naming scheme but different archs...like A4000 and A4000 ADA...it confusing and annoying.
This is not my primary machine; It will be a general box to take the workload and resources off of my other computers, mostly when I am mobile. One example is local AI code server; it stead of running it on my laptop eating up resources...I can run it on the MS-01 and connect to it. I think you can build a small ITX "Lunchbox" type machine and get a full fat A4000 for your needs.

RTX A4000 ADA/ 140Watts : NVIDIA RTX A4000 Graphics Card (Newer card, same arch, more power)
RTX A4000 Mobile(Laptop) / 115Watts : NVIDIA RTX A4000 Mobile Specs (Nvidia remove the product page???)
RTX A4000 ADA Mobile 150Watts: https://www.notebookcheck.net/NVIDI...Laptop-GPU-Benchmarks-and-Specs.744889.0.html
RTX A4000 ADA SFF / 70Watts : NVIDIA RTX 4000 SFF Ada Generation Graphics Card

 

[oharag]

The card retails for $1250; What you are seeing is markup from the demand...I was able to get it at retail before the price hikes
The Vulkan score an limitation has nothing to do with the x8 slot...but rather the TDP of the card being only 70 watts. The scores you see on geekbench are for the RTX A4000 and RTX A4000 Laptop GPU and not the RTX A4000 ADA SFF.

Geekbench is just one synthetic benchmark and sometimes it results can be skewed...that is why I am running test with software I will be using on it.

IF you look at their testing...this card scored even or higher with the a4000 (non ADA) at half the power

Review: Nvidia RTX 4000 SFF Ada Generation

Nvidia RTX 4000 SFF Ada Generation workstation GPU delivers in all types of workflows, but this level of performance doesn't come cheap

develop3d.com

You should only look at this card if you were previously looking at the RTX a2000, it is the same size, shape, and has the same power draw...just more VRAM and is faster. This card is very niche and only solves a hand full of use cases.

I blame Nvidia's purposely misleading naming conventions on the confusion of the cards. There are cards with the same naming scheme but different archs...like A4000 and A4000 ADA...it confusing and annoying.
This is not my primary machine; It will be a general box to take the workload and resources off of my other computers, mostly when I am mobile. One example is local AI code server; it stead of running it on my laptop eating up resources...I can run it on the MS-01 and connect to it. I think you can build a small ITX "Lunchbox" type machine and get a full fat A4000 for your needs.

RTX A4000 ADA/ 140Watts : NVIDIA RTX A4000 Graphics Card (Newer card, same arch, more power)
RTX A4000 Mobile(Laptop) / 115Watts : NVIDIA RTX A4000 Mobile Specs (Nvidia remove the product page???)
RTX A4000 ADA Mobile 150Watts: https://www.notebookcheck.net/NVIDI...Laptop-GPU-Benchmarks-and-Specs.744889.0.html
RTX A4000 ADA SFF / 70Watts : NVIDIA RTX 4000 SFF Ada Generation Graphics Card

I'm intrigued by the Oculink solutions also being tested and shown on the tubes. There is a discussion on STH as well. Run a beefy external graphics card outside of the MS-01 with its own power supply. I bought the MS-01 as a powerful Firewall/VPN server - but I'm intrigued to also run Windows and see what I can squeeze out of it for Engineering work. The work you are doing - and I think maybe coming down the line with Intel Arc/Nvidia for AI cards running at half-height will be interesting.

I'm definitely interested in the Card applications over the storage/networking cards people are testing. The PCIE slot in the MS-01 is only interesting for the graphics potential. I need to research further Oculink solutions - I know there are Nvme solutions - but I think the MS-01 kind of sucks for that. SSD1 is 4x4 SSD2 is 3x4 and SSD3 is 3X2. I wish they would have removed one of the SSD slots and maybe had 2X @ 4x4.

 

[playlogic]

I'm intrigued by the Oculink solutions also being tested and shown on the tubes. There is a discussion on STH as well. Run a beefy external graphics card outside of the MS-01 with its own power supply. I bought the MS-01 as a powerful Firewall/VPN server - but I'm intrigued to also run Windows and see what I can squeeze out of it for Engineering work. The work you are doing - and I think maybe coming down the line with Intel Arc/Nvidia for AI cards running at half-height will be interesting.

I'm definitely interested in the Card applications over the storage/networking cards people are testing. The PCIE slot in the MS-01 is only interesting for the graphics potential. I need to research further Oculink solutions - I know there are Nvme solutions - but I think the MS-01 kind of sucks for that. SSD1 is 4x4 SSD2 is 3x4 and SSD3 is 3X2. I wish they would have removed one of the SSD slots and maybe had 2X @ 4x4.

I am looking into the Oculink myself...although it doesn't fit my use case, I am interested interesting it. I won't be running Windows on this box as
I am sick of Microsoft and what they are doing is the OS these days. As far as the storage slots...I am ok with that we got...I repurpose my drives
most of the time from older builds into new builds, ordering newer based on link speed and purpose.

My fastest drive (980 PRO) on the 1st slot and get close to the max advertised speed. This has the main OS on it...the other drives (2nd) is the applications drives, which hit at about 5000-4000 and the last (2000-1000) is for "cold" storage ...the Garage. They are all 2 TB drives, the last 2 being 2 being P5 Plus drives, which top out at 6600MB/s...am I losing a lot of speed on the last one...yeah...but to be honest, no offense...I think people are too hung up on the numbers of everything instead of maximizing what is in front of them...I doubt most can tell a difference in the speeds in their work loads any way.

 

[playlogic]

Update 2: Link Speeds / Drive States w/GPU

A user posted in another thread, that they were having issue with a MS-01 unit (13900 ver.) not posting with an ARC GPU. This
seem to cause issue with M.2 SSDs loading and after working with support came to the conclusion that it may be a BIOS issue,
compatibility issue w/the SSD, and/or lane issue (too many being used)

https://forums.servethehome.com/index.php?threads/minisforum-ms-01-pcie-card-and-ram-compatibility-thread.42785/page-62

It's a comment in a thread, so I feel I might have missed some things in is replies or there is scattered information, i.e:
- What else was connected, externally?
- Was the power states re-adjust in the bios after connecting the GPU?
- System OS, (Windows, *.NIX)...I assume windows

Lucky I have not had these issue...don't want them...but my initial feeling is that this is a power state issue...either at
the bios level or in the OS. They did indicate that they were working with support on a bios fix

for ref...my unit is:
- MS-01 13900h / 96GB / RTX A4000 SFF /3 M.2 SSDs (980 PRO, x2 P5 Plus) 2TB each
- Ubuntu 24.04
- PL 1/2 back to 60000 - 80000 from the 40000 - 65000 when the card was first added.

Here are my stats, note that the (downgraded) on the LnkSta is the power state management at work
when devices are idled.

From lspci -vv | grep -P "[0-9a-f]{2}:[0-9a-f]{2}\.[0-9a-f]|LnkSta:"




Stats from "dmidecode --type 9 | egrep "Usage|Type|Designation"

 

[anewsome]

Update 2: Link Speeds / Drive States w/GPU

A user posted in another thread, that they were having issue with a MS-01 unit (13900 ver.) not posting with an ARC GPU. This
seem to cause issue with M.2 SSDs loading and after working with support came to the conclusion that it may be a BIOS issue,
compatibility issue w/the SSD, and/or lane issue (too many being used)

https://forums.servethehome.com/index.php?threads/minisforum-ms-01-pcie-card-and-ram-compatibility-thread.42785/page-62

It's a comment in a thread, so I feel I might have missed some things in is replies or there is scattered information, i.e:
- What else was connected, externally?
- Was the power states re-adjust in the bios after connecting the GPU?
- System OS, (Windows, *.NIX)...I assume windows

Lucky I have not had these issue...don't want them...but my initial feeling is that this is a power state issue...either at
the bios level or in the OS. They did indicate that they were working with support on a bios fix

for ref...my unit is:
- MS-01 13900h / 96GB / RTX A4000 SFF /3 M.2 SSDs (980 PRO, x2 P5 Plus) 2TB each
- Ubuntu 24.04
- PL 1/2 back to 60000 - 80000 from the 40000 - 65000 when the card was first added.

Here are my stats, note that the (downgraded) on the LnkSta is the power state management at work
when devices are idled.

From lspci -vv | grep -P "[0-9a-f]{2}:[0-9a-f]{2}\.[0-9a-f]|LnkSta:"
View attachment 36449

View attachment 36450

Stats from "dmidecode --type 9 | egrep "Usage|Type|Designation"

View attachment 36452

Did you tape the pins on the arc gpu before putting it into the pcie slot?

 

[playlogic]

Did you tape the pins on the arc gpu before putting it into the pcie slot?

I think this is a question for the other member @thekobk; I believe he is using the ARC GPU. The
GPU I have installed in the Nvidia RTX A4000 AFA SFF...I have not seen anywhere that you need to modify
the pins...I also think this is true for the A2000, since it is the card they use in their marketing...with a custom
cooler

 

[playlogic]

Update 3:

Ran a the Blender Benchmark headless using the 4.0.0 preset, seems that the 4.1.1 files are not avaliable:
- CPU AVG 1% - 3% while preloading scenes
- Avg temp 79c, Max 73c
- Max VRAM usage 13% of 20 Gigs

Blender - Open Data Benchmark

Blender Open Data is a platform to collect, display and query the results of hardware and software performance tests - provided by the public.

opendata.blender.org

 

[thekobk]

I think this is a question for the other member @thekobk; I believe he is using the ARC GPU. The
GPU I have installed in the Nvidia RTX A4000 AFA SFF...I have not seen anywhere that you need to modify
the pins...I also think this is true for the A2000, since it is the card they use in their marketing...with a custom
cooler

I didn't need to tape any pins the GPU was plug and play except for the odd nvme slot 1 issues I have.

 

[anewsome]

I didn't need to tape any pins the GPU was plug and play except for the odd nvme slot 1 issues I have.

Other users have reported that they needed to apply tape to some pins as a workaround for a bug in the MS-01 systems.

 

[thekobk]

Other users have reported that they needed to apply tape to some pins as a workaround for a bug in the MS-01 systems.

That's different, that bug presents itself as soon as the card is installed. The bug I have only happens when I have the card in and a crucial m.2 card in nvme slot 1. I can have that nvme in slot 2 or 3 no problem or I can have either a Samsung 990evo or a 970 pro in any slot no problem. That nvme also works in slot 1 without the card installed. It's a head scratcher.

 

[anewsome]

That's different, that bug presents itself as soon as the card is installed. The bug I have only happens when I have the card in and a crucial m.2 card in nvme slot 1. I can have that nvme in slot 2 or 3 no problem or I can have either a Samsung 990evo or a 970 pro in any slot no problem. That nvme also works in slot 1 without the card installed. It's a head scratcher.

Well I have Crucial M.2 NVME in all three slots of all five MS-01 and I do plan to use PCIe cards in all 5 units. So I'll probably need to solve this one as well, or buy 5 new NVME drives for slot 1 in all the boxes. I'll probably try taping the pins on the card as a first step, or I may end up using an Oculink 4 lane card, which I don't think would trigger this issue. Only one way to find out though.

 

[thekobk]

Well I have Crucial M.2 NVME in all three slots of all five MS-01 and I do plan to use PCIe cards in all 5 units. So I'll probably need to solve this one as well, or buy 5 new NVME drives for slot 1 in all the boxes. I'll probably try taping the pins on the card as a first step, or I may end up using an Oculink 4 lane card, which I don't think would trigger this issue. Only one way to find out though.

Is it the crucial T500? It might be model specific.

I haven't herd anything from support in a few days so I sent a email and got an automated response they are off until the 5th.

 

[anewsome]

Is it the crucial T500? It might be model specific.

I haven't herd anything from support in a few days so I sent a email and got an automated response they are off until the 5th.

Yes, I have 15x Crucial T500 2TB. Actually I have 20 of them. 5 are in Intel i7 NUCs and 15 in the MS-01s.

 

[playlogic]

Yes, I have 15x Crucial T500 2TB. Actually I have 20 of them. 5 are in Intel i7 NUCs and 15 in the MS-01s.

This is most likely model specific...When I first got my system,I ran off a P5 Plus w/the GPU before changing it to a 980 PRO and moving them to the second and third slot

 

[oharag]

Yes, I have 15x Crucial T500 2TB. Actually I have 20 of them. 5 are in Intel i7 NUCs and 15 in the MS-01s.

@anewsome I would like to be your PC Supplier from here on out! WOW - you like to tinker - I guess that's why you are posting here on STH.

 

[anewsome]

@anewsome I would like to be your PC Supplier from here on out! WOW - you like to tinker - I guess that's why you are posting here on STH.

A little over a year ago, I decided to try and self-host as much as I could, less reliance on cloud storage and apps. Filled 12u of rack space with cheap Xeon servers, about 2TB total ram and a bunch of cores. Plan worked out well until I couldn't stand the noise, heat and power draw. So I replaced it all with a cluster built with 5 MS-01 and 5 Intel NUCs I've had for years. Only 800GB of RAM across the cluster but I'm making it work.

The PCs are all just sitting in there now but I'm working on designing the mounts to get all 10 boxes arranged nicely and with maximum airflow from the fans at the front of the rack. Wiring is a mess, I know. I'm working on it. But, this ULTRA short depth rack is only 12 inches deep and it's quiet enough to sit beside my desk.

 

[GDB_Herk]

Hey all,

Just for anyone curious about the MS-01 Geekbench numbers under a non-windows OS
CPU 13900H / 64GB (DDR5 5200) - Which puts it at about an top M2 PRO Max performance

Micro Computer (HK) Tech Limited Venus Series - Geekbench

Benchmark results for a Micro Computer (HK) Tech Limited Venus Series with an Intel Core i9-13900H processor.

browser.geekbench.com

GPU RTX A4000 SFF / 20GB VRAM

OPENCL Micro Computer (HK) Tech Limited Venus Series - Geekbench

VULKAN Micro Computer (HK) Tech Limited Venus Series - Geekbench

Comment a photo from a previous post I did to show the current setup:
After a small bit of modding to the case…was able to get the 4000 SFF in without modifying the card itself. Couple of things:

had to remove the bracket at the back of the card…even the half height bracket would make it a tight fit.

Remove a secure clip from the top of the front panel and the PCie card screw bracket…it prevent most cards from going in.

shaved the rubber pads down with a razor do to the card having a solid backplate unlike the a2000 mobile

The bottom plate unscrews from the top, creating two halves…I am planning on modding the since if I want to replace it down
the line…it would be much easier the void the GPU warranty. Looking at doing a 3D print top

Will update as I do more testing

View attachment 36438

Could you provide some more photos to capture some of the modifications you made? Ie. Remove back plate, remove a secure clip, shave rubber pads might be touch to show).