Minisforum MS-02 Ultra Performance
Before we get too far into performance, let us quickly touch upon TDPs. While Minisforum runs the Core Ultra 9 285HX chip with a TDP (PL1) of 100 Watts (and 140W for PL2), this is only valid for systems without a discrete video card installed. If a video card is installed, then the BIOS will automatically dial down those limits to 90W for PL1 and 110W for PL2, respectively. Which is to say that CPU performance in both bursty and sustained multi-threaded workloads will take a hit if a video card is installed.
It is an important distinction since, although the MS-02 Ultra does not come with a video card, the system is specifically designed to accommodate one (and indeed, it is one of the system’s distinctive design features). So many (if not most) users in Minisforum’s target market will be impacted by these TDP limits. It is an AI system, after all, and it needs a good discrete GPU in order to be one.
Geekbench 6 CPU
Since we are looking at a system that is primarily being marketed as an AI system, we will go ahead and compare it to a similar AI system: Minisforum’s own MS-S1 Max, their Ryzen AI Max+ 395-based system. This offers us an apples-to-apples comparison between the highly integrated AMD platform and the more discrete Intel platform that Minisforum is using here.
Starting things off with a look at the high-level Geekbench 6 cores, the two systems trade blows depending on whether it is a single-threaded or multi-threaded workloads. The 285HX-powered MS-02 Ultra holds a ST lead by about 5%, but then it trails the AMD-based MS-S1 Max by around 5% as well.
A look at the sub-scores shows that these results are pretty consistent across the individual benchmarks as well.
The MS-02 Ultra is pretty consistently ahead in ST workloads. But more often than not is behind in MT workloads.
Geekbench 6 Compute
The Geekbench 6 Compute scores are nowhere close, however. The integrated GPU on Intel’s 285HX is not meant to stand up to the Ryzen AI Max+ in either memory bandwidth or compute throughput, and sure enough, it does not.
MLPerf 1.5
We have also run MLPerf to get a better look at machine learning/AI performance. For these tests, we have run it once with just the integrated Intel GPU available, and a second time with a low-profile GeForce RTX 5060 card installed. This is a prime example of the type of card that can fit in the system’s double-wide HHHL PCIe bay.
Here are the results with the Gigabyte NVIDIA GeForce RTX 5060 OC LP card. (Amazon Affiliate)
The integrated GPU doesn’t fare all that poorly here. Still, it is a step below even a discrete RTX 5060-class card. Which is a big part of this system’s reason to be: to have the space to install such a card.
I wish they’d have managed to make this fit into 2U so you could fit 2 of them side-by-side in a rack (via a presumably-optional bracket). As it is, it’s about 8mm too tall.
The MS-01 has a similar problem, where it’s ~3.5mm too tall for 1U.
These are slick. Wish I had a use case for one.
I would love to see a comparison between native Intel ECC memory support and the software based ECC capabilities offered on some of the mini platforms. I have soft ecc enabled via bios on 4x alderlake mini boxes and they have been stable for 1 yr+.
I wonder when MCIO pinout and cable finally landed in SFF or normal desktop mainboard? So many benefits come from it.
The article says that the 2 M.2 slots on the PCIe board are limited to PCIe3. The Minisforum site says that it is limited to PCIe3 for 8TB, and PCIe4 for <4TB. Was a less that 4TB SSD tested to see if it runs at PCIe4?
The lack of a TB5 name on the front most likely means some part of the data path is make by someone other than Intel. While they released TB to be an “open standard”, they still own the name and are the only ones that can grant its use. It’s been a good thing, allowing USB4 and now USB4v2 to be compatible. It’s dumb of them, really; in the couple of years nobody is going to bother paying Intel for the name at all, and as a result it will become worthless.
15W power delivery is nowhere near the TB5. Please turn you bs knob down a little bit.
That slot for the 25 Gbit NIC I see as a long term curse: you can’t use it for a 16x device even though it’d physically fit. There are several reasons for it. The first is that there are two sources of PCIe lanes going to it which only a few devices actively support in this fashion. The second is that it is 12 lanes in total. Leveraging 12 lanes in a 16x physical spot is permitted under the PCIe spec but few 16x devices support. Furthermore, those 12 lanes I believe are arranged in a nonstandard fashion for those devices that do support 12x. Thus the best fit for that slot would be a device that inherently needs bifurcation but users have to realize that one of the middle devices is not going to work as it doesn’t have PCIe lane going to it. For example, a quad M.2 carrier card would not have the second M.2 slot functional while the first, third and forth work fine.
I’d almost advocate for having all the lanes from that PCIe slot come from the PCH but it’d be bandwidth constrained due to the DMI link. I’d have implemented the slot as a 4x physical slot and then leverage an 8 lane MCIO connector to a proprietary card to get the additional two M.2 slots off of it. The 8 lane MCIO could find other uses, though the system would likely be physically space constrained for other cards.
Dual 10 Gbit NIC on the motherboard would have been nice to see.
I do hope that Nova Lake adds more PCIe lanes to the desktop platform. The oddities here are all examples of lane shortages forcing compromises into the design.
Why would anyone prefer this crap over Minisforum’s won BG-795SE MoBo with extra 25GbE NIC ?
Or mini-ITX MoBo with 9950X and NIC as needed ?
Why are you hyping those included NICs so hard ?
Hi Tinkering Ted – Just as a FYI as I have a system with the X3D version of that motherboard next to me, and we have another one with the non-X3D motherboard as well. The MS-02 has better I/O like the USB4 ports, it can also take an internal RTX 5060 LP while it has the 25GbE and 10GbE/2.5GbE ports, and more SSD. That AMD board is great, but this has way more expansion potential.
@Andrew
I based that section off of their user guide, which has the most comprehensive details on the matter.
To quote said guide: “For compatibility, the E810 expansion card’s NVME slots are set to PCIe 3.0 x4 by default. You can change them to PCIe 4.0 in BIOS, but after switching it is recommended to run a storage benchmark. If you see anomalies, revert the slot to PCIe 3.0.”
At least there, Minisforum isn’t guaranteeing anything. It can run at PCIe Gen4 speeds, but they clearly aren’t 100% confident about it. Which is why it defaults to Gen3 speeds (and why you’d need to go out of your way to enable Gen4).
Still, I’ve tweaked the language in the article a bit to make it clearer that Gen3 is the out-of-box setting rather than a hard technical limitation.
@Patrick Kennedy:
Who cares bout USB4 on such box.
Also BD-795SE has way faster PCIe5 lanes, that can be split with a simple splitter and use port bifurcation.
PCIE5x4 is enough for 100GbE. One could use 8 lanes with PCIe4 and be still left with 8 PCIe5 lanes for GPU.
Not to mention there are mini-ITX boards with AM5 that have 2x M.2 PCIe5 directly on CPU + PCIe5x16 for GPU etc.
@Illrigger . Read the block diagram. The USB4v2 80Gbps ports are supported with an Intel JHL9580 which is TB5 certified. Minis Forum just chose not to call it TB5.
@Tinkering Ted i care.
@spuwho: Without Intel (paid) certification for the whole device they can’t call it TB5.
When you say adding the Sparkle Arc A310 was a mistake, do you mean compared to the 5060 or were there problems more broadly?
Have one currently running transcoding/encoding duties on a plex server and was potentially planning to build an MS-02 with the A310 as a replacement. Bad idea?
Billy Baroo – I think the biggest reason was that the A310 is not a huge upgrade over integrated graphics if you just want basic GPU/ transcoding capabilities. It also does not offer the bigger gaming and small local AI model jump that you get with the RTX 5060. Better said, it is not a bad card. It just left us feeling like it was not a big enough jump to warrant adding that GPU to this system. Here is an example https://browser.geekbench.com/v6/compute/compare/5697728?baseline=5349589
I go KVM for more flexibility and have been running Debian KVM and Proxmox, so VM license is no longer my challenge.
The 350W power supply makes this system a non starter for anything AI-optimized: there is no way it can accommodate one of the nice 2-slot blower style GPUs (~800w), not counting the CPU, RAM, 4xNVME and NICs. Once again, Minisforum snatches defeat from the jaws of victory.