Looks like there’s room for more tuning:
1 Like
Based on what I read from users, it seems like the GMKtec model has way worse thermals, and that doesn’t spell good for hardware longevity. Since it’s less repairable than a normal PC, for me it’s a big no-no. Also, the community around Framework seems way more active.
7 Likes
I do wonder what “worse thermals” means. Like, if it’s faster, it would run hotter, naturally, and would need to ‘exhaust’ more heat, naturally. Or does it mean it ‘retains’ more heat, and therefore hitting thermal limits sooner, and running slower (not likely slower, base on the benchmarks)?
Depends on the component selection and quality. e.g. You don’t expect servers to run idle, they run, and run and run. Like CPUs can run at 90c+ all year round…but the board itself, really depends. Granted that this isn’t a ‘server’…but given the strong push on AI….it might as well be an home AI server.
Need to classify why it’s more active. Active for what reason. Even Framework recognized that if things work well, people naturally write less about things working well.
With that said, I’m going to wait for more reviews and direct comparisons. But raw performance alone seems to be a tad bit behind in the Framework camp, based on sample size of one. Main point of the first post: There’s some performance to be picked up.
Plus, really, the real long term options are only HP and Framework at this point (support, BIOS…). GMKtec…I’m not sure, when was the last time you saw them releasing BIOS update for a 4year+ product?
1 Like
Why would the Framework Desktop show weaker performance than the EVO-X2? It is the same APU. Some reviews presumably were running on “Balanced” power profile (~10sec 115W boost/100W sustained) as this was the default setting on the Framework but it is dead easy to go to “Performance” (140W boost for 10 min/120W sustained). If you dare, with ryzenadj you can also do undervolating, overclocking and a lot of things, also unsafe ones.
Exactly…why did it behave the way it did? Some digging required here.
Not really the point here.
…hum…the BIOS maybe?
But is it really slower?
I can’t really compare my Framework Desktop results directly to L1 Techs Windows results, but on 1440p “High” preset (where FSR is automatically turned on to “Quality”, and in the graph that is shown in the video it specifically says that upscaling might be included if it is part of the preset) get 74 min FPS and 86 FPS average on openSUSE Tumbleweed. On “High” without FSR (which is considered a “custom” setting by the game) I get 53 min FPS and 61 average FPS.
From a sample size of one vs sample size of one, done by a single reviewer….yeah, it seems to be slower.
Would be interesting to know then why my Linux Framework Desktop is running as fast as his Windows Evo-X2 though, even though I am using the settings that limit it to basically 100W TDP. I doubt Linux is able get so much more performance out of Cyberpunk but what do I know.
But even more interestingly, why would the 64GB version of the Evo-X2 be slower than the 128GB version?
Rather puzzling.
Judging by other tests, it hits its thermal limit and has to clock way down, and its cooling solution doesn’t have the ability to dissipate the power that it generates. It’s built like a laptop, it’s designed to runs in bursts and any long processes will exceed its ability to dissipate the heat. I don’t think he’s the only reviewer who has found this with the GMK, it can beat the framework in short tests, but loses in long tests.
On the other hand, the framework can run full out all day without exceeding its thermal dissipation capacity. I’ve run long tests and never had it go much over 65c.
I wonder if it would be able to if it were overclockd similarly to the GMK?
This one shows the GMK winning in GPU but HP and Framework doing better in CPU tests.
3 Likes
Need to figure out why the FD is not winning out the burst tests then.
(Watching the video now).
Update:
Finished watching the review. If Framework can address the burst performance, then it’ll be the clear winner in raw performance. HP is only winning in BIOS support, on-site support, extended warranty and case (subjectively. I suppose people can buy a better case with FD’s mainboard-only purchase). At the price of the HP, you can almost get two FD.
I’m still a bit torn when it comes to the AI Max+ 395 (regardless of which system builder). The raw non-AI workload performance is not really better than a 13980HX (my current 2-year old laptop…it can sound like a jet engine). Maybe I should view the FD as a quality of life upgrade, quieter…with local AI:
CB 2024
CB R23
I bought the GMKtec EVO-X2 on Amazon as soon as it was available.
I tested it for three days and returned it because it was really too noisy!
Now I would like to test the new Beelink GRT9 Pro as soon as it is available.
According to some tests, it seems to perform quite well…
I have been working with a Beelink GTR7 Pro (OctalCore AMD Ryzen 9 7940HS) for 2 years and I am quite satisfied with it.
It should not be a problem to run the Framework Desktop like the GMKtec EVO-X2. In Windows and Linux you should have easy access to power settings. By default the Framework Desktop is set to Balance (115W boost/ 100W sustained), you can set it to Performance (140W boost/120W sustained) with a simple mouse click, and that option will be remembered at reboot. 140W will push the cooling system of the FD to the limit as well but apparently it can handle it without temperatures pushing beyond 90°C. The APU throttles at its Tjmax of 100°C.
I don’t know the precise power settings of the EVO2 but if you are so inclined you can even overclock the FD beyond that, for example setting it to 140W sustained. For that you would need ryzenadj though, a tool available for Windows and Linux. You should only use it though if you know what you are doing as some of the settings you could do there could potentially damage your machine.
I’ve ordered a Framework Desktop for its combo of compute and portability. I run physics simulations that max out all the cores for many hours and use a lot of memory, but I also want to easily move the machine back and forth from home to office frequently. There are few other similar machines coming in the market (mostly from China), such as the EVO-X2, with these AMD Max 395 CPUs, but the FW wins for three reasons: likely better thermal management, very likely better long term driver/firmware support, and repairability if I drop the thing (I will commuting with it several times a week). Pay about a 20-25% premium over the Chinese competition though, but I think that’s acceptable. The only non-FW model that tempted me was the Beelink GTR9, which I think may have decent thermals, but I’m not convinced there won’t be driver and repair support pain.
1 Like
What kind of compute/processing unit do you use for those simulations?
How was laptop form factor ruled out from your selection process?
Why not the HP?
The physics simulations I do are 2D and 3D finite-element analsyis electro-magnetic simulations for RF circuit design. They can get complex. Currently running on a bunch of different machines, mostly mid laptops and desktops, and one pretty nice threadripper based machine. Compute time isn’t always critical, but when it is, it sure is.
I looked hard at laptops, and there certainly are some nice 32-thread capable laptops out there, but either I have grave doubts about their thermal capability (lot of gaming laptops similarly priced to the Framework Desktop), or they are way more expensive. Which also rules out the HP Z2 G1A - in my country it’s 60% more expensive than the FWD (but, hey, you get a keyboard and mouse!). I’m sure it’s pretty decent, just pricey. So the FW wins on being portable enough (just), compute, and I think longevity (drivers, repairs, maybe upgrades).
Maybe in six months or a year there will be loads of machines around that would also fit the bill, I don’t know.
So it’s the traditional processor core compute you’re after, and not really about the Radeon 8060S? Not sure what got you interested in the 395 then. There are plenty of options for you in the traditional compute horse power.
Update:
I’ll give an example:
Give this chart (take it with a boat load of salt): PassMark CPU Benchmarks - Multithreaded - All - Page 1
EPYC 4585PX, score 71689
AI Max 395, score 53635
Grab a board (say, Supermicro): AMD EPYC 4005 Series | Newegg.com
Grab the processor: AMD EPYC 4005 (5th Gen) 4585PX Hexadeca-core (16 Core) 128MB L3 Cache Socket AM5 170W - 3D-V-Cache Technology Processor 100-100001561WOF - Newegg.ca
Grab some ECC (if you need that, your work sounds important) RAM and the usual PC stuff, and you’re off to the races. Even a 9950X (X or X3D) would come out ahead of the 395.
But yeah, if you just want a mostly pre-built solution then maybe Framework fits the bill…and everything soldered on the board. Just saying, raw x86 compute is cheap if you don’t need the AI stuff with minimal GUI / graphical requirements…because it sounds like you won’t benefit from the 8060S and memory bandwidth to the iGPU. That portion of the cost might be better spent on real modularity and repairability if that’s also a consideration. The moment you mentioned ‘max out all the cores’…that to me sounded like a non-AI workload.
As I’ve said, sweet spot for compute, form factor, price, thermals, support and repairability. I’ve looked at putting together a comparable ITX form factor machine and it will be slower, less memory and I like that someone else has sorted out the thermals (non trivial and widely underestimated).
1 Like
Yeah I was also getting some cold feet after seeing that Lvl1 video, especially when I found it available for immediate shipping at the same price as the FD, with a 2TB disk included.
Of course we would all like to assume that it’s just a matter of settings somewhere, be it BIOS or OS or some combination, but I’d still like to have actual confirmation whether or not that’s the case. Because the performance differences weren’t what I’d call minor…
BTW I worked on Cinebench 24, greetz from Maxon