News Intel's newest E-core-only "Twin Lake" CPUs are on the way, starting with Intel N250

[bit_user]

That's for barebones, no case etc. You're probably going to end up at $300. It's not too hard to find AMD fast 8-core mobile mini PCs starting at $300 barebones, but with the case. For example, here's Minisforum UM773 Lite at $294.

That's without RAM or storage, so all we have to do for the ODROID-H4 Ultra to match it is add case and power supply. For a 60 W PSU, which they claim is sufficient if you're not running hard disks, add $9.40. Then, for the smallest case, add another $10. The optional case fan adds another $4. Equivalent price: $243.40.

Out of curiosity, I priced out a Crucial 16 GB DDR5-4800 SO-DIMM at $45 and a Team Group MP33 M.2 2280 512GB PCIe 3.0 SSD at $37. That brings the price of a working ODROID-H4 Ultra to $335.

For my system, I went with a 32 GB (dual-rank) DIMM and a SK Hynix P31 Gold SSD. The main reason for the RAM upgrade isn't because I need that much capacity, but rather that dual-rank performs slightly better.

Note: I'm not saying which is the better value or better in absolute terms. I just like to make the correct data available so people can make accurate value judgements for their needs.

 

[brandonjclark]

This:

​

Amazon.com

​

I don't have any other meters to compare it with, but it seems to get good ratings.

It's plugged into a power strip that's plugged into my APC SmartUPS, which does line-interactive power filtering. So, my power shouldn't be terribly noisy.

I'm going to buy it and test it out. I have an n100 stood up to run a few containers and I'm interested now.

 

[bit_user]

I'm going to buy it and test it out. I have an n100 stood up to run a few containers and I'm interested now.

Try installing a package called stress-ng and running it with options: --cpu=4 --cpu-method=matixprod That's one of the most effective methods I've found for stressing the CPU. By itself, this runs my power meter up to 34 W.

I also found a good GPU torture test: glmark2 --run-forever -s 1280x1024 -b buffer:columns=200:interleave=false:update-dispersion=0.9:update-fraction=0.5:update-method=subdata I'm not sure how many of those parameters really matter, but the key one seems to be the last. It's a lot more stressful than most of the other glmark2 benchmarks. I discovered it based on watching the entire set of benchmarks while watching my power meter to see which ones spiked it the most. For whatever reason, that specific test case is a lot more stressful than any of the other "buffer" benchmarks. Another counterintuitive thing I discovered is that it burns significantly more power to have it render onscreen than to an off-screen buffer. In general, I think the reverse tends to be true. By itself, this benchmark runs my power meter up to 43 W.

When running both CPU and GPU stress tests, try running the CPU at between 1 and 3 threads, because that GPU stress test seems to need most of a core. When I combine both, I'm getting about 48 W.

To see SoC package power & temperature, try this: turbostat --quiet --Summary --show=PkgTmp,PkgWatt --interval=0.5 Then, you can capture it to a logfile and later plot it for seeing how fast your CPU throttles based on either power or thermal limits.

So far, all of this is with the stock Intel-recommended PL1 of 12 W and PL2 of 15 W. If you're curious, here's a tool I found for overriding the PL settings on your machine. On my Dell i9-12900 it absolutely works as advertised. On my N97, I don't really see much difference, I suspect because there's a different set of thermal-oriented limiters that comes into play (note: I do mean something different than the overall throttling temperature) and I'd need to find out how to override those MSRs.

https://github.com/horshack-dpreview/setPL
​

In the course of modifying the values, it shows you what the old values were, which is how I confirmed what it was running at. Each time you reboot, the values get reset, so don't worry about playing around with them.

When I set that to some silly values and run the CPU stress test on 2-3 cores + GPU stress test, that's when I get the peak reading of 54 W. With these settings, running a standalone 4-core CPU stress test uses 40 W (instead of 34 W). By itself, these settings result in the GPU stress test using 45 W, instead of 43 W, but the reading jumps around enough that maybe there's no real difference for the GPU.

I think the main effect I see of tweaking package power on this board is really that it seems to make the CPU cores throttle a bit less, either when they're running by themselves or along side a GPU workload. Weirdly, it doesn't seem to have much effect on how much power the GPU gets.

BTW, these tools are not new, but the OS I'm using is Ubuntu 24.04. Oh and I booted with the kernel parameter mitigations=off !
; )