glmark2 is not a great benchmark for measuring actual GPU performance. FPS numbers are so high that it mostly tests the display server/compositor and job submission latency, and the geometry load is far lower than real applications.
On Radxa Zero2 I have gotten glmark2-es2 scores as high as 1272 with Panfrost, which is almost twice what you report.
Though I canât comment on price, as Radxa sent me a Zero2 gratis, Iâll note that as a device which can show reasonable CPU performance whilst being powered from the host USB port of another computer, I donât know what more Iâd want from the board, apart from faster USB OTG and a better wifi antenna. Oh, and a more reliable power setup that doesnât hang the entire system or reset when power usage gets too highâŚ
Why do you think anybody would care about your opinions on SBC design and market strategies? At least I honestly donât care about any of these aspects and GPUs in general.
All I did was questioning your claim A311D would be âchokedâ by âmemory arrangement/bus/architectureâ (quoting you) and all that based on some mbw numbers posted somewhere that are in massive conflict with whatâs generally known about the SoCs in question.
In case you havenât tried yet⌠the dumb 15W USB-C power brick by RPi Trading Ltd. is a relatively good choice for 5V powering since at least allowing for a stable voltage even when +10W (+2A) consumption is exceeded.
Usually the cable between power brick and board is the culprit based on AWG ratings of the power lines:
any test case that we can try to reproduce this issue? also cooling might help which was unavailable when we sent out samples.
$ echo 2208000 | sudo tee /sys/bus/cpu/devices/cpu[02]/cpufreq/scaling_max_freq
$ for i in {1..6}; do openssl speed md5 & done
will hang after a few seconds, at least when powered from a USB port on another computer, rather than a proper power supply.
Usually I have the maximum frequency set to 1704000 kHz, where the system is stable enough. I think that images Iâve tried have defaulted to a lower frequency than that anyway. So maybe the fix to this problem is just: âUse a better power supply if you want to increase the frequency.â
Whatâs the maximum consumption when executing stress-ng --cpu 6 --cpu-method matrixprod? With A311D at 2.2/1.8 MHz maybe 5W above idle consumption, right?
Then use an average USB-A to USB-C cable, send 1A over it and measure the voltage at the receiving side. The voltage drop caused by super tiny power lines in average USB cables is usually the biggest problem source but ofc with a power source already limited to 900mA (the average USB3 port on a PC) everything above 4.5W will also cause problems regardless of contact and cable resistance along the power path.
Quite the opposite at least with sustained loads. Throttling and DVFS will result in cpufreqs being reduced when starting to overheat and as such consumption will drop significantly. But ofc this doesnât help with consumption peaks since throttling needs some time to kick in.
I was sent a pre release to evaluate as I keep repeating, that is what I am doing, who are you to tell a community member what is or isnât needed.
Its a bad product as its premium doesnât fit its designed remit of âZeroâ, it been stuck on hold because of this for months and now no longer is an intermediary product to the Rock5b as it will now be released after.
Its GPU historically across the internet never met expectations and having a 4gb single version where its premium negates the possibility of lower ram versions, where another soc will have to be used.
The format had to be changed so it is only zero in name and whilst a huge wealth of IO is lost to pigeon hole it in the most none Zero product ever.
I was sent a product to evaluate it it and I am evaluating it and as far as I am concerned my feedback to Radxa is its an absolute stinker on just about every level from concept, implementation, probable cost and likely performance.
That is my honest end evaluation of the product I was sent to evaluate and I have, it just has far too many chinks and flaws
You seem to be the âthere exists only one single use caseâ guy, right? GPU, GPU, GPU, GPU.
The reasoning for Zero 2 was made pretty clear by Radxa: CSI/DSI in a compact form factor and âtarget AIoT market with NPUâ. Others explained their interest in Zero 2 being a super compact server (though w/o wired networking) or even just the sheer CPU power at this format (currently amongst the somewhat affordable ARM SoCs only A311D2 and RK3588/RK3588s are faster).
And the reasoning for potential SKU variants with S905D3 is also pretty obvious: pin-compatible drop-in replacement while providing same features: CSI/DSI and (weak) NPU.
I would believe once @RadxaYuntian (busy with RK product launches right now as everybody else at Radxa) will have time to play around with Amlogic stuff again then sorting out the FIP stuff I bet A311D cpufreqs get a boost of a few hundred MHz each and S905D3 will boot as well (with different BLOBs ofc).
That is exactly what I am thinking as likely they have the ability to source much more cost effective CSI/DSI SoCs with maybe the RK3566 being a much better fit in function and cost.
That the Zero idea with the s905y2 dongle SoC was a great idea but has detractors due lack of CSI/DSI but it has nothing to do the the RK launch as the pre production samples where received far in advance end of April / start of May and had a great working image and was ready to go but has sat frozen because its âpremiumâ is just ill fitting of a Zero format.
Because of this, even though I was a big initial backer I honestly think it missed its opportunity and forcing this now is a bad idea.
I still have no idea why you compare the sales strategies (pricing and marketing) of two different SBC makers but if you think what Khadas charges gives us an idea what Radxa will charge have a look at these refreshing numbers for the just now announced RK3588S thingy by Khadas:
Bare board with RK3588S + 8GB RAM + 32GB eMMC + 2 antennas for 240 bucks. Now compare with what Radxa said to charge for Rock 5A (existing at least as samples since months BTW): less than 100 bucks (for sure with no or a less capable Wi-Fi/BT chip but IMHO itâs pretty obvious how pricing and marketing differ between Khadas and Radxa).
Yeah whatever TK.
Thnx.
That is just huge. Not the same form factor not the same market. I need the radxa 2.
