| 09:27 | topi`> | it seems the GC7000 Lite, which the 8m has, only has 16 shader cores. This translates to 64 GFLOPS, which is just 3 times the nominal perf of the Videocore-IV in Raspberrypi |
| 09:28 | topi`> | but it's not bad either, it's a bit short of premium smartphone perf in 2012 (Adreno 320 in Snapdragon 600) |
| 09:56 | jnettlet[m]> | yes that is correct. I will hopefully have some real world specs soon |
| 10:03 | Ke> | when will you have the new cubox available? |
| 10:03 | Ke> | what is the cpu freq btw.? |
| 10:41 | topi`> | we have evaluated both the Compulab utilite2 (QS600 based) and the Hummingboard for a customer, and the 3D stuff on the QS600 is clearly way faster |
| 10:42 | topi`> | I hope the 8M will do much better |
| 10:42 | topi`> | (otherwise, I cannot recommend the Utilite2 to anyone) |
| 10:42 | topi`> | mainline kernels fail to boot mystically |
| 10:44 | suihkulokki> | I wonder which SoCs survive the Qcom-NXP-Broadcom merger for long term availability.. |
| 10:44 | topi`> | too many mergers going on... |
| 10:44 | topi`> | they should complete the qcom-nxp merger first |
| 10:45 | topi`> | from my point of view, all these APQxxxx attempts from Qualcomm to make decent embedded SoCs have failed to a certain extent. |
| 10:45 | topi`> | from linux point of view, everything from NXP is much, much better documented |
| 10:46 | topi`> | I'd venture to say it's impossible to get proper docs from Qualcomm. If Compulab was unable to get the required docs, then who can? |
| 10:47 | topi`> | and Rob Clarke has to do reverse engineering on the Adreno |
| 10:47 | topi`> | that's just pathetic |
| 10:47 | topi`> | (OK, the same applies to Vivante, but anyway the GC parts are like black boxes and not everyone needs the 3D stuff) |
| 10:48 | topi`> | I think the worst outcome from the merger would be worse availability of docs for the NXP parts |
| 10:51 | suihkulokki> | topi`: re documents, you are correct |
| 10:51 | suihkulokki> | topi`: but lots of the freedreno kernel code comes from codeaurora these days |
| 10:55 | suihkulokki> | really the most effective way to force the hand of vendors |
| 10:56 | suihkulokki> | 1. reverse engineer enough that mainline has somewhat working driver |
| 10:57 | suihkulokki> | 2. wait for customers contact vendor "we have problems with mainline.." "don't use mainline, use our driver..." |
| 10:58 | suihkulokki> | 3. "we tried your kernel but it is old and has problem and we miss all the nice features of newer mainline..." |
| 10:59 | suihkulokki> | 4. "sigh, we'll assign someone look and fix issues in mainline, BUT WE WON'T SUPPORT IT" |
| 11:03 | Ke> | Qualcomm categorically means not user controllable signed bootloader |
| 11:03 | Ke> | good companies get eaten by the more predatory ones |
| 11:05 | topi`> | or the ones who have the biggest army of lawyers |
| 11:05 | Ke> | quite |
| 11:06 | topi`> | I'm waiting to see the battle at the datacenter field when Qualcomm is going to attack Intel and both have huge armies of lawyers |
| 11:06 | topi`> | Qualcomm might be a newcomer to the field with lots of disadvantages, but they also have a pretty decent war chest and experienced, veteran lawyers |
| 11:10 | suihkulokki> | That is assuming the broadcom leveraged buyout doesn't choke the whole combined company under debt |
| 11:33 | topi`> | yeah that sounds like a stupid buyout to me |
| 11:34 | topi`> | why can't they just compete like everyone else :) |
| 15:17 | jnettlet[m]> | topi`: but you were evaluating the etnaviv 3d acceleration right? It is still about 50% the performance of the binary driver. I expect the iMX8M to be about the same, most likely worse. |
| 15:21 | Ke> | flops sounds like spec evaluation |
| 16:23 | jnettlet[m]> | one of the big restrictions of the iMX6 series was the screw up they made on the memory controller. That has been fixed so the iMX8M should have a lot more memory bandwidth available for the GPU. This alone will help a lot |
| 16:24 | jnettlet[m]> | fyi, the BSP is also now using the imx-drm driver for the iMX8M. the binary GPU driver can handle dmabuf file descriptors and the likes. Still digging in to how far the integration goes, and also the inner workings of the new VPU |