Hi,
thanks for all the comments.
I was coming from that end:
https://electrek.co/2017/09/28/tesl...g-platform-autopilot-media/#jp-carousel-52503
So the entertainment PCB has a Intel SOC (Quadcore X86 + Intel GPU) in combination with an additional automotive Micro (NXP) on board (and additional chips for communication and interfacing). This PCB is "sandwiched" with the ADAS PCB to join the same cooling system (water cooling) . The ADAS PCB has the two NVIDIA Parker (in HW2.5) on board plus the NVIDIA Pascal GPU and one Safety MCU (Infineon AURIX) (and additional interfacing devices).
So for me personally the Intel SOC potentially could also be replaced with something different ... say by a Samsung Equinox, which would make more sense, since this device is known supporting Infotainment application. Whatever, that's only a side note and explains the background of my question. Is still assume that the Samsung SOC in on the ADAS PCB. If remember correct still in coexistence of NVIDIA chips? I read something in reddit? Was it also
@verygreen?
As to the replacement of NVIDIA SOC by a Samsung SOC on the ADAS PCB: I object to the statement, that this a no-brainer. Switching from one to another MCU or SOC is always huge pain:
1) never pin-compatible!
2) SW stack totally different even basing on the same ARM architecture (or at least similar). NVIDIA has his DRIVE OS in place which I believe TESLA is using to "drive" Parker and Pascal (the latter especially because of CUDA). So moving away from that environment is more or less a full restart in SW development.
I cannot judge on the quality of the Samsung SOC and by that I don't know whether it is superior to the Parker SOC. But I can tell you: Samsung has to digest the automotive DNA as NVIDIA has done in the past. Samsung has to learn a lot!
It might be easy to support an infotainment solution but the field of ADAS is a lot more demanding in terms of:
1) reliability in general
2) functional safety
3) temperature ranges
4) security (here also the traditional OEMs lack a lot in experience to be fair)
To compare the Parker/Pascal approach with one unknown Samsung / TRIP approach is also not fair. It should be compared with Xavier / Volta.
In general there is an industry trend to move "everything" to AI. There is a huger discussion what is the perfect infrastructure for handling the problems of autonomous driving. Candidates are:
CPU will be always play a role as the master controller, AI should not run here
GPU was the first step of acceleration but GPU architecture has to evolve a lot to compete with other technologies
FPGA has some benefits because of it's flexibility but also is not the ultimate solution, needs also to evolve
ASIC which is a broad field, and I count the TRIP as an ASIC.
What I learned: Key is to avoid DDR memory access when executing the algorithm. Chips like the TRIP claim - besides the massive parallel computing - to always keep the data flow in the chip and the local SRAMs. Studies show that any storage to DDR memory severely hurts performance, power and latency. That to my knowledge is one of the weakest point of the GPU approach. That is also the reason why GPUs always have the top notch memory types and with that the highest bandwidth available.
If a TRIP like chip can really solve that issue described above must be proven in the wild. The Google TPU chip is known to follow this concept and is considered as a success.
Finally, there is a big race ... and every single day there pops a new start up (in automotive) which wants to address the automotive market.
Best regards,
Frank
