FSD "solved" sounds scarily like funding "secured".
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
I'm only logically comparing and contrasting. If chip A runs all the NN models necessary for LVL 5 perception including way more NN models than chip B using 75% less processing power. Then Chip b should theoretically be able to handle the amount of NN models that chip A does even if it came with some overhead (even with the difference in efficiency and accelerators). The only other contingency would be planning, but again like i said i don't believe Tesla software are as advanced as some people believe they are.
Not if chip B is memory constrained. The whole network has to fit in memory. And while it may be able to run, it may not be able to run in real time to meet the requirements of driving.
just to be clear. everything you need to know about this process was posted further upthread. I do not distribute Tesla binary firmwares for legal reasons, so you need to find some other source than me. There was one and I listed it, now that it's closed, I do not know of any other public sources I could direct people to. If anybody knows better, perhaps they'll update here in the thread with more sources.
DrivePX2
8 GB LPDDR4 (50+ GB/s)
Xavier
16 GB 256-bit LPDDR4 137 GB/s
Pegasus
16GB HBM2 RAM 1TB/s
I don't have datasheet for eyeq4.
I know elon mentioned memory but you can't just take him at his word. The biggest problem with DL GPU chips was when you had multiple chips trying to access memory which was problematic with a bus bandwidth speed of 50 GB/s which nvidia fixed by utilizing nv link 2 (1TB/s).
But again its not a reason why Tesla's current software CAN'T run on ap2.0.
it's a complete rewrite. But I am not big NN expert to dive into any details here.
Moderator note: It looks like 2018.39.5 qualifies as the first 9.0 release, and someone already created a thread for it, so I've moved 50-some-odd posts about 2018.39.5 to that thread:
Software Update 2018.39.5 a1c2332 (first 9.0 release)
Posts relevent to this particular release should go to that thread, as with most firmware threads.
I might break out some other topics to their own threads later. Please note that manually splitting threads like this is an imperfect process, and I apologize in advance for any confusion created.
Bruce.
Software Update 2018.39.5 a1c2332 (first 9.0 release)
Posts relevent to this particular release should go to that thread, as with most firmware threads.
I might break out some other topics to their own threads later. Please note that manually splitting threads like this is an imperfect process, and I apologize in advance for any confusion created.
Bruce.
KArnold
Active Member
That was fast. It seems to imply they are monitoring this thread.
Wish they had the gonads to comment now and then.
The total power consumption of all the onboard processing amounts to a piddly amount and would be insignificantly different from it even being idle.
Remember those NDAs that were mentioned for Early Access/Adopters, whatever you want to call it... the R&D team at Tesla likely has even much more restrictive and easily enforcable NDAs. So please don’t question their gonads as I’m sure they have massive gonads to take on this incredibly difficult problem as their day job when they could easily go work at Google, Apple, or any other tech firm in Silicon Valley working on AI and/or computer vision and not have the weight of Tesla’s mission and Elon’s twitter habit on their shoulders.
If they are reading this... hats off to you guys on what you’ve accomplished to date.
The AP computer can pull about 200W, which means over an hour it takes 0.5-1 mile off your range depending what car you have.
Shock-On-T
⚡️⚡️⚡️⚡️⚡️
Our eyes are as far apart as practical without incurring other evolutionary penalties.Thanks goodness our eyes are three feet apart or else we might not have depth perception.
Cars can go wider, and perhaps should.
Trebek1762
Let's play jeopardy
I keep checking the site...just to make sure it doesn't 'accidently' surface...
