roon wrote: ↑
Sun Jun 02, 2019 9:33 pm
TankMarvin wrote: ↑
Sun Jun 02, 2019 11:36 am
That is just for a single chip, the associated hardware such as memory/storage, connectivity, etc will increase that into the 100s.
Then there will be multiple chips/packages; for both performing different tasks as possibly for resiliency/failover.
will be significantly less than the development vehicles, but still much higher than 72w.
This is production. In the cars already. Samsung in Texas manufacture the chip. Memory less power than processor. Connectivy in what sense consumes power? "100s" is a number you made up. Processor is built as two concurrently running halves for redundancy--leading to their "shootable" claim. This is all freely available info released months ago. I cannot vouch for any of it other than to say: why would they start lying at this point in their history.
Processor = ~72W
8GB GDDR5 (for example) = ~20-25W. Given the size of the datasets and models the systems could be running multiple of this, so 20-25W or 40-50W or maybe even 80-100W just in RAM (or less RAM per server but more servers)
Solid state storage = ~2W per TB active power consumption. Given AVs are predicted to generate >10TB/day (but not store all of it) there would likely be muliple TBs for the current, working dataset, so say another 10W.
Networking/connectivity = ~10W - this would be internal connectivity - sharing data (100s Mb/s up to Gb/s) between systems in the car (such as acqusition of images from camera to processor or do another idenitical processor running in lockstep as part of a resilient system or to another system that performs a different function in the workflow) and also perdiocally syncing/sharing data with cloud/backend systems for model/training updates, analytics and accountability likely over a 5G antenna.
So that is somewhere between ~100-200W for a single computing unit (processor + memory/storge + connectivity)
Also, that is only one implementation. Just because one company claims a implementation with X power consumption doesn't mean there are other implementations with different features/performance that consume more.
My point was, in general, in the context of different technology providers, there will be popular, non-niche production grade systems on the market where the power requirements can and will be greater than that figure.
A single processor built as 2 concurrently running halves in not a resilient system - it might given 2 independantly calculated results for any given task to help detmine probabilities/weights when all components are functioning correctly, but there are single points of failure spacially (they are physically in the same place), electrically, thermally and logically.
A truely resilient system needs physically seprated components which do not so share any common infrastructure - therefore multiple computing units active in the vehicle (multiples of that 100-200W figure).
Why would a company in a capitalist market trying to sell a product in a rapidly expanding sector make any sort of claims about a product they are developing... Well...
I'm not saying their figures are wrong, just a little disingenuous as it does not paint the entire picture... and tech companies always "sell forward" that is part of demand generation - make the buying public think it's ubiquitous.