stopcrazypp
Well-Known Member
You are forgetting the integrated GPU in Parker with 256 Cuda cores. That's why Parker is quoted at 1.5 TFLOPS (FP16).
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
lunitiks
Cool James & Black Teacher
EarlyAdopter
Active Member
On the topic of speculation of enough computing power and/or redundancy in what we see here in this module for FSD - do we know there isn't a second module elsewhere in the car?
lunitiks
Cool James & Black Teacher
Only GP100 (Tesla P100) and the Tegra X1 and X2 SoCs supports FP16 at double speed.
GP106 offers INT8 at quadruple speed. So 4,5 TFLOPs FP32 = 18 TFLOPs INT8.
You dont't want to train the neural networks in the car, that you do on a supercomputer somewhere, what you need in the car is fast inference for the network, and for that the INT8 capabilities of the GP106 is good.
Quadro-version of GP106 is not clocked as high as the gaming-version so 3.0 TFLOPS FP32 and 12.0 TFLOPS INT8.
Still has 1024 CUDA cores, from the pictures posted of the GPU it looks like it uses 4 GB GDDR5 memory on a 128-bit memory bus.
I thought so too initially, but @haakonks has explained about the performance above. Also the GP102 requires a lot more power and cooling.
Reference Drive PX2 boards from Nvidia has 2x GP106 boards (similar to Quadro P2000, but with less memory 128-bit bus and 4GB memory). If you need more memory and compute cores, you can change the GPUs out with 2x under clocked GP104 boards (similar to the GPU found in Tesla P4). Both have a TDP at 75W.
A single GP102 have 250W TDP.
A single GP102 have 250W TDP.
I disagree. Nvidia DLTOPS is referring to training and not inference. The entire drive px2 unveiling conference focused on it being a one stop shop for training and running models live. They even compared how long it took them to train 10 million images drive px2 versus a regular gpu/cpu.
Their alexnet test where drive px2 came out 6 times faster than titan x(2015) was based on training not inference.
Titan x (2016) for example has 44 int 8 inference and no deep learning library supports int 8 at the moment.
The 24 DLTOPS of drive px2 is training.
EDIT: but since we never got a fp16 number. We will never really know for sure. Nvidia has a problem of creating new labels and comparisons like the 150 macbook pro = drive px 2
Last edited:
JeffK
Well-Known Member
They may have said that but they didn't mean it. Here's a quote from the Nvidia website:
People don't normally train at 8 bit ints. Google's former TPU for instance was only for inference. The new one supports FP16 so they are finally about to do training on it.
It's possible they could augment the network or train a smaller network based on local patterns, but the major networks are going to have been trained outside the car.
Just think about the need for the Service center capacity to swap out all the boards. By the end of this year there will be more than 100 000 AP 2 cars in the wild..At the same time fixing the teething pains in the Model 3..
lunitiks
Cool James & Black Teacher
Btw Teslas FRT on this procedure is one hour (remove and replace AP2ECU, incl reset). So 100.000+ man-hours?
Btw Teslas FRT on this procedure is one hour (remove and replace AP2ECU, incl reset). So 100.000+ man-hours?
But keep in mind it would only apply to cars that purchased FSD, the retrofits wouldn't need to happen at the same time and they could likely be done in the parking lot with the customer present. It seems Tesla made the unit easily accessible for just this reason.
I see it being no worse than a minor recall really.
Is everyone in agreement that the hardware as it sits likely is not going to be capable of full level 5 autonomy?
Yes I agree.
By the way, I'm posting more pictures tonight. I'm going to pull out the daughterboard, clean off the thermal paste, and post pics of the GPU and memory ICs. Also, I'll remove the CPU heat sink and take pics of it.
Check back in then...
By the way, I'm posting more pictures tonight. I'm going to pull out the daughterboard, clean off the thermal paste, and post pics of the GPU and memory ICs. Also, I'll remove the CPU heat sink and take pics of it.
Check back in then...
lunitiks
Cool James & Black Teacher
You my friend is nothing less than a super hero.
Request: Plz capture a close up of the Ublox device
Btw Teslas FRT on this procedure is one hour (remove and replace AP2ECU, incl reset). So 100.000+ man-hours?
Service tech's are on salary anyways... and the take rate for FSD is probably pretty low up until now. And again, the FSD option likely already builds this into the price. But they should have a handle on this *before* massive Model 3 volumes.
@techmaven thanks for your post. I think Musk is pretty clear that Tesla will do whatever is required to ensure that anyone purchasing the FSDC option now will have their car upgraded in the future with whatever software AND hardware is required to enable that feature when it starts to roll out (and of course it will be incremental, not happen all at once).
@Matias as others have pointed out, it is very likely that only a small fraction of all Teslas sold include the FSDC option. Most buyers probably don't select that option (which should not be surprising, since the functionality does not exist yet and the initial implementation is likely over a year away, some say "years away", and even some Tesla fans say "never going to happen"). So when the time comes for FSDC to start to roll out, only a fraction of all Teslas sold since October 2016 will need upgrades.
Similar threads
