First, Google touts $150 AI dev kit. Now, Nvidia's peddling a $99 Nano for GPU ML tinkerers. Do we hear $50? $50? It's that time of year again. Jensen Huang, Nvidia’s CEO, paraded a range of goodies for GPU nerds, including a forthcoming credit-card sized AI computer board and software updates for its CUDA platform, during his company's annual GPU Technology Conference (GTC) in Silicon Valley on Monday. There is nothing that’s too eyebrow …
> It peaks at 472GFLOPS, we're told, and consumes about 5W to 10W.
Given the size of the heatsink on that puppy I'd say it's expecting to suck much more than 10 Watts.
And as with all these sorts of product, the success or failure will depend completely on the level of support (both from the manufacturer and user community) that it gets. The documentation, development tools, examples and ease of use are crucial. We saw this with the RPi: there are much better boards out there - cheaper ones, too. But the RPi dominates because of its community - not because of its technical capabilities.
Very much, Re: community support.
The Pi is actually a bit rubbish for what most people tend to use it for (which isn't what it was intended for, so not holding that against it!), but the community around it goes an order of magnitude beyond making up for it. The BeagleBone and Odroid communities do similar for those platforms. .... Then you have the vast ocean of technically better and often also cheaper boards that, support wise, well... best of luck!
With that performance and price, I see a potentially promising future as a CNC controller, with decent look ahead and route planning. Travelling salesman like path optimisation would be nice to implement in hardware, wonder how it would cope with that. Any word on numbers and capabilities of the GPIO pins?
Traditionally, yes. Travelling salesmen type paths are slow on traditional hardware, so would be interesting to see if the cuda cores could speed things up.
But in terms of motion planning (for acceleration profiles) is often done in the controller hardware, and also requires the hardware to look ahead at the next moves and account for them. Traditional DIY controllers are limited by slow processors and limited ram (few hundred kB), which limits the number of moves held in memory while still allowing head room for the calculations.
work quite happily with a 32 bit or smaller MCU, with an FPU if they're lucky, why would you need 400+GFLOPS for one? Especially why would you use a system without a predictable cycle time? There's a good reason that MCUs are slower than conventional CPUs, because that way you can guarantee how long each chunk of code takes to run.
Predictable cycle time is only relevant for actual step generation. Most commercial systems I've seen (including our newish Hurco machines we use at work) have a hybrid system with a CPU based system generating the acceleration profiles and a real time motion controller generating the control signals as queued by the CPU.
The motion control is a lot smoother and more finessed then our old Hurcos, and a lot better than my home machine running a smoothie board. We get better cycle time and surface finish (less chatter) out of our new machines than the old ones, even though they are using the same tooling.
Damn, that Jetson Nano looks like the wish-list spec for a Pi 4.
Linux4Tegra looks like a custom-built version of Ubuntu 14.
https://developer.nvidia.com/embedded/linux-tegra
A cluster of those would be interesting.
Like this:
https://www.techworm.net/2018/03/learn-build-supercomputer-raspberry-pi-3-cluster.html
Safety force field:
Awesome! A force field. Cool.
Nvidia’s Drive boffins have come up with a set of policies and mathematics, dubbed the Safety Force Field, that can be used by autonomous vehicles to, hopefully, avoid crashes and injuries.
Oh wait, so it's not an actual force field? I was all excited there for a sec. Bummer.
I've been looking around at a few things. The ARM3 have the fast Neon code but even running 4 cores doesnt achieve the power of the GPU which is 24GFlops , actually probably 12 in real life - which at a £5 for a zero compares well with the Google $150 USB which offers 16Gflops doing (pointless) 64 bit calcs.
The ARM custom NN design claims to be Tflops/W but may be a lot faster if they implement 8 bit FP - IBM recently showed this worked pretty well for most NN stuff.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
