How many Flops is one Tops?

VIM3 is 5 TOPS, but other SBC boards are measured in FLOPS. Does anybody knows how many FLOPS is one TOPS.

1 TOPS = 1 Tensor operations per second

by the way TOPS is diffrent from FLOPS
here TOPS is referring to the NPU, FLOPS is used for the raw cpu, gpu processing power

Yeah, I know. But I want to be able to relate them.,

TOPS is in integers and FLOPS is in floating point numbers. So It’s a bit hard.

No, TOPS is Tensor operations per second not numerical integer operations
it is a scale of AI and deep learning frameworks and at what speed it operates

1 Like

@DROBNJAK On NPU, floating-point operation is rarely used . Almost all data will be converted to integer type for operation . So in NPU, TOPS cannot be equivalent to FLOPS by simple conversion . This is related to the array arrangement of NPU .

3 Likes

Do you mean 2D matrix when you say “array”?

@DROBNJAK
It basically depends on the GPU architecture. In pascal architecture 1TFLOP = 4TOPS.

What Frank was trying to say was NPUs typicaally contain what are called MAC unit (Multiply accumulate units) which is what used for AI interference…

If you want a reference this is has more performance than Google coral and Intel Neural compute stick 2 both have 4TOPS peak performance, Google coral has matched a desktop GTX 1080 in some workloads when connected to a desktop processor(in particular a i7 5gen processor). but in vim3 I have seen it match 1060 like performance

1 Like

I read this article quite long ago, very good article, really illustrates the scale of the processing capabillities across so many devices

OK, so by that comparison NVIDIA Jetson Nano is doing about 2 TOPS, which is 2.5 times slower than VIM3. And Jetson Nano and VIM3 are approximately in the same bracket, although VIM3 has Wi-Fi ( which interferes with USB 3.0 ), m.2 socket, 64GB of eMMC and two camera ports built in.

Yeah, hardware to hardware VIM3 beats Jatson Nano. Only thing left is software stack. Does VIM3 support CUDA or OpenGL and Visual Studio Code?

1TFLOP = 4TOPS in a on paper performance conversion. In practice since it’s a GPU and not an NPU it has over head and stuff that will reduce the performance…

If you want ease of use in AI related application I would suggest you go with Jetson Nano…

CUDA is only for Nvidia GPU, so naturally VIM3 doesn’t support cuda…

VIM3 supports OpenGL.

1 Like

1 TOPS = 1 Trillions Operations per second, 1 TFLOPS = 1 trillion floating point operations per second

Ahh yes, sorry for the Mishap, I got confused with Tensor Operations Per Second,

If 1 TOPS is 1 Tera operations per Second then the VIM3 has a NPU comparable to that of Apple’s A12 bionic

So NPUs are faster than GPUs for purely AI calculations, like Computer Vision?

Yes they are specifically designed circuits, that enhance the performance of AI

But the GPU is like a factory of 1000 mini cpus that can do more than just AI vision, and graphics output, it can accelerate almost any task 100x fold

yes they are… even though GPUs are superior, since it’s not designed for AI it has a lot of overhead making it less efficient in AI workloads.

That said now a days GPUs are also optimized for AI workloads like the one in new Nvidia Xavier NX SBCs

1 Like

Nvidia Xavier NX SBC is really force to recon with. I am currently hesitating between Xavier NX and VIM3. Xavier NX is not that much more expensive, but it is x4 times more powerful ( 21 TOPS vs 5 TOPS ) and it obviously doesn’t have Wi-Fi / USB 3.0 problem. But VIM3 is much more compact and probably can work with less power.

But from what I read on Nvidia’s forum people are having problems installing the OS. Pathetic.

NX is a new platform it will take 6-8 months to iron out the problems…

NX is way ahead of VIM3, It has more MIPI lanes, much more powerful decoder and encoder, I mean it can encode/decode 14 1080p Streams !!! It supports 6 cameras in MIPI lanes and camera support is also better.

But the cost is also 4X the cost of VIM3 (including base board) and also consuming 3X the power…

It’s not really x4 more expensive. With NX you get heat sink and fan included, while with VIM3 they are extras, plus you get more USB ports, 8 GB RAM etc. VIM3’s cameras cost $75 each, while with NX you can get cam for $20-30. As well, NX has tons more of learning materials.

But VIM3 has a smaller form factor, which is really good.