What does 24GFLOPs GPU speed mean?

I found this definition. As an example, apparently the Nvidia GeForce 8400GS (2007 era entry-level desktop computing GPU) had approx 40GFLOP capability.

Many GPUs have hundreds of processors on the chip apparently. (I'm no expert in this area).

Per Wikipedia a GFLOP is a Giga(Billion) Floating-point Operations Per Second. That can be calculated by

http://en.wikipedia.org/wiki/FLOPS

All info provided by Boardcom about the BCM2935 which uses the VideoCore 4 GPU.

http://www.broadcom.com/products/BCM2835

I the link below states that the default GPU is set at 250MHz

http://elinux.org/RPiconfig

That is about all I was able to dig up. I will move this thread over to the Raspberry Pi Group to see if anyone there can provide furnther input.

Thanks,

Jason G.

Hi Jason,

   I really did intend this as a question for Element14 to answer via Feedback and Support,

asking what does Element14 mean when they claim 24GFLOPS performance,

rather than a question for discussion between RPi users as to what we think

Element14 is claiming.

   I don't see anything at the Broadcom page you mentioned that claims 24GFLOPS.

   User shabaz cites a definition requiring 64-bit precision, which I agree is standard,

but I don't think I've seen any claims that the GPU can even do 64-bit floating point.

   I presume the claim is that a single RPi does 24GFLOPS, so I'm not sure why you

selected Wikipedia's long formula that includes sockets/node and nodes/chassis.

   There is a claim here:

http://www.element14.com/community/groups/raspberry-pi/blog/2012/02/29/raspberry-pi-gpu-performance-dominates-the-iphone-4s-and-tegra-2-based-smartphones

that the RPi VideoCore IV is "dual core", but I think that's a mistake.

It's been a long time since I was active in supercomputing, but here's my understanding.  AFAIK, there's no requirement that FLOPS have to be double-precision.  There's no reason for a GPU to do double precision, so I can't imagine why Broadcom or anyone else would do so.

Generally, GFLOPS is a marketing number so the specs are as optimistic as possible.  It's generally a "peak" number, and according to one wag it's "a guarantee from the manufacturer that you won't go faster than this".  So they generally just multiply FLOPS times the clock rate, and say "sure, you could go this fast if you could get the operands to the arithmetic hardware".  That's usually very hard to do in general, so it's only with highly structured computations like inner products that you can actually get close to the peak.  If the problem is embarassingly parallel enough and you do some hand-tuning, you can get pretty close.

Another spec game: the main processing element of a floating-point engine is usually multiply-accumulate, since it's easy to do both at the same time.  These are counted as two FLOPs, so the GFLOPS rating immediately doubles.

As far as benchmarks are concerned, I note that chez RasPi a certain moderator often claims that VideoCore 4 is higher performance than its competitors, but I've never seen him quote an independent study.  If anyone has seen any, let us know.

The core has a 3D hardware acccelerator which consists of a bundle of procesors each capable of running SIMD (Single intruction mulitple data) 32-bit floating point calculations. If you multiply the processors, the data words and the clock you get the 24Gflops/sec.

Post edit: forgot that this is alongside the GPU itself which has a vector floating point processor which does 32, 64 or 80 bit precision floating point operations.

Which again is along side two general purpose (integer) cores. All of that is working with special H264 hardware accelrated video encode/decode HW and a plethora of other special multi-media cores.

Gert,

  Thanks for the reply.

We've been told previously that the GPU has a vector unit, but that

it is integer only, so it wouldn't be possible to develop floating-point

libraries, such as BLAS.  James wrote:

  "BLAS has lots of floating point, and the vector core is integer only unfortunately.

  I'll take a look at the interfaces posted though - should save a bit of specing time."

http://www.raspberrypi.org/phpBB3/viewtopic.php?p=14633

So is there a vector floating-point unit in addition to a vector

integer unit?  And if so, would it be theoretically possible to

develop libraries like BLAS that could demonstrate something

close to 24GFLOPS?

I was part of the team that build the  3D engine so I know it has vector floating point.

There are a number of 'issues with that:

1/ You are unlikely to get access to the details how to program it. It is very special and very valuable IP.

2/ The core is designed for graphic (pixel) calculatons so it has only very few of the IEEE exceptions build in.

I think it is well understood that the GPU IP is proprietary, but many people

have asked whether it was possible for Broadcom to develop an

interface that exposes the 24GFLOPS of performance, but the answer

has always been that it could perhaps be done for integers but not for floats.

So this is an interesting relevation if in fact a floating-point array library

could in theory be developed, even without IEEE exceptions.