Bitcoin Forum

I'm wondering about the power efficiency of PowerVR chips for mining. As an example, the SGX545 has an OpenCL compatible pipeline, and average power draw of .9 watts. I don't have data on comparative computing, but the pixel fillrate is 1/28 of a 5850 with less than 1/150 of the power consumption.

Thoughts?

do you mean mining on a mobile device?

well, on a mobile chipset. But I was thinking ordering stand alone boards with an embedded linux mining setup.

Ok, so I did some number crunching.

The Radeon Mobility 4670 gets 674% higher performance than the Intel GMA 500 (based on PowerVR 535) in Fear, which is a very shader intensive game.
http://www.notebookcheck.net/Intel-Graphics-Media-Accelerator-500-GMA-500.12614.0.html

According to Anand, the modern PowerVR chip used in the iPad 2 gets roughly 5.5x the performance of the GMA 500 chip in shader intensive tests.
http://www.anandtech.com/show/4216/apple-ipad-2-gpu-performance-explored-powervr-sgx543mp2-benchmarked/1

This means we can expect a modern PowerVr chip to 81% of the shader performance of a Mobility 4670. The M4670 is equivalent to a desktop 4650, which can manage 31 Mh/S.
https://en.bitcoin.it/wiki/Mining_hardware_comparison

We can therefore expect the PowerVR to manage 25 Mh/S at .9w. This gives us a Mh/J of 27.7, compared to 1.8 for a 5850.

This has potential.

And how do you plan on writing drivers to use this in mining? You need Cuda or OpenCL which the chip dosent have. Secondly, You cannot base shader performance and gamer performace as basis for hashing. Example of why

I have a Nissan truck. It climbs up a 15% incline over 10 feet in 2 seconds. I can run that distance and incline myself. Thus, I am a Nissan?

Doesn't look promising dude.

First off, yes it does. "Combining our many years of experience in the embedded, mobile and PC-based DirectX graphics worlds, POWERVR SGX 545 takes the possibilities of hand-held graphics to a new level by delivering a full DirectX 10.1 and OpenGL 3.x feature set as well as delivering GPU powered OpenCL heterogeneous parallel processing capabilities for the mobile and embedded markets. This makes POWERVR SGX545 a compelling solution for application processor SoC designers targeting the next generation of netbook and MID mobile products demanding exceptional graphics capabilities."

Second, it has linux driver, since Android is based on a Linux core.

Third, shader computations are done on the same pipelines which are used by OpenCL for computation. This is why unlocking shaders on the 6950 improves Bitcoin performance.

Cuda and OpenCL are just standards that allow you to access hardware through a standard set of libraries rather than writing custom code for each possible hardware/OS combination. Making a miner for other hardware is not impossible, but rather something that requires more work.

Have you tinkered with this any more? Something like a Pandaboard would make a great test platform.
http://www.pandaboard.org/content/platform

I've thought of getting one to control an FPGA miner and for other projects.

That does look promising. I looked into the technical side and saw no reason why such a project wouldn't be possible. I was more constrained by lack of a test platform.

I think I will get on of those pandaboards and see what kind of performance numbers I get.

According to this way of thinking, the Nvidias have a lot of potential too. Graphic performance =/= mining performance.

The only way you can know if the PowerVR are good for mining is trying them. Different graphic architectures have different mining performance. If you do Im interested in knowing how they do.

Color me impressed, I don't believe it will either give out that hash power or use that little power. How could they provide a chip that is more than ten times for efficient than high-end desktop versions, for a device needing less complexity than a desktop PC.

Thats exactly it. They are much less complex, which means smaller chips, which means less power consumption. Mining is a very simple operation, most of the stuff on a desktop chip (high speed RAM, ROP's, Texture Units) is extraneous and wastes power.

But I agree, hard numbers are needed. I will try and get a dev board so I can test it.

Since they are less complex, they should be oriented towards this specific process (hashing operations). Yes, I am aware mobile chips use much less power, especially in sleep mode.

According to the wikipedias, SGX545 has a 1 Gpixels/s fill rate for 2 x 200Mhz and 1W, to use such a metric that can be compared to desktop GPUs...
An HD6990 has 53 Gpixels/s rate for 2 x 830 MHz and 375W on the other hand, so 50 times fill rate, 4 times clock rate and 375 times the power usage.

So by idiot made-up math, I would say the SGX545 would be able to make HD6990 hash rate divided by 53 times 4, or 3.77 Mhash/s @ 1W? Really good, but not that extreme as previously suggested?

The issue is that pixel fillrate is related to the ROP's, which are not used in bitcoin mining.
http://www.gpureview.com/pixel-fill-rate-article-365.html

Thats why I did the complicated 3 step comparison to try and get shader performance number. I haven't been able to find any good direct comparison for shader performance between PowerVR and desktop chips.

This will be fun to try out, since I am going to get a Pandaboard soon anyway.

I hope mining for long periods of time doesn't burn up the OMAP4430, the PowerVR SGX540 is on die with the rest of the CPU. Additional cooling is probably in order, I know the OMAP in phones can get quite hot when running full speed.

http://focus.ti.com/general/docs/wtbu/wtbuproductcontent.tsp?templateId=6123&navigationId=12843&contentId=53243

Very nice. Please, let us know the results, Im very curious about it.

I'm not sure how soon it will be, but I'll post any results I get. I'm more interested in using the board to control FPGA test devices. I think the coming ASICs will make using any graphics chipsets pointless, save for hobbyists.

bookmarked