AMD RX 480

[VictordeHolland]

Originally Posted by James Heinrich

They apparently released a statement today saying:So basically the problem is there and they'll make sure future versions of the driver limit performance to keep within power envelope.

Yes, I think this is the closest to an acknowledgement of the problem as we're going to get from AMD. In their quest for performance they (most likely knowingly) took the gamble, but it backfired. They'll have to get rid of 10-20W, I can't see them doing this without impacting performance.

I hope custom cards come with 2x PEG-6 or 1x PEG-8 connector to accommodate overclocking.

[bgbeuning]

Have a VisionTek rx480 8GB GPU card.

mfakto-0.14-win would not run on it.
It said to change VECTOR_SIZE from 4 to 2.
After getting the source and AMD APP SDK, it is now running.

System idle power 120 Watt (Kill-A-Watt)
Running mfakto system power 255 Watt for GPU using 135 W
(GPU-z says it is using 110 W.)

Its first assignment is M78891077 from 2^73 to 2^74 and
the estimated run time is 1h08m.

Please use your own judgement if the PCI-e power load is safe.
Mine is too new to tell.

[James Heinrich]

Originally Posted by bgbeuning

Its first assignment is M78891077 from 2^73 to 2^74 and
the estimated run time is 1h08m.

When it's done can you please send me the data?
http://mersenne.ca/mfaktc.php#benchmark

[bgbeuning]

Benchmark form sent.

Says 502 GHz-days / day

[James Heinrich]

Originally Posted by bgbeuning

Benchmark form sent.
Says 502 GHz-days / day

Much appreciated, thank you.

Hope you don't mind if I share the GPU-Z screenshot you sent, as it's quite informative.
Clock is very steady at 1191 MHz, which is above the stock 1120 but below the 1266 max boost.
Power is steady at 109W, just within the 110W envelope for the GPU only (the other 40W is for the RAM and rest of the board).

I would, however, like to see more benchmarks from other users. I'm not sure how much the VECTOR_SIZE has to do with it vs architectural differences, but throughput is ~15% lower than I had been predicting based on GFLOPS-GHzday ratios of previous-generation Radeons.

[LaurV]

Originally Posted by bgbeuning

After getting the source and AMD APP SDK, it is now running.

Yeah! Good job! You may inform Bdot about that, drop him a PM he is busy and does not read these threads. He is the mfakto/opencl specialist here, and may come with some optimizations and/or nice solutions.

[bgbeuning]

When I run 0.14 64-bit mfakto, usually the simple selftest (108 tests) will fail.
I have seen the number of failures between 0 and 5. When it had 0 failures
and started, it reporting 300 GHz-days / day. But this version passed the
long selftest (3092 tests) twice.

The version I compiled from source always passes the simple selftest,
but has failed the long selftest both times I ran it.

The card is in a machine I was willing to lose if it fried the motherboard
but the machine has been having other problems. I should try it in another
machine but that may not happen until next weekend.

I sent a message to the author and waiting to hear back.

I would be suspicious of the results misfit sent from this GPU last night

no factor for M78891077 from 2^73 to 2^74 [mfakto 0.15pre6-Win cl_barrett15_74_gs_2]
no factor for M78891277 from 2^73 to 2^74 [mfakto 0.15pre6-Win cl_barrett15_74_gs_2]
no factor for M78891493 from 2^73 to 2^74 [mfakto 0.15pre6-Win cl_barrett15_74_gs_4]
no factor for M78891797 from 2^73 to 2^74 [mfakto 0.15pre6-Win cl_barrett15_74_gs_4]
no factor for M78891877 from 2^73 to 2^74 [mfakto 0.15pre6-Win cl_barrett15_74_gs_4]
no factor for M78892039 from 2^73 to 2^74 [mfakto 0.15pre6-Win cl_barrett15_74_gs_4]
no factor for M78905969 from 2^72 to 2^73 [mfakto 0.15pre6-Win cl_barrett15_73_gs_4]
no factor for M78905969 from 2^73 to 2^74 [mfakto 0.15pre6-Win cl_barrett15_74_gs_4]
no factor for M78891811 from 2^73 to 2^74 [mfakto 0.15pre6-Win cl_barrett15_74_gs_4]

[0PolarBearsHere]

So apparently some of the retail 4GB units actually have 8GB of memory. The extra 4GB of which can be unlocked via a firmware modification.
http://www.kotaku.com.au/2016/07/peo...nto-8gb-cards/

[LaurV]

Why should they do that? (sell a double amount of silicon for the price of single)
The only explanation is that the card had 8G and it didn't pass the memory tests, therefore half of the chips (the unreliable part) were disabled "via a firmware modification", as you put it. This is common practice, it is called binning, it is cheaper to disable the defective part and sell it as a lower end, than desoldering the parts, or repairing it, which can be unsafe (other things get damaged) or costly (qualified repairing work is not cheap, I remember my former boss getting totally out of control when he was seeing engineers repairing/cleaning the mouse, at that time the mice had balls (hehe no link here, but you can google for photos ), i.e. they were not optical like now, and that ball was collection dust, hairs, grease, transpiration from your hand, etc, it had to be taken out periodically to be cleaned, the screws opened to remove the internal collected hairs, and this boss of mine always said that is cheaper to send the driver out to buy a new mouse, total cost 12 dollars, than an engineer paid with $20 per hour spending one hour to "repair" the mouse).

TL;DR, I won't enable the memory on those cards, except when using them for gaming. There may be a reason why it was disabled.

[henryzz]

Originally Posted by LaurV

Why should they do that? (sell a double amount of silicon for the price of single)
The only explanation is that the card had 8G and it didn't pass the memory tests, therefore half of the chips (the unreliable part) were disabled "via a firmware modification", as you put it. This is common practice, it is called binning, it is cheaper to disable the defective part and sell it as a lower end, than desoldering the parts, or repairing it, which can be unsafe (other things get damaged) or costly (qualified repairing work is not cheap, I remember my former boss getting totally out of control when he was seeing engineers repairing/cleaning the mouse, at that time the mice had balls (hehe no link here, but you can google for photos ), i.e. they were not optical like now, and that ball was collection dust, hairs, grease, transpiration from your hand, etc, it had to be taken out periodically to be cleaned, the screws opened to remove the internal collected hairs, and this boss of mine always said that is cheaper to send the driver out to buy a new mouse, total cost 12 dollars, than an engineer paid with $20 per hour spending one hour to "repair" the mouse).

TL;DR, I won't enable the memory on those cards, except when using them for gaming. There may be a reason why it was disabled.

It may just be a clock rate issue. It also might be possible to enable portions of the disabled memory. You do want to avoid a situation like the 970 with fast and slow memory.

[0PolarBearsHere]

Originally Posted by LaurV

The only explanation is that the card had 8G and it didn't pass the memory tests, therefore half of the chips (the unreliable part) were disabled "via a firmware modification", as you put it. This is common practice, it is called binning

That's what I'd assumed was happening. I too would want to make sure it was rock solid before I used it for anything important.