AMD RX 480 Fails PCI-E Specification

[SouthEastBlue]

Well if AMD have redirected the power draw of the card to pull more from the PCIIE 6 pin then well done. The card is still kept cool by a piece of crap cooler with the awful excuse for a heatsink underneath it. Then there's the fact you're extremely lucky to get even an 85mhz overclock out of it if at all.

These cards clock so bad I can't see even the Sapphire Nitro one making that much of a difference unless it allows for some serious over voltage.

 

[bru]

looks like the fix does exactly what it needed to do, so good job AMD for sorting it out nice and quickly.

Just how this issue wasn't picked up by their internal testing in the first place, is another matter entirely.

 

[D.P.]

Can you provide a source that Polaris has had it stripped? Anandtech states Fiji still has FP64, although only a reduced, amount; while still having full FP32 performance.

Anandetach explicitly state that FIji only has 1/16FP64 support, as does AMD )who should know, because they made the chip). That is why the profesional Fiji based card also only had 1/32 FP64 support.

http://www.anandtech.com/show/10209/amd-announces-firepro-s9300-x2

GCN 1.3is an evolution of GCN1.2 no where has AMD said they have increased the double precision rate. What AMD have said is there is half precision FP16 support though at 2:1.

 

[D.P.]

since it did pass pci-e spec the thread title is still wrong.

It passed pcie spec when set to use a lower power draw with lower clocks. The current card doesn't pass PCI specifications, even with the driver fix.

 

[Scougar]

How can you miss the comments in here about spec updates, and then still claim it doesn't meet the spec? Looks well within spec to me.

 

[N19h7m4r3]

Anandetach explicitly state that FIji only has 1/16FP64 support, as does AMD )who should know, because they made the chip). That is why the profesional Fiji based card also only had 1/32 FP64 support.

http://www.anandtech.com/show/10209/amd-announces-firepro-s9300-x2

GCN 1.3is an evolution of GCN1.2 no where has AMD said they have increased the double precision rate. What AMD have said is there is half precision FP16 support though at 2:1.

I never doubted that, and you did not link to me Polaris as I asked. Please link to me to where it's stated GCN4.0 having it completely stripped as you said.


I even stated in post you replied to that Anandtech said the FP64 compute of Fiji was reduced.
Edit:
I found one of the rare RX 480 reviews that shows varying compute performance.
Shame it does worse than Hawaii still, but not too much, and still beats the Titan X. If GP204 was 50% faster than TX, it would still be under the 480 in terms of Double Precisions for some tasks; while Pascal would pull ahead in a similar manner as in Folding@Home.

https://techaltar.com/amd-rx-480-gpu-review/2/

 

[Sargatanas2511]

It passed pcie spec when set to use a lower power draw with lower clocks. The current card doesn't pass PCI specifications, even with the driver fix.

Do you have proof that it passed spec at lower clocks and then was changed afterward?

 

[drunkenmaster]

Yes, but it is still technically out of "spec". I am not saying that I agree that it matters ( i don't really) but it is still technically outside PCIe specifcations.

It's not, PCI-SIG will tell you quite pointedly that they are guidelines and they were established in the mid 00s, they were basically saying make sure you have enough power circuitry to suit a 75W or a 150W input cable. It makes no difference if you draw 25, 75, 150W over a cable, you just need to make a safe card that can handle 150W incoming over the cable and it's fine. The RX480s VRMs/mosfets can deal with 150W over that cable with quite a bit of headroom on top.

 

[Catanonia]

Nice, buy a shiny new card to run stock and a few weeks later get it gimped to be safe to run stock.

Personally I would send the card back if I had of bought one.

 

[Sargatanas2511]

Nice, buy a shiny new card to run stock and a few weeks later get it gimped to be safe to run stock.

Personally I would send the card back if I had of bought one.

How is it gimped? Did you even read the thread? Sure it isn't ideal but using the word "gimped" seems horribly uninformed.

 

[N19h7m4r3]

Nice, buy a shiny new card to run stock and a few weeks later get it gimped to be safe to run stock.

Personally I would send the card back if I had of bought one.

Gimped so much the new driver gave an average 3.6% improved performance. If that's what you consider gimping, I'll take a gimping from AMD all the time.

Or if you run with the new Compatibility mode, no performance difference, but a reduction in power usage.

 

[LtMatt]

Gimped so much the new driver gave an average 3.6% improved performance. If that's what you consider gimping, I'll take a gimping from AMD all the time.

Or if you run with the new Compatibility mode, no performance difference, but a reduction in power usage.

I really have to bite my tounge sometimes with some of the stuff i read.

 

[agent00x]

Nice, buy a shiny new card to run stock and a few weeks later get it gimped to be safe to run stock.

Personally I would send the card back if I had of bought one.

at least we can actually buy the cards instead of being on a long waiting list for weeks on end without any idea of when the card will actually arrive.

and oh, the drivers IMPROVE the cards performance while decreasing power consumption, but you probably knew that already, you're just here to stir.

 

[humbug]

Has this been fixed?

 

[N19h7m4r3]

Has this been fixed?

Thoroughly, with some games getting a performance improvement on the drivers, and no performance difference if you choose to run the optional "Compatibility Mode" that further reduces powerdraw.

Note: PCIe 12 = PSU
Mainboard 12v = PEG/PCIe Slot

The new drivers to fix the 480 power draw are out.

http://support.amd.com/en-us/download
Here's Tom's tests on the new driver.

http://www.tomshardware.de/amd-radeon-rx-480-polaris-grafikkarte-treiber,testberichte-242146.html

 

[Seanspeed]

Gimped so much the new driver gave an average 3.6% improved performance. If that's what you consider gimping, I'll take a gimping from AMD all the time.

Or if you run with the new Compatibility mode, no performance difference, but a reduction in power usage.

Well, the driver performance improvements are a separate thing. If the new compatibility mode has indeed lowered the clock speeds, then yes, it has technically been 'gimped'. Basically, it would mean that if you didn't have the clock drop, you'd get even more performance from the driver update. This is assuming there actually is a clock drop(which I saw somebody post an example of, but not sure if that's representative of every 480).

I'm not sure that would be grounds for returning a card, but I also didn't think the 3.5GB fiasco was worth returning a card for either, so I'm probably less 'principled' than some.

 

[N19h7m4r3]

Well, the driver performance improvements are a separate thing. If the new compatibility mode has indeed lowered the clock speeds, then yes, it has technically been 'gimped'. Basically, it would mean that if you didn't have the clock drop, you'd get even more performance from the driver update. This is assuming there actually is a clock drop(which I saw somebody post an example of, but not sure if that's representative of every 480).

I'm not sure that would be grounds for returning a card, but I also didn't think the 3.5GB fiasco was worth returning a card for either, so I'm probably less 'principled' than some.

According to Guru3d when testing the new driver. The main issue is still throttling from that tiny little heatsink more than anything else. They noticed no performance difference at all.

They also state that the new "Compatibility Mode" is irrelevant, as the normal driver mode already reduces power load to be within spec.

Now in the driver you'll notice a new 'Compatibility mode' as well as a new 'Power Efficiency mode'. These modes are irrelevant in the sense that you do not have to change them.

I've also added a core clock recording of the card in a 3DMark FireStrike loop. Though there is less varying throttling, the card runs below 1200 MHz with that benchmark (which is normal behaviour). I verified and compared this with against Unigine Heaven (also included), which runs steady at 1240~1250 MHz in an Extreme mode. So the card does still reach its maximum boost frequency.

http://www.guru3d.com/news-story/radeon-rx-480-performance-benchmarks-with-16-7-1-driver.html

 

[humbug]

Thoroughly, with some games getting a performance improvement on the drivers, and no performance difference if you choose to run the optional "Compatibility Mode" that further reduces powerdraw.

Note: PCIe 12 = PSU
Mainboard 12v = PEG/PCIe Slot

Thanks

Right so 68 Watts on the PCIe, down from 82. Panic over, don't overclock it.

 

[N19h7m4r3]

Thanks

Right so 68 Watts on the PCIe, down from 82. Panic over, don't overclock it.

Some people have downvolted the GPU and memory even further already, and gotten even better performance from it.

I do wonder why the hell AMD overvolted the card so much on the reference design. Someone in the Owners thread is running it at just under 1V as well.

Installed 16.7.1

Slight boost in firestrike score:
(16.6.2) = 9613 (gfx=12547, phys=7039, comb=4360)
(16.7.1) = 9868 (gfx=12906, phys=7127, comb=4509)

I am undervolting core to 990mV and memory to 970mV and set core max to 1240 and left memory at 2000

 

[Griffildur]

A bit unrelated to the discussion but I felt I'd post this anyway.

One of the reasons I like this forum is that you get people from AMD, EVGA and OC providing info and actively helping out.

Major thumbs up to people like AMDMatt, Ben@EVGA, Gibbo@OC (who actually helped me overclock my shiny new card).

The forums can be a bit ... rough ... sometimes so don't let that get to you ...