AMD Radeon RX Vega Plug-and-Play in 10.13

[theitsage]

The eGPU vendor, Mantiz was the first to report RX Vega compatibility in macOS High Sierra on August 9th. I finally got my RX Vega 56 today after multiple delays from Best Buy. It was plug-and-play in 10.13b.

The biggest news is the link speed, PCIe 3.0 at 8 GT/s [false alert - OpenCL OpenWave shows PCIe 1.1 speed]. Perhaps this was part of the firmware update recently. Anyone knows how this is possible on a machine that's supposed to have PCIe 2.0?

I'm currently powering the card with a pair of mini 6-pin to 8-pin PCIe power cables. No other power modifications to this system. OpenGL apps crashes similar to what the RX 480 was doing last fall. I ran some quick benchmarks and it's slightly faster than the R9 Fury.

The card was identified as R9 XXX in About This Mac. I edited the controller file in AMD1000Controller.kext to show the correct name. IMO if you can find one of these for $399, it's worth it. Otherwise wait until the end of the year for the iMac Pro to launch so that's the drivers are more optimized.

 

[h9826790]

Thanks for the info. Do you mind to run OpenCL OceanWave to confirm the actual bandwidth transfer speed?

Anyway, the score looks very promising to me. Almost 20000 points on Luxmark with a single GPU that can easily power by the mini 6pins only. I really wonder how it perform in BruceX.

 

[theitsage]

Looks like false alert then?

 

[h9826790]

Looks like false alert then?

View attachment 716298

Agree, most likely system info can't correctly reflect the actual negotiation speed. This is the PCIe 1.1 speed.

Anyway, even with PCIe 1.1 x16, the performance is still outstanding (at least true on those benchmarks).

Also, do you have any monitoring software installed? Is it possible to post the power draw during Luxmark test?

 

[MisterAndrew]

Sweet! I sold my GTX 1080 today and have a Vega 56 coming from Newegg. I'm not sure if it matters which AMD board partner it comes from, but I selected Sapphire since that's the brand Apple seems to work with.

 

[theitsage]

Agree, most likely system info can't correctly reflect the actual negotiation speed. This is the PCIe 1.1 speed.

Anyway, even with PCIe 1.1 x16, the performance is still outstanding (at least true on those benchmarks).

Also, do you have any monitoring software installed? Is it possible to post the power draw during Luxmark test?

This is a test drive and I don't have many apps installed. Which monitoring utility do you recommend?

 

[h9826790]

This is a test drive and I don't have many apps installed. Which monitoring utility do you recommend?

You may utilise the iStat menus 14 days free trial.

HWSensors can also do the job, but the identification usually not matching the real source, may need some time to figure out which is which.

 

[theitsage]

Thanks for the info. Do you mind to run OpenCL OceanWave to confirm the actual bandwidth transfer speed?

Anyway, the score looks very promising to me. Almost 20000 points on Luxmark with a single GPU that can easily power by the mini 6pins only. I really wonder how it perform in BruceX.

I did five runs in BruceX just now. The average was right around 16s. Here's iStat Menus while Luxmark was running. This seems very manageable for the stock arrangement.

 

[h9826790]

I did five runs in BruceX just now. The average was right around 16s. Here's iStat Menus while Luxmark was running. This seems very manageable for the stock arrangement.

View attachment 716305
View attachment 716306

Right at my expectation! (dual HD 7950 score about 22000, and finish BruceX in 15s)

Sure this is the next "best" GPU for cMP (best balance in all price, power draw, driver support, performance......)

 

[joebclash]

Any news of a low cost vega card? Something under $150?

 

[Draeconis]

Any news of a low cost vega card? Something under $150?

RX Vega Nano will be released soon, but no date confirmed. Based on price of Vega 56, I'd say UK RRP of around £300/350 or so.

The eGPU vendor, Mantiz was the first to report RX Vega compatibility in macOS High Sierra in August 9th. I finally got my RX Vega 56 today after multiple delays from Best Buy. It was plug-and-play in 10.13b.

The biggest news is the link speed, PCIe 3.0 at 8 GT/s. Perhaps this was part of the firmware update recently. Anyone knows how this is possible on a machine that's supposed to have PCIe 2.0?

I'm currently powering the card with a pair of mini 6-pin to 8-pin PCIe power cables. No other power modifications to this system. OpenGL apps crashes similar to what the RX 480 was doing last fall. I ran some quick benchmarks and it's slightly faster than the R9 Fury.

The card was identified as R9 XXX in About This Mac. I edited the controller file in AMD1000Controller.kext to show the correct name. IMO if you can find one of these for $399, it's worth it. Otherwise wait until the end of the year for the iMac Pro to launch so that's the drivers are more optimized.

View attachment 716290
View attachment 716295 View attachment 716286
View attachment 716287 View attachment 716288

This is really interesting, but with a TDP of 210w, this card could pull up to 150w or more from either of those 8-pin connectors. Hopefully your board spreads the load evenly and draws the rated 75w from the PCIe link to reduce power draw from your mini 6-pin connectors.

 

[throAU]

Any news of a low cost vega card? Something under $150?

You're looking for an RX460/RX470. Not Vega in that price range.

 

[Kris Kelvin]

Thanks, theitsage!
If the values reported in iStat are accurate (which I somewhat doubt):

RX Vega 56
LuxMark score: 20000
AUX A: 82 watts (above spec for ATX 6-pin power)
AUX B: 87 watts (above spec for ATX 6-pin power)
PCIe slot: 22 watts
Total power consumption: 191 watts

RX 580
LuxMark score: 14000
AUX A: ? (shown as 0, sensor defective; probably 55 watts)
AUX B: 55 watts
PCIe slot: 56 watts
Total power consumption: 165 watts (assumed)

What setting was the BIOS switch set to on the Vega?
Did you experience any system instability?

 

[h9826790]

Thanks, theitsage!
If the values reported in iStat are accurate (which I somewhat doubt):

RX Vega 56
LuxMark score: 20000
AUX A: 82 watts (above spec for ATX 6-pin power)
AUX B: 87 watts (above spec for ATX 6-pin power)
PCIe slot: 22 watts
Total power consumption: 191 watts

RX 580
LuxMark score: 14000
AUX A: 0 watts
AUX B: 55 watts
PCIe slot: 56 watts
Total power consumption: 111 watts

So with current macOS drivers, the Vega 56 is basically 45% faster in LuxMark while requiring 70% more power than a RX 580.

What setting was the BIOS switch set to on the Vega?
Did you experience any system instability?

Thanks for the info!

The Vega 56 numbers looks very correct to me.

Luxmark 3.1 usually won't draw as heavy as 3D gaming or the Unigine benchmark. So, for a 210W GPU, drawing 191 is right at the expected range.

For this kind of power draw. I expect the Vega 56 will draw about right at about 75W in FCPX. And make it the perfect card to do that job. 16s BruceX, single 8GB VRAM.


However, RX580 looks a bit strange.
1) did you connect only one mini 6pin to the 8pin? If yes, why? It's more appropriate to connect 2x mini 6pin -> single 8pin in general.

2) if all CU working properly in the RX580 (at stock voltage and clock speed), that's really very good efficiency. Even though the Polaris are more build for efficiency but not performance. This kind of performance / power ratio still looks promising to me.

I think we need to run Furmark on the RX580 to check if it's really drawing anything close the 185W (WARNING: I only recommend this test when the 8pin connected to BOTH mini 6pin)

 

[Kris Kelvin]

However, RX580 looks a bit strange.

I agree, it does look strange.

I'm using the following cable, which I assume is wired correctly: http://www.ebay.ch/itm/-/112410059104?. Additionally, I've also tested with a regular 2 x 6pin to 8pin adapter. It's exactly the same.

With FurMark 0.7.0 under macOS @2560x1440, I see a total power consumption of 165 watts, evenly split between AUX B and PCIe slot (no power draw on AUX A). That's above spec for both the AUX and PCIe slot.

 

[MisterAndrew]

This is really interesting, but with a TDP of 210w, this card could pull up to 150w or more from either of those 8-pin connectors. Hopefully your board spreads the load evenly and draws the rated 75w from the PCIe link to reduce power draw from your mini 6-pin connectors.

Why would the card pull that much over one connector only? Would it be best then to have a dual mini 6-pin to single 8 pin cable connected to a 8-pin to dual 8-pin y-splitter cable to evenly distribute the power?

 

[Draeconis]

Why would the card pull that much over one connector only? Would it be best then to have a dual mini 6-pin to single 8 pin cable connected to a 8-pin to dual 8-pin y-splitter cable to evenly distribute the power?

6-pin PCIe power connectors are for up to 75w. Technically they can go beyond this, but it's not recommended.
8-pin PCIe power connectors are for up to 150w.

Realistically, you'd expect the draw to be balanced from both 8-pin connectors, and if you assume that Vega 56 draws 75w from the PCIe slot on the logic board, that leaves 135w, which evenly split between two ports is only 67.5w.

But there's no guarantees for any of this is how it'll work real-world.

 

[MisterAndrew]

6-pin PCIe power connectors are for up to 75w. Technically they can go beyond this, but it's not recommended.
8-pin PCIe power connectors are for up to 150w.

Realistically, you'd expect the draw to be balanced from both 8-pin connectors, and if you assume that Vega 56 draws 75w from the PCIe slot on the logic board, that leaves 135w, which evenly split between two ports is only 67.5w.

But there's no guarantees for any of this is how it'll work real-world.

The 6-pins on the cMP are not standard though. They have wires on all connections so are more like mini 8-pins. People on this forum have stated they reliably support up to 120w each.

 

[h9826790]

I agree, it does look strange.

I'm using the following cable, which I assume is wired correctly: http://www.ebay.ch/itm/-/112410059104?. Additionally, I've also tested with a regular 2 x 6pin to 8pin adapter. It's exactly the same.

With FurMark 0.7.0 under macOS @2560x1440, I see a total power consumption of 165 watts, evenly split between AUX B and PCIe slot (no power draw on AUX A). That's above spec for both the AUX and PCIe slot.

This is really strange.

Since the Vega can draw power from Aux A, and the reading is completely normal. So, it's safely to assume that iStat is reporting the correct power draw.

In this case, it seems the RX580 is not programmed correctly. Rather than utilise the 8 pin. It simply evenly distributed the power draw between 8 pin and slot. AFAIK, this is exactly what's happening on the RX480, and eventually kill some gaming PC. Because the initial version of RX480 only has a single 6pin. By considering both the 6 pin and slot are rated up to 75W, the logic may really simply always evenly distribute the power draw between them.

If the graphic card manufacture do absolutely nothing on the firmware to alter the power draw logic, but simply put a 8pin connector on the card, the last 2 pin may not be even connected properly. Then the result is explainable.

1) the card always evenly distributed the power draw, because it was designed for single 6pin card.

2) the 8 pin is not properly connected or programmed to draw power from the last 2pins

3) the dual mini 6 -> single 8 pin cable is seperateed inside. Aux A connected to 2 power supply pins inside the 8 pin. And Aux B is connected to another 2 power supply pins inside the 8pin. They are independent inside.

4) since 2 of the power supply pins inside the 8pin never been used properly. So, iStat report zero reading.

Of course, all the above are just my guess, but this is the only explanation I can think about which match all the results.

 

[Kris Kelvin]

In this case, it seems the RX580 is not programmed correctly.

I agree. On Windows, AMD's Crimson Edition driver contains a software fix to change the power distribution for RX 480 reference boards. The same fix is probably applied when using a RX 580. However, since there's no equivalent driver for macOS, this means that using a RX 580 may be dangerous to a Mac Pro. That being said, so far I haven't experienced any problems.

 

[fendersrule]

Ugh!

Back to wanting a Pascal.

 

[Draeconis]

The 6-pins on the cMP are not standard though. They have wires on all connections so are more like mini 8-pins. People on this forum have stated they reliably support up to 120w each.

If you want to run 120w down them, that's down to you. But when the traces on your logic board burn out, and Apple won't replace it because the model will soon become vintage.. you'll be a bit stuffed.

 

[MisterAndrew]

If you want to run 120w down them, that's down to you. But when the traces on your logic board burn out, and Apple won't replace it because the model will soon become vintage.. you'll be a bit stuffed.

That doesn't seem to make any sense. The TDP for the Vega 56 is only 10 watts higher than the HD 7950 Mac Edition.

 

[goMac]

Hoping we see a Mac Pro ROM someday. Is there a reason the 400 or 500 series never got a flash? Did the iMac ROMs move too far away from what's needed for a Mac Pro? Or did they build the ROMs into EFI?

 

[Squuiid]

On Windows, AMD's Crimson Edition driver contains a software fix to change the power distribution for RX 480 reference boards. The same fix is probably applied when using a RX 580. However, since there's no equivalent driver for macOS, this means that using a RX 580 may be dangerous to a Mac Pro.

Wow. That is bad news. It is beyond belief that AMD didn't address this issue in hardware with the release of the 580 rather than rely on yet another software fix.
I know the 480 and 580 are virtually the same card but this is still very sloppy.