Radeon Instinct MI25 16GB HBM2

[pio]

8 hours ago, Lifted said:

 

 

Thank you for confirming

I am debating between this and GTX 1080ti/RTX 3070

How much would you want for it?

 

I probably won't need the water block, I simply don't run custom loops in my PC, since I move components too much.

But that fan adapter, I would love to see that. Would be nice to see what you came up with.

Not trying to be rude, or even a moderator..., but best to take that conversation to PM at this point.    

As far as your debate, you already know the answer I'm sure.  RTX 3070 > GTX 1080Ti > Vega.  The cards will be priced accordingly I'm sure too.  Knowing damric though, I'm sure he'll hook you up with a fair price on that MI25.  On the other hand, you COULD say screw it, why not both?  For the price of the RTX 3070 you could have a 1080Ti and the MI25 I'll bet.  Both cards would be pretty comparable; personally I think it's kinda neat having a card from both camps that are similar.

[damric]

1 minute ago, pioneerisloud said:

Not trying to be rude, or even a moderator..., but best to take that conversation to PM at this point.    

As far as your debate, you already know the answer I'm sure.  RTX 3070 > GTX 1080Ti > Vega.  The cards will be priced accordingly I'm sure too.  Knowing damric though, I'm sure he'll hook you up with a fair price on that MI25.  On the other hand, you COULD say screw it, why not both?  For the price of the RTX 3070 you could have a 1080Ti and the MI25 I'll bet.  Both cards would be pretty comparable; personally I think it's kinda neat having a card from both camps that are similar.

I'm probably going to keep it a bit longer to play with the nime drivers after the team comp is ended

[pio]

Just now, damric said:

I'm probably going to keep it a bit longer to play with the nime drivers after the team comp is ended

Ahh, well there we go then.   

It is a fantastically neat card.

[Lifted]

4 hours ago, damric said:

I'm probably going to keep it a bit longer to play with the nime drivers after the team comp is ended

That settles it, I was leaning towards RTX 3070 or GTX 1080ti or wait and get 6950xt or something.
I want to make sure all parts play together first.

With E5-26XX v4 series of Xeons I don't want to go too strong on the GPU side, as I am limited by CPU, unless I bump up to 1440p, or 4k, then I can actually utilize those resources. Or I will just throw away GPU resources.

In fact with 9900k, I was already bottlenecked with RTX 3070 at 1080p, by about 20-30%

 

I am thinking MI25 is too weak. GTX 1080ti beats it, and costs about $150 used. 
HBM2 is good for machine learning from what I know, but for gaming as well, it lacks a lot of ease of compatibility.

 

Edited August 29, 2023 by Lifted

[damric]

10 minutes ago, Lifted said:

That settles it, I was leaning towards RTX 3070 or GTX 1080ti or wait and get 6950xt or something.
I want to make sure all parts play together first.

With E5-26XX v4 series of Xeons I don't want to go too strong on the GPU side, as I am limited by CPU, unless I bump up to 1440p, or 4k, then I can actually utilize those resources. Or I will just throw away GPU resources.

In fact with 9900k, I was already bottlenecked with RTX 3070 at 1080p, by about 20-30%

 

I am thinking MI25 is too weak. GTX 1080ti beats it, and costs about $150 used. 
HBM2 is good for machine learning from what I know, but for gaming as well, it lacks a lot of ease of compatibility.

 

I agree with your points. The MI25 is truly a labor of love project at this point. For performance at just about anything your money is better spent elsewhere now that the worst GPU crunch is over.

[LabRat]

Dangit, @pioneerisloud!


I don't even know how I got here... oh wait... yes, I do.

Edited August 31, 2023 by LabRat

[pio]

2 hours ago, LabRat said:

Dangit, @pioneerisloud!


I don't even know how I got here... oh wait... yes, I do.

  Your welcome! 

[LabRat]

Update!

After

"Mixed-CrossfireX Adventures"

and
finding a 'cheap' AIB/aftermarket cooler for a 65x65mm mount-pitch GPU

Both, went about as smoothly as I'd expect...

 

I decided to just re-paste and re-mod my MI25's cooling.

 

'stock' TIM - looks like it dried up some, and wasn't a very good application.

 

Clean n' pretty - Vega 10, and its two stacks of Samsung 8GiB HBM2

 

The Mod - Heatpipe > Copper | Allotropic Carbon > Heatpipe 

 

Used a sheet of Thermal Adhesive-backed Allotropic Carbon-coated Copper Foil

AKA: Space Magic (NVMe) Heatspreader

 

TIM is 4+ year old MX-4, that I re-homogenized using a drill and paper clip

(Same TIM is on my 5600. lol)

 

Previously, the card would crash if sustaining over 1400core-1100mv (190-210W sustained, 245-250W peak)

Now, I can sustain 1528core, 1100-1150mv (240-250W sustained, 280W+ peak).

'light loads' like Quake 2 Remaster (@ 240fps) don't even get to 45c now. The fan rarely if ever becomes audible.

Note: PPT loaded (and edited) from Vega FE 16GB Air .rom. There's still room for improvements.

 

Lastly (sorry for no pics @TM) - I re-oriented the fan.

Instead of the 120x38 axial bloweymatron being 'semi-flat' against the card (making a 3+ slot card)

I set the fan 'up on its end', and taped a 'plenum' around it (making a non-symmetrical* 5+ slot card.

 

*With the shorty 6500XT in my X570-gen4x4, everything 'fits'

The Vega's fan 'rests' on my HDD cage (doubling as a strain-relief support), and the shipping tape plenum angles to Vega's heatsink; juuust 'touching' the edge of the 6500XT.

[will edit-in pics later. "If it looks stupid, but works; it aint stupid" needs pics to properly convey)

 

 

 

 

 

Now. I have a question.

Why are some 16GiB MI25s using 2x8-pin power (like mine)

and

Some are using 1x6-pin+1x8-pin

????

If I wasn't concerned that a 6+8 Mi25 might clock worse, I'd buy another for matched CrossfireX/AMD MGPU

[Wow... was gonna post pics of an eBay listing I asked that question to. He pulled the pic that showed the power inputs. Guess he wants to 'move' those cards...]

Edit 1: Went ahead and sent the eBayer a 'liquidation-pricing' offer on 2.

Edited September 17, 2023 by LabRat
Update on end of post comments

[damric]

1 hour ago, LabRat said:

Update!

After

"Mixed-CrossfireX Adventures"

and
finding a 'cheap' AIB/aftermarket cooler for a 65x65mm mount-pitch GPU

Both, went about as smoothly as I'd expect...

 

I decided to just re-paste and re-mod my MI25's cooling.

 

'stock' TIM - looks like it dried up some, and wasn't a very good application.

 

Clean n' pretty - Vega 10, and its two stacks of Samsung 8GiB HBM2

 

The Mod - Heatpipe > Copper | Allotropic Carbon > Heatpipe 

 

Used a sheet of Thermal Adhesive-backed Allotropic Carbon-coated Copper Foil

AKA: Space Magic (NVMe) Heatspreader

 

TIM is 4+ year old MX-4, that I re-homogenized using a drill and paper clip

(Same TIM is on my 5600. lol)

 

Previously, the card would crash if sustaining over 1400core-1100mv (190-210W sustained, 245-250W peak)

Now, I can sustain 1528core, 1100-1150mv (240-250W sustained, 280W+ peak).

'light loads' like Quake 2 Remaster (@ 240fps) don't even get to 45c now. The fan rarely if ever becomes audible.

Note: PPT loaded (and edited) from Vega FE 16GB Air .rom. There's still room for improvements.

 

Lastly (sorry for no pics @TM) - I re-oriented the fan.

Instead of the 120x38 axial bloweymatron being 'semi-flat' against the card (making a 3+ slot card)

I set the fan 'up on its end', and taped a 'plenum' around it (making a non-symmetrical* 5+ slot card.

 

*With the shorty 6500XT in my X570-gen4x4, everything 'fits'

The Vega's fan 'rests' on my HDD cage (doubling as a strain-relief support), and the shipping tape plenum angles to Vega's heatsink; juuust 'touching' the edge of the 6500XT.

[will edit-in pics later. "If it looks stupid, but works; it aint stupid" needs pics to properly convey)

 

 

 

 

 

Now. I have a question.

Why are some 16GiB MI25s using 2x8-pin power (like mine)

and

Some are using 1x6-pin+1x8-pin

????

If I wasn't concerned that a 6+8 Mi25 might clock worse, I'd buy another for matched CrossfireX/AMD MGPU

[Wow... was gonna post pics of an eBay listing I asked that question to. He pulled the pic that showed the power inputs. Guess he wants to 'move' those cards...]

Edit 1: Went ahead and sent the eBayer a 'liquidation-pricing' offer on 2.

Looking good! I'm guilty of keeping and using very old thermal grease as well. I finally finished a big tube of the MX-4 2019 edition last month

 

That is weird with the 6+8 pins.

[damric]

Alrighty I finally got a chance to work on this again after having to put it on the shelf for a while during the Team Cup.

 

Looks like TRIXX allows higher clocks via higher power limits, but it's a bit sketchy on whether it will crash or not once I apply. If the bench can start, it runs flawlessly though.

 

 

[LabRat]

5 hours ago, damric said:

Alrighty I finally got a chance to work on this again after having to put it on the shelf for a while during the Team Cup.

 

Looks like TRIXX allows higher clocks via higher power limits, but it's a bit sketchy on whether it will crash or not once I apply. If the bench can start, it runs flawlessly though.

 

 

It's been my experience that the 'fall back' P-states, have to be customized too. AFAIK, these MI25s are the only Vega that *needs* tweaking beyond P6 and P7, for max Watt/Perf.

I mean, it makes sense. We're basically 're-binning' Vega 10 16GB into another product line

note: Samsung HBM2 seems to run 1+ghz no problem, even with a 1025-1050mv 'HBM voltage floor' set.

 

I played STALKER: Anomaly for *hours* yesterday, sustaining 1390-1528Mhz core*, at less than 1100mv.
220ish-260ishW sustained, 280ish peak.

If I push voltages to 1100mv or higher, I can get 1600+, but hotspot or HBM temps eventually cause instability.

(Also, the 2-less power phases trying to push 300+W might be related. I've seen 350W+ pulled from my card, before overheating)

 

*After using OverdriveNtool to 'find' stable clocks, I used RadeonWattman to create a profile for STALKER Anomaly that locks it in p-states 5-7
(some areas in the game don't 'load' the GPU, even though the clocks are needed for smooth framerates)

 

Any idea on how to edit the vBIOS's clocks/voltages? I'm very close to finding my card's "sweet spot", would like to make it 'permanent'.

 

 

PS: I've got a 6+8pin MI25, on the way; I'm very curious what the differences between the 2 MI25s are.

 

 

 

Edited September 28, 2023 by LabRat

[damric]

10 minutes ago, LabRat said:

 

It's been my experience that the 'fall back' P-states, have to be customized too. AFAIK, these MI25s are the only Vega that *needs* tweaking beyond P6 and P7, for max Watt/Perf.

I mean, it makes sense. We're basically 're-binning' Vega 10 16GB into another product line

note: Samsung HBM2 seems to run 1+ghz no problem, even with a 1025-1050mv 'HBM voltage floor' set.

 

I played STALKER: Anomaly for *hours* yesterday, sustaining 1390-1528Mhz core*, at less than 1100mv.
220ish-260ishW sustained, 280ish peak.

If I push voltages to 1100mv or higher, I can get 1600+, but hotspot or HBM temps eventually cause instability.

(Also, the 2-less power phases trying to push 300+W might be related. I've seen 350W+ pulled from my card, before overheating)

 

*After using OverdriveNtool to 'find' stable clocks, I used RadeonWattman to create a profile for STALKER Anomaly that locks it in p-states 5-7
(some areas in the game don't 'load' the GPU, even though the clocks are needed for smooth framerates)

 

Any idea on how to edit the vBIOS's clocks/voltages? I'm very close to finding my card's "sweet spot", would like to make it 'permanent'.

 

 

PS: I've got a 6+8pin MI25, on the way; I'm very curious what the differences between the 2 MI25s are.

 

 

 

I'm not sure about the voltages. I can't seem to read them. I tried those 3rd part drivers but failed to get them to work right so I went back to older crimson drivers.

 

I did get a time spy benched today while sustaining around 1670MHz. That was using overdriventool and trixx to push the power limit to +100%. At one point I had it at 1700MHz while testing but it was hit or miss on whether or not Time Spy loads would crash on start. If I can get it to start the bench, it's fine. Maybe I need to back the HBM off since I had it at 1GHz. I'll play around more tomorrow.

[damric]

I am finishing up my benchmarks on this card today, then I think I will list it for sale in the forums. You can get these for about $70 on ebay, so mine will be $50 and I will include the universal GPU block and 120mm fan adaptor.

 

This was one of the more challenging projects I have completed. First the stock cooler had to be modified since it is a fanless design only meant for low clock speeds. I feel like someone could make it even better by adding more cooling to the back side VRMs. Speaking of which, the VRMs are underpowered on this card compared to other Vega 56/64 and they run very hot, so at least a fan is needed to point at the back of the card at the socket area. Then the card needed a firmware flash to make the mini-display port usable. I chose the WX9100 BIOS, but there are other options depending on what exactly you want to do with it. Then I had to find the right drivers. There's multiple options there. Then lastly to overclock it I used OverdriveNtool and an old version of Sapphire Trixx.

 

I pulled a bunch of hardware gold cups, but one benchmark just wouldn't run without crash and that was 3Dmark Ice Storm EXTREME. I have no idea what's up with that particular bench, but I tried all kinds of things and I think it was something to do with low power states but I could not figure it out.

 

But that's it, I think I'm done with it. Probably still a good card for AI or descent if tuned for folding in the 100W range. I know some people try and game on this card since it's 16GB but I think there's easier options for that performance level.