AMD: Speculation, Rumors, and Discussion (Archive)

[Rodéric]

The goal of those new API is to get rid of the ineffiency of current API, and the mismatch of their representation vs what happens in real.

I'm not too keen on moving to new API, (but I'd pick Mantle/Vulkan over D3D12 any day) because GPU still aren't at the level I'm interested in when it comes to virtual memory management.
At least there are a lot fewer black boxes and magic going on ;p

 

[CSI PC]

Have you seen the power consumption. This is hugely disappointing because it seems like Polaris needs the same level of power to achieve the same performance as a R9 390X.

Although this is when it is pushed outside its optimal performance envelope (better than previous AMD gens), still here it seems Nvidia still has an advantage due to the inherent efficiency of their architecture-design.

A critical point in the video, to reach that 1500MHz they had to put the voltage up to 1.35V, and they mention this is not good from a durability or 24/7 perspective, importantly it is possible to see how quickly this wall was hit and just how narrow the operating window is because 1430MHz required 1.25V and they mentioned at 1.35V the GPU internal thermals were too much (limits of the die).
The performance window characteristic and the 1.35V absolute ceiling seems to be a character of these smaller FinFet nodes (both 14nm and 16nm), because Nvidia 1080 when pushed to the extremes was also reported to hit a wall around 1.35V and produced only small OC boost (relative to its own clocks) compared to around 1.25V.
Previous generation could target up to around 1.6V at the very extreme implementations, caveat being that just like the 1.35V clocking wall for 14nm/16nm is never designed for durability or long sustained use, along with deliberately breaking any sensor-protections-limitations put into the components/IC/GPU of the card.

So I think again we will see an OC advantage for Nvidia with the latest generations (may matter more with the very top custom AIB boards from a buyer's perspective or expectations of beating/matching certain higher tier models and cards), but Maxwell 2 I think will go down as the last 'golden' overclocker.
Cheers

 

[Anarchist4000]

This should have been the low power node, so having voltages over even 1V seems surprising. I'm still wondering if the memory voltage is limiting how low they can drive the chip. Other possibility seems the transistors are so sensitive they can't reliably control them. I don't suppose anyone knows the memory voltage difference between the 4/8GB cards? I'd think the 7Gbps chips are probably 100mV or more lower which likely shaves off some wattage.

 

[CarstenS]

Here you go.

 

[Razor1]

damn so AMD going to 4gb for reviews wouldn't have saved any real power then.

 

[CSI PC]

This should have been the low power node, so having voltages over even 1V seems surprising. I'm still wondering if the memory voltage is limiting how low they can drive the chip. Other possibility seems the transistors are so sensitive they can't reliably control them. I don't suppose anyone knows the memory voltage difference between the 4/8GB cards? I'd think the 7Gbps chips are probably 100mV or more lower which likely shaves off some wattage.

This is Samsung's LPP 2nd generation though, which is meant to compete with TSMCs 16nm replacement to high performance.
But that said, as I mention both seem to have limits of around 1.35V compared to previous gen, which should not be unexpected in that it would be more of a tighter performance window.
Anyway this lets us know the expected performance that can be seen between enthusiast and absolute ceiling.
Cheers

 

[Anarchist4000]

So 1.5V or 1.35V for 6Gbps, maybe 7Gbps(TBD). Ballparking 60% to drive them, so in theory that's an extra 100mV of core voltage. Problem being the voltage floor for the core to properly interface with the memory. Overvolting, or being unable to adjust, the memory voltage, may have kept the core artificially high. That's a pretty significant cut, ~20% total power, if that's the case.

damn so AMD going to 4gb for reviews wouldn't have saved any real power then.

It may have depending on the exact problem. I doubt the memory power is the issue. Best case they would maybe shave double digit watts which doesn't really solve the problem.

This is Samsung's LPP 2nd generation though, which is meant to compete with TSMCs 16nm replacement to high performance.
But that said, as I mention both have limits of around 1.35V compared to previous gen, which should not be unexpected in that it would be more of a tighter performance window.
Cheers

The specifics of LPP are the one thing I'm not sure on. All I've seen in papers are first gen and the graphs stopped at 0.9V. Still I would have expected lower target voltages only because they would need to add the same architecture to APUs eventually. Would be interesting to see Polaris 11 just to see what makes it the power efficient version.

 

[A1xLLcqAgt0qc2RyMz0y]

AMD Radeon RX 480 Power Measurements Repeated And Clarified

http://www.tomshardware.com/reviews/amd-radeon-rx-480-power-measurements,4622.html

 

[CSI PC]

The specifics of LPP are the one thing I'm not sure on. All I've seen in papers are first gen and the graphs stopped at 0.9V. Still I would have expected lower target voltages only because they would need to add the same architecture to APUs eventually. Would be interesting to see Polaris 11 just to see what makes it the power efficient version.

Also just to re-emphasise, that 1.35V performance ceiling ignores all safety/specs/limitations just like the 1.65V on previous gen did as well, meaning they are seriously reducing the life of the chip-components-IC, but this is about finding absolute ceiling/wall compared to say 1.21V, and provides some figures in terms of scaling and of course extreme OC benchmark competitions.
As they mention you would not run at these voltages/clocks if looking for durability and failure-free operation.
Cheers

 

[CSI PC]

AMD Radeon RX 480 Power Measurements Repeated And Clarified

http://www.tomshardware.com/reviews/amd-radeon-rx-480-power-measurements,4622.html

Thanks for the link, glad they also mention that GPU-Z measurement interval is only 500ms, which is still not really enough for an accurate reading into characteristics and they calculated the missing data points, not much of an issue as it sort of lines up.
110W for the GPU core does sound about right because at 1.25V it consumes roughly 144W, obviously this is the core without any other considerations.

Cheers

 

[CarstenS]

GPU-z is at 0,1 seconds in more recent versions. Still, far cry from some serious equipment employed by the reviewers.

damn so AMD going to 4gb for reviews wouldn't have saved any real power then.

1.35 volt vs. 1.5 volt should be a few watts for all 8 memory chips combined, I guess. IF there was a realy 4-GB-card.

 

[mczak]

The high resolution image of the die, it's pretty clear they are TMUs.

Also, going by that overclocking video above, 1500mhz on the core won't be easily attainable at all, even with AIBs. 1.35 volts is too much even under water for 24/7 use.

If that die shot is any indication, the MC PHY would be around 30-35mm^2, which is quite a lot. Or in other words, not shrunk at all compared to Tonga (in fact maybe even slightly larger per partition, but I suppose that was to be expected).

 

[Malo]

Never before in 20 years of building gaming PC's and dealing with every generation of cards have I seen caveats like these or encountered these types of warnings, which haven't been provided by the manufacturer.

4. Current hardware should be able to handle this amount of current without taking any damage, as long as the motherboard’s slots are clean and not corroded. It’s also advisable to make sure that the graphics card sits precisely in its slot. This should always be the case, though, even with significantly lower amounts of power.

There shouldn’t be any problems unless cheap, dirty or outdated components are used, the card isn’t installed correctly, or the amount of power drawn is increased by overclocking the card.

Basically mirrors bunch of other sites as well for their findings. So do motherboards and power supplies have to be certified for AMDs new cards? Is there a compatibility list of motherboards or a list that are considered too cheap? Why is the consumer responsible for possible damaged components if the card isn't within spec? This doesn't make sense for someone like Toms to state that the card is pushing 23% above the limit but that's ok, as long as you have quality hardware. If this is intended design by AMD, do these warnings go on the box by every manufacturer?

 

[mczak]

1.35 volt vs. 1.5 volt should be a few watts for all 8 memory chips combined, I guess. IF there was a realy 4-GB-card.

It is difficult to get the numbers, but samsung once had some slides saying power consumption of 1.35V gddr5 to be 30% lower than with 1.5V (at the same speed). That could amount to maybe 10W or so. However, I don't think 1.35V 7 gbps gddr5 is possible yet. All you could probably do is lower the voltage from the seemingly slightly overvolted value of 1.55V back down to 1.5V (together with the slower speed, might still be good for a 5W saving maybe in total).

 

[CarstenS]

It is difficult to get the numbers, but samsung once had some slides saying power consumption of 1.35V gddr5 to be 30% lower than with 1.5V (at the same speed). That could amount to maybe 10W or so.

Interesting find, thanks.

However, I don't think 1.35V 7 gbps gddr5 is possible yet. All you could probably do is lower the voltage from the seemingly slightly overvolted value of 1.55V back down to 1.5V (together with the slower speed, might still be good for a 5W saving maybe in total).

Note that the "TBD" in the 4 GB/1.35v row is following a /. My interpretation would be, that 8 gbps at 1.35v is still TBD, while 7 gbps is available.

 

[seahawk]

Basically mirrors bunch of other sites as well for their findings. So do motherboards and power supplies have to be certified for AMDs new cards? Is there a compatibility list of motherboards or a list that are considered too cheap? Why is the consumer responsible for possible damaged components if the card isn't within spec? This doesn't make sense for someone like Toms to state that the card is pushing 23% above the limit but that's ok, as long as you have quality hardware. If this is intended design by AMD, do these warnings go on the box by every manufacturer?

Imho the big problem is that we are talking about a mainstream card and not something aimed at enthusiasts. The good price point means it will be attractive to owners of older PCs including some OEM PCs. It would not be uncommon to see this card for a PC still using Sandy Bridge or even Core2 CPUs and this could mean OEM mainboards and power supplies around 5-8 years old.

 

[xEx]

Imho the big problem is that we are talking about a mainstream card and not something aimed at enthusiasts. The good price point means it will be attractive to owners of older PCs including some OEM PCs. It would not be uncommon to see this card for a PC still using Sandy Bridge or even Core2 CPUs and this could mean OEM mainboards and power supplies around 5-8 years old.

Ironically yes.

 

[Anarchist4000]

Imho the big problem is that we are talking about a mainstream card and not something aimed at enthusiasts. The good price point means it will be attractive to owners of older PCs including some OEM PCs. It would not be uncommon to see this card for a PC still using Sandy Bridge or even Core2 CPUs and this could mean OEM mainboards and power supplies around 5-8 years old.

So someone needs to take a cheap PC that doesn't meet specs and throw another product that may not meet specs into it at which point it's still likely to fail the same way it would under standard operation. Any failure case will be some guy using a $20 power supply whose mobo must have failed while he overclocked his 480. The only people pushing this are a bunch of alarmists trying to score marketing points or clickbait.

 

[CSI PC]

GPU-z is at 0,1 seconds in more recent versions. Still, far cry from some serious equipment employed by the reviewers.

Ah thanks Carsten,
good to know.
Cheers

 

[silent_guy]

So someone needs to take a cheap PC that doesn't meet specs and throw another product that may not meet specs into it at which point it's still likely to fail the same way it would under standard operation.

You can't prove the non-existance of something with one sample. Especially when it could be a failure that could only show up after a while.

Any failure case will be some guy using a $20 power supply whose mobo must have failed while he overclocked his 480. The only people pushing this are a bunch of alarmists trying to score marketing points or clickbait.

Maybe. But who enabled them?