AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread

WTF?

987z3sml.jpg


http://www.anandtech.com/show/7457/the-radeon-r9-290x-review
 
The base clock is supposed to be 850 MHz, not 727 MHz
Some say they observed 727MHz in Furmark, at which point the fan speed setting got violated to keep it there. Anyway, it's just the exact value for one of the absolute limits, it doesn't change the general idea of the user accessible priority list for the Powertune algorithm.
 
After thinking a bit about the priority thing with Powertune, it would actually be nice to offer a setting in the control center for what takes precedence over what.
I think there may be some nice properties to the hierarchy and static temperature target.

Long-term reliability is easier to ensure when the chip and package are subjected to fewer thermal cycles.
Controlling temperature adds a constraint when modeling the behavior of an unknown thermal solution, which the controller algorithm would use to determine how close it can take things before exceeding the cooler's ability to absorb a power spike.

Limiting a control scheme with microseconds of latency by the mechanical limitations of a fan's ramp or the perceptual unpleasantness of spastic RPM modulation slows thermal and clock handling.

In terms of responsiveness, the other constraints enforce a hierarchy of very fast DVFS with microseconds, transient thermal behavior of the inteface that operates beyond a thermally significant period (some number of microseconds with a floor close to the controller, then increasing fractions of a second when it comes to the sink), then the constrained variation of the fan that might not fully transition for longer.

Guard-banding for the temp target would likely have to go back to what it was before if the clock and voltage scaling has to defer to the slower elements.
 
@3diletante:
One doesn't have to make it more complicated as it is. The number of thermal cycles wouldn't be changed at all by such a priority list. The R290X is specified to run at 1 GHz at 95°C and the default powertune behaviour ensures it reaches said 95°C under load. It can't get that much worse from the durability point of view. Setting a more aggressive fan profile wouldn't make things worse. I'm not against keeping the properties of the rather smooth regulation of the fan speed and such stuff.

===========================

Back to the actual behaviour of Powertune, according to a report in my home forum (3DC), Powertune actually behaves like some kind of a blend between my scenarios (v) and (vi), at least with reduced temp target and increased fan speed limit. It downclocks a bit before it reaches the fan limit, so it keeps the temp target with a combination of downclock and fan speed (with a bit more weight on fan speed apparently). It would be nice if the CCC would provide a switch to decouple the fan control from the temp threshold for downclocking. That way, one could keep the GPU at lower temperatures for "normal" loads and woud only approach 95°C for very demanding stuff (Furmark).
 
Last edited by a moderator:
It is sad, if true. :cry:
New 290X with 1:8 DP for $549 it is misunderstanding if 7970 with 1:4 DP for launch price $549 and 6970 with 1:4 DP for $369.

Why? Beyond the fact that double is clearly 2x single (only that is sort-of is not). Who cares about this in the desktop / gamer space? And if anybody says "BUT FLIGHTSIMS TOTALLY NEED QUAD-PRECISION FOR COORDINATES" I will cry.
 
It is sad, if true. :cry:
New 290X with 1:8 DP for $549 it is misunderstanding if 7970 with 1:4 DP for launch price $549 and 6970 with 1:4 DP for $369.
This is not a meaningful comparision. R9 290X offers 704 GFLOPS DP. That's more than HD 6970 and also (4-times) more than GTX 780 (=direct competitor). Compared to HD 7970, it's 75 %. That's not bad at all. With some of the architectural improvements real-world DP performance can be quite close to the original HD 7970 in some cases.
 
Why? Beyond the fact that double is clearly 2x single (only that is sort-of is not). Who cares about this in the desktop / gamer space? And if anybody says "BUT FLIGHTSIMS TOTALLY NEED QUAD-PRECISION FOR COORDINATES" I will cry.

But flight sims totally need quad-precision for coordinates!


But mostly, previous top Radeons (Cypress, Cayman, Tahiti) would let you build very powerful, very cheap clusters. They also made for nice, cheap GPGPU development platforms. I guess it was only a matter of time until AMD decided to protect its FirePro sales.
 
I think that when you're aiming for top gaming GPU with a ~30% area disadvantage, everything else goes out the window in order to achieve that.
 
Because it's probably still hardware limited to 1/4?

According to Toms' the hardware is capable of 1/2 DP

We've also come to learn that AMD changed the double-precision rate from 1/4 to 1/8 on the R9 290X, yielding a maximum .7 TFLOPS. The FirePro version of this configuration will support full-speed (1/2 rate) DP compute, giving professional users an incentive to spring for Hawaii's professional implementation.
 
No, not explicitly.

I've only seen the slides that point out pixel pipes, but the drawings showed everything hooking into a vague compute unit blob.

If what was described verbally as a pixel pipe matches what is now marketed as a shader engine, it indicates a more dramatic shift in the diagram than in the architecture.


@3diletante:
One doesn't have to make it more complicated as it is. The number of thermal cycles wouldn't be changed at all by such a priority list. The R290X is specified to run at 1 GHz at 95°C and the default powertune behaviour ensures it reaches said 95°C under load. It can't get that much worse from the durability point of view. Setting a more aggressive fan profile wouldn't make things worse. I'm not against keeping the properties of the rather smooth regulation of the fan speed and such stuff.

My speculation centers on how the design is hovering in a relatively narrow temperature band, with some potential thresholds that the slower to react power controller methods shy away from.
There is a nebulous upper temperature bound where other silicon chips would rather shut down the system before they damage something. I see CPUs trip something only a little higher than what the GPU runs at, so I admit I'm assuming that ceiling isn't drastically higher.
Depending on the physical design of the package, there may be some kind of threshold for the materials in the stack that the other designs generally keep below.
Rather than hop back and forth across these more sensitive regions due to dynamic activity, the thermal policy could keep the chip consistently on one side or the other over a longer period of time.

The thermal modeling would be easier to calculate if the GPU is allowed to bake a steady state.
At a fixed temperature that has been holding steady for a while, a number of variable behaviors can be simplified at the time frame the controller operates at.
I'm thinking of the temperature-based terms for leakage and power consumption employed by the logic monitoring on-die activity counters, as well as the behavior of the thermal solution.
Those terms would fall out more readily if the GPU forces a known target that is kept stable long enough to iron out history effects in the heatsink and fan speed.
 
Back
Top