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

AMD Radeon R9 290X Hawaii - The Configurable GPU?

LOL
clockspeed4.png

Were those 780Ti clocks? The author was carefully vague....
 
There has to be a reason why AMD is using those coolers. Maybe they are sitting on shedloads of stock, or have some deal with whatever company provides them for x amount of years?

Whatever, they've shot themselves in the foot yet again. Such needless hassle over a simple part, and they could easily have charged $50 more with a better one. Instead they've basically snatched a draw from the jaws of victory.

Cards with a shitty cooler is intended to lower the price as much as possible..

This is something that people willing to install a custom cooler (air/liquid) will appreciate a lot
 
http://forums.anandtech.com/showpost.php?p=35699215&postcount=24

And incidentally, Ryan Smith reports the base clockspeed of the 290 as 662 MHz. Whether or not that's in the official documentation I have no idea. "Up to" marketing is misleading, I'll give you that.
AMD does not officially specify a base clockspeed (or a TDP for that matter). Those numbers are taken from turning down the fan speed limit and hitting the card with FurMark; the card will drop to its base clockspeed and then override the fan maximum to keep it there.
 
The turbo functionality introduced with Kepler is where at least some of the leeway was exploited.
Yes, you're right.
The thermal diodes used to feed into typical monitoring software measure ramps on the order of significant fractions of a second. That's a value modified by the thermal capacity of the heat sink, the imprecision of the diodes, and conduction across the GPU over a pretty long period of time.
But the capacity of heat sink must be the biggest factor. And that's something that may vary quite a bit due to variances in manufacturing, assembly, and, of course, usage of third party coolers. That is not something you can model by collecting the toggle rate of various units across the chip. (Well, you could create such a model, I guess, by correlating toggle rates with observed temperatures? Let's assume they don't do that. ;) )
The instantaneous power draw and localized heating for spots in the logic can go up and down more quickly, particularly if other elements like fan speed ramp or cooler quality are constrained. This may not be a particularly close proxy, given its age, but for a modeled CPU in an era of lower power density, there are localized temperature ramps that happen in hundredths of a second. There's an example of a 5-10 degree ramp that is basically over in .1 seconds.
But that's still a 3 orders of magnitudes larger than the microseconds' granularity to regulate voltage.

I'm trying to see how this typically would work:
- some unit gets a burst of activity
- this raises an alarm of potential localized critical temperature
- lower voltage by a few mV?
- lower clocks all across the die as well because of slower transistor switching speed. (I'm assuming that this will lower the voltage across the whole die.)

Voltage is a very big hammer to play with, and normally used at the macro level. If you're concerned about local temperatures, wouldn't it make more sense to have some clock pulse dropping mechanism at the smaller unit level? This way, you get very granular control and don't impact the whole die.

Maybe they already have that, but then, again, that raises the question of how much additional benefit you get by furiously regulating voltage.

(I'm not questioning AMD for doing this, they obviously are experts at this, just trying to understand the mechanics of it.)
 
Im curious.... what makes Nvidia stock cooler that much better? Looks the same-sized vapor chambered centrifugal blower design? Does the metal enclosure helps that much...? Amd is even using some special thermal pad iirc....are we sure its not just poor leaky bins?
 
Were those 780Ti clocks? The author was carefully vague....
Probably a GTX780. And according to their review, the 290X still beats the 780 on average with the 40% fan profile (with the lower clocks shown there). The clock speed may be lower, but it is still faster. With better cooling (or more noise with the 55% fan profile) it decisively beats the 780. I wouldn't read much into the drop of clock speed for the 290X (which shows, that there is a lot of headroom with a more capable cooler) other than one has to be very aware of the individual environment one puts the card into (to a larger degree than it's the case also with an nV GPU) to not get surprised by the resulting performance difference to reviews.
 
Last edited by a moderator:
Im curious.... what makes Nvidia stock cooler that much better? Looks the same-sized vapor chambered centrifugal blower design? Does the metal enclosure helps that much...? Amd is even using some special thermal pad iirc....are we sure its not just poor leaky bins?
While the metal enclosure may help a bit (but it is not directly connected to the actual cooling surface, so it's probably a small effect), I think it's mostly the fan itself. It is slightly higher (I think the VRM cooling is done differently saving a few millimeters height for the baseplate which can then be used for the fan). That means it needs a lower speed for the same throughput. Furthermore, it looks like they optimized the number and angle of the fan blades. It uses more blades at a slightly lower angle, but no idea how much difference this makes compared to the higher fan height. But it looks like the AMD design keeps the VRMs cooler in exchange. But the added noise is a fairly large sacrifice for that

One area of improvement one could think of is the dual DVI output blocking maybe 40% of the slot opening. One probably wants to keep the two DVI outputs for the time being. But I've seen a custom plug on some Dell computer some time ago which integrated 2 dual-link DVI plugs into one higher density plug (just slightly larger than a single DVI output) and supplied a breakout cable for the two DVI outputs. That way one could get back the full slot for the exhaust, which probably enables the same airflow at a slightly lower fan speed.
 
Were those 780Ti clocks? The author was carefully vague....

Either that or the 780 "Ghz edition" if it really exists. No way it's the standard 780 with it's typical boost of 900Mhz and even less so the Titan.

Incidentally that was a very good article and everyone should be taking note. The advertised clock speeds of the 290 cards are flat out misleading. A system of base clock + typical boost like NV would be much more informative.

In nvidia terms the clock speeds of the 290x are actually more like 727Mhz base and 865 boost.
 
Incidentally that was a very good article and everyone should be taking note. The advertised clock speeds of the 290 cards are flat out misleading. A system of base clock + typical boost like NV would be much more informative.

As long as reviewers inform themselves properly and avoid publishing numbers from "cold" cards the advertised speed discrepancy is less of a problem.
 
Igor Wallosek often does articles for the US toms as well, so hopefully that will see the light of day there soon. Got a lot of time for him, he writes some good stuff.
 
Have to admit it's an atrocious release for AMD. It looks like a potentially fantastic GPU especially for the price but all the initial issues with review samples and throttling could really put a damper on their initial sales boost.

How many reviewers do only one run ? and you think they stop the bench, put the pc in idle for 15minutes or even power off the PC, before launch the second benchmark ? You know how many bench they are doing in follow ?
 
As long as reviewers inform themselves properly and avoid publishing numbers from "cold" cards the advertised speed discrepancy is less of a problem.

My bigger concern is (as stated in the article) how 3rd parties could abuse the "up to" rating. We may not have the luxury of benchmarks of GPU's from 3rd parties so what's stopping them putting a different (but no better) cooling solution on their card, upping the max speed to 1100Mhz in the bios but never actually achieving better speeds than the reference 290x outside of the first 5 minutes of gaming?
 
My bigger concern is (as stated in the article) how 3rd parties could abuse the "up to" rating. We may not have the luxury of benchmarks of GPU's from 3rd parties so what's stopping them putting a different (but no better) cooling solution on their card, upping the max speed to 1100Mhz in the bios but never actually achieving better speeds than the reference 290x outside of the first 5 minutes of gaming?

Reviews? If events of the last few days have taught us anything it should be that we'll see a lot more in-depth cooler data.

What you might actually be worried about is some of them stating max performance at maximum fan speed. This kind of stuff was the natural progression of the turbo can of worms that GK104 opened. Should probably add that AMD has a much tighter reign on their partners than Nvidia has so it probably won't be a huge issue.
 
Last edited by a moderator:
Reviews? If events of the last few days have taught us anything it should be that we'll see a lot more in-depth cooler data.

I don't expect every GPU variation from every vendor will get it's own full review though, and even if they do they'll probably end up difficult to find and limited to 1 sites point of view.

When I buy a GPU I generally look down the list at my favorite online store for the fastest variation at the best price of the GPU model that I want. It looks like I wouldn't be able to do that with the 290 series because vendors will be virtually free to claim any "up to" speed they want without it having much reflection on the cards true performance.
 
I'm sure we all know what coolers to look out for now. IceQ, Windforce's etc. In terms of ultimate performance it's going to be difficult because of the nature of the silicon. If you're really worried about it I'd say take the max advertised clocks and drop it by 50 MHz to avoid disappointment. Don't forget that you might be pleasantly surprised also and end up with a better than advertised sample. It's just the nature of the game now.
 
Does AMD allow partners to go above 10% on overclocks yet anyway? Or is it a timed thing, ie no 10%+ overclocks for the first 6 months? I remember reading something to that effect a while ago.
 
Does AMD allow partners to go above 10% on overclocks yet anyway? Or is it a timed thing, ie no 10%+ overclocks for the first 6 months? I remember reading something to that effect a while ago.

Asus/partner has 50% TDP (not 50% of 208w), = 20% oc from 1000mhz, or ~50% oc from 800mhz.
 
But the capacity of heat sink must be the biggest factor. And that's something that may vary quite a bit due to variances in manufacturing, assembly, and, of course, usage of third party coolers. That is not something you can model by collecting the toggle rate of various units across the chip. (Well, you could create such a model, I guess, by correlating toggle rates with observed temperatures? Let's assume they don't do that. ;) )
Tackling this in reverse order, there should be some correlating between event counters and temperatures. If not at run time, it would be part of the power estimation engine, which estimates power based on event counters and values that are derived from physical characterization of the device. I think the latest versions of that include a temperature factor.

This means there's a fast approximation to actual power draw where those factors generate a value of Joules that will be heating up the chip and heat sink.
The behavior of the heat sink does have specifications for how much power it needs to dissipate within a given period of time.
Having a target temp can allow for certain non-ideal behaviors to fall away--especially if you sample fast enough.

An approximate Joules out value coupled with the thermal design specification can give a rough estimated power to temp change.
Corrective feedback from measured runtime values would keep the chip bouncing within some error bound of its target, where the sampling period is very short and so its error is correspondingly small.
A thermally significant time for a chip and heat sink is something I think is measured in milliseconds, so brief mistakes are limited in scope and such transients are required of the chip and heat sink to handle.

But that's still a 3 orders of magnitudes larger than the microseconds' granularity to regulate voltage.
It would allow the GPU to take a bite into the voltage guard-banding, since it can respond more quickly to demand variations that could cause voltage to drop for a few milliseconds.

Another possibility: the logic that controls the clock rate can probably shift at that rate or faster.
What this might do is allow for more clock options in a given time frame.

If it's stuck waiting for 100 milliseconds, a range of clocks that could have been hit if not for voltage limits are ruled out.

I'm trying to see how this typically would work:
- some unit gets a burst of activity
- this raises an alarm of potential localized critical temperature
- lower voltage by a few mV?
- lower clocks all across the die as well because of slower transistor switching speed. (I'm assuming that this will lower the voltage across the whole die.)
It might be more proactive, or it could if AMD's CPU turbo has any indication.
The hardware might be able to use a heuristic based on qeue occupancy to guess at how much more activity the GPU will receive. This, coupled with the current counter information, can provide a window for roughly how much things can scale up before one of the temp or power limits are hit.
Even if it's a little wrong, the correction can happen ~10us later.

Given the sampling rate, perhaps a determination can be made based on the slope of the temperature curve. A current or imminent temperature violation with an apparently low slope might respond to a clock reduction.
A more severe slope might require a voltage drop as well.

The rapid approximation probably means some of the sources of such a steep slope like an upclock at a high voltage step that hits unexpectedly high utilization won't happen as much, because there's less time for the error to compound itself.

Voltage is a very big hammer to play with, and normally used at the macro level. If you're concerned about local temperatures, wouldn't it make more sense to have some clock pulse dropping mechanism at the smaller unit level? This way, you get very granular control and don't impact the whole die.
That would be a more CPU-like thing to do, looking at the P4's duty cycles or things like Foxton. GPU DVFS is still inferior to CPUs. Hawaii is just not insanely inferior.
The sheer number of nodes and less bursty behavior might figure into why this hasn't been done, the more rigid pipelines and less physical design effort for the ASIC relative to a complex core may be contributors as well.
Perhaps they do, or will someday have duty cycles--although if people are freaking out about the clocks right now, they are not going to enjoy it when the global clock becomes constant at the price of being even less meaningful.

Maybe they already have that, but then, again, that raises the question of how much additional benefit you get by furiously regulating voltage.
It can shave off voltage margins that need to absorb demand spikes, which are worsened by variable clocks. Otherwise, a more conservative voltage has to be selected for most of the time it's not needed, or activity has to be scaled back. The clock/voltage choices become far more limited like they are with Kepler.
 
Back
Top