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

Why do we care about clock speed? Dont we care about throughput? So long as highest utilization brings the highest throughput who cares what the clock is.

Do i care if my plane travels @ 30,000 ft @ 800 MP/H or @ 25,000 @ 750 MP/H. No i dont, I only care what time i get to my destination.
 
Why do we care about clock speed? Dont we care about throughput? So long as highest utilization brings the highest throughput who cares what the clock is.

Do i care if my plane travels @ 30,000 ft @ 800 MP/H or @ 25,000 @ 750 MP/H. No i dont, I only care what time i get to my destination.

Err, does your example not answer your question? Read it a few times...Throughput and clocks are not orthogonal concepts, but rather directly related. So yes, clocks are important for throughput just like being able to do an extra 50 MPH is important for someone who cares about the time of his arrival (even when factoring in the extra time needed for descending from the superior altitude).
 
I dont think we can compare to CPU turbo mode....with my i7 overclock, what i set is what i get under load. In BF4, it may not load all cores, but the CPU speed runs at 4Ghz. In IBT, it does loads 100% all cores, and my CPU still runs at 4Ghz. The only time i seen CPU throttling is because of poor quality motherboard....activating its own safety measures.

With Hawaii turbo, its not the case. I can set at 1.1ghz and it still runs between 945-1.1ghz....most of the time it does not stay at 1.1ghz...at times it even dips to 840 (assuming GPU-Z is relatively accurate)....so frustrated because it did impact my game with some stuttering.

CPU overclock also allows us to turn off the turbo offset mode..

Another quirk of Hawaii turbo is that the idle voltage is now 0.91v and max is 1.224v via GPU-Z. Way above what my old Tahti did at 0.81v to 1.174v (the load vddc is acutally around 1.158v). Furthermore, Hawaii GDDR5 seems to have only 2 states 300mhz and 1250mhz and it keep switching to load stage more often than Tahiti....So i assume the heat and consumption goes up...not sure if such volts and fast GDDR5 switch is necessary...

Imo it could just be immature bios/CCC...Hawaii feels a lot like 4890...more refinement is needed...i still love my 6970 on how refined it was on launch...i also love the 7970 because of how it overclocks and being rather refined even if it was AMD first 28nm...Hawaii...not feeling it atm...

Another rant...why is the default idle fan speed only at 20% / 950rpm? Thats too low....when will AMD allow us to use custom fan profile in CCC OD?
 
Why do you are about the voltage ranges? All Hawaii's are pretty close in power, but those are the voltage ranges to get your ASIC to run at those speeds; another part may well have lower voltage ranges but it will still have a very similar power profile and run in a very similar manner. The same was the case with you Tahiti, although that relationship was a little more coarse than on Hawaii.

The memory status is no different from previous GPU's, we only use "high" and "low" speeds and low is basically only activated at the desktop when there is little activity.

With regards to overclocking, yes, you can increase the power limits and increase the clocks and you will still get power state changes, again this is expected. The end result (performance increase) is whats necessary.
 
Dave, I wanted to ask, could my 290X run at an specific clock and an specific voltage bin and yours run at the same clock but in a different voltage bin? Both having the same base voltage? I remember that an old Kepler review inferred that they could run like this, having the same board achieving same clocks and different voltages.
 
Hi Dave, do u think AMD can share the ASIC quality range of voltage?

I would also like to see the ability to set custom fan profile in CCC instead of having install third party apps.

Last of all, do you think we will be seeing another never settle performance gains on the Hawaii?

On the surface, the paper specs gain from Tahiti to Hawaii seems bigger than Cayman to Tahiti....Tahiti started with a 25-35% gaming improvement over Cayman but today its like 50-100% faster than Cayman..
 
About the Tesla boost thing: for CPUs, a workstation that runs Ansys one day may also run Visual C the other. But the amount of applications for CUDA is comparatively very low which also decrease the chance of use case overlap for the same person.

So contrary to CPUs, I believe that GPUs as compute devices are typically bought for 1 task and 1 task only. This makes it much easier to profile your device for cases where you can crank up the clock.

Also, I can imagine that there are plenty of customers who don't care about being 100% predictable and would like to see more aggressive up and down clocking... So I'm a bit surprised they don't offer that kind of option also.
 
Dave, I wanted to ask, could my 290X run at an specific clock and an specific voltage bin and yours run at the same clock but in a different voltage bin? Both having the same base voltage? I remember that an old Kepler review inferred that they could run like this, having the same board achieving same clocks and different voltages.
The binning of Hawaii actually leverages learning from our CPU side by doing what I call "power binning". This means that speed, static power and dynamic power are all taken into account in the binning process. The onus is to, as much as possible, keep the same power utilisation profile from one ASIC to the next (at the same clocks) meaning that one ASIC will actually be operating at different voltage from another while performing very similarly from both a clock speed and power perspective.

This is one of the reasons why we went non-deterministic with PowerTune because we've actually pushed more of the deterimism into the product selection process.
 
Hi Dave, do u think AMD can share the ASIC quality range of voltage?
ASIC "quality" is a misnomer propobated by GPU-z reading a register and not really knowing what it results in. It is even more meaningless with the binning mechanism of Hawaii.

I would also like to see the ability to set custom fan profile in CCC instead of having install third party apps.
The control of the fan is intrinsically tied to the PowerTune implementation.

Last of all, do you think we will be seeing another never settle performance gains on the Hawaii?
Its a new high end ASIC, with a different engine configuration (much more fillrate) with some new architectural features, so for sure there is plenty to play with and tune from here. Tahiti was a significantly different architecture from the previous generation, though, so the room for optimisation was likely greater on the whole.
 
Err, does your example not answer your question? Read it a few times...Throughput and clocks are not orthogonal concepts, but rather directly related. So yes, clocks are important for throughput just like being able to do an extra 50 MPH is important for someone who cares about the time of his arrival (even when factoring in the extra time needed for descending from the superior altitude).

stop taking the analogy so literally, i could have replace speed with engine rpm which would vary thrust at different altitudes as does resistance but then i would of had to explain that.......

so take two



Do i care if my plane travels @ 30,000 ft @ 3000 engine RPM or @ 25,000 @ 2750 RPM. No i dont, I only care what time i get to my destination.
 
The binning of Hawaii actually leverages learning from our CPU side by doing what I call "power binning". This means that speed, static power and dynamic power are all taken into account in the binning process. The onus is to, as much as possible, keep the same power utilisation profile from one ASIC to the next (at the same clocks) meaning that one ASIC will actually be operating at different voltage from another while performing very similarly from both a clock speed and power perspective.

This is one of the reasons why we went non-deterministic with PowerTune because we've actually pushed more of the deterimism into the product selection process.

Thank you.

I see, I expected to see more power variation between Hawaiis, but this way you can possibly defend yourselves from having to explain it to the consumer (there are no golden samples) while making overclockers and undervolters happy.
 
stop taking the analogy so literally, i could have replace speed with engine rpm which would vary thrust at different altitudes as does resistance but then i would of had to explain that.......

so take two

Do i care if my plane travels @ 30,000 ft @ 3000 engine RPM or @ 25,000 @ 2750 RPM. No i dont, I only care what time i get to my destination.

I think the point you were making is that GPUs throttle because of high power draw, and usually if they have high power draw, it's because of high utilisation, which means they still get a lot of work done, even with the throttling.

Is that it? If so I'm not sure the plane analogy helps, but you have a fair point.
 
I'm beginning to wonder if Hawaii's high operating temperatures are more to do with AMD's aggressive transistor density and less to do with the alleged crappy cooler. AMD is pushing nearly 14.2m transistors per mm/2 on a 438mm die compared to about 12.6m trannies per mm/2 on a 561mm GK110.

So NV is cooling a chip with only 14% more transistors with about 28% more surface area to contact the heatsink. The actual vapor chamber(not the shroud) heatsink on the 290x can't be that much more inferior, if any, to the unit on NV GTX 780+ series.
 
I'm beginning to wonder if Hawaii's high operating temperatures are more to do with AMD's aggressive transistor density and less to do with the alleged crappy cooler. AMD is pushing nearly 14.2m transistors per mm/2 on a 438mm die compared to about 12.6m trannies per mm/2 on a 561mm GK110.
As explained before, the ASIC temp is targeted from a performance perspective. The fan control mechanism we (and you) can target any temp you want, but from a performance/acoustics perspective higher ASIC temps are going to offer greater benefits if the overall solution is thermally TDP limited in any way.
 
I still dont get why Nvidia cooler is so much better??
They are the same 75mm fan blowing the same vapor chamber aluminum slab iirc...ime Hawaii fan noise signature is acoustically better than Tahiti.

In future i do hope AMD consider some form of AIO H2O cooling addon...every threads i saw on H2O on GPU, the temps are significantly better even with simple AIO...
 
I still dont get why Nvidia cooler is so much better??
They are the same 75mm fan blowing the same vapor chamber aluminum slab iirc...ime Hawaii fan noise signature is acoustically better than Tahiti.

In future i do hope AMD consider some form of AIO H2O cooling addon...every threads i saw on H2O on GPU, the temps are significantly better even with simple AIO...


NV's cooler looks nice but in reality it probably doesn't perform that much better than the 290X unit. I bet if you could slap it on the 290x, the chip would still be in the same temp ballpark. AMD's temp and peformance target for the chip with a stock factory blower style heatsink/fan is simply pushing it, if you want a quiet card that is.

As Tomshardware has shown though, with aftermarket cooling the chip is quite a beast.
 
Back
Top