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


Not sure if this has been talked about anywhere else (if so, apologies), but this article: http://www.xbitlabs.com/news/other/..._Regarding_16nm_FinFET_and_20nm_Progress.html quotes the TSMC CEO as claiming that 20nm has had 5 tapeout already and more scheduled for mobile, CPU, and PLD segments. For 16nm, the claim is that tapeouts are scheduled for mobile, CPU, GPU, PLD, and networking.

I'm probably reading way to much into things, but it sounds like AMD/NV have either already taped out their 20nm products (or maybe AMD is going GloFlo for 20nm and below?), or are going straight to a FinFET process.
 
Last edited by a moderator:
hhvj078.png
Well it would be more credible if the other folding results shown there would be consistent, but the implicit/single precision one also shows the r290x being slower than 7970. So that number there doesn't have much credibility...
Looks though like Dave confirmed it :).
 
Since we know that there are two shader engines in Tahiti that don't show in AMD's diagrams, is there a substantial change in the non-marketing arrangement of units in terms of front end, CU arrays, or render back end subdivision that differentiates a Tahiti shader engine from a Hawaii one?

Has there been a change in how things interface with the memory controllers, other than there being more of everything?
 
The shader engines have been called Pixel Pipes in the M&M architecture presentation at AFDS11.
 
I think they would rather lose a little on perf/W and win by a bit more on perf/mm2, which is likely to be the reason why they went for the latter.

Not really. Performance/mm2 can be killed if RMA return is too high from returns. Its clear that this card is being pushed closer to its limits than gk110, so I would expect higher returns.

In addition, perf/mm doesn't matter at in the professional market.

Performance/watt and heat are king.

IN super computers for example, chips are often passively cooled and such a hot chip would be unusable in such a system.

Both negative characteristics also kills any chance for it to be used in a mobile card and for its chance to make a dual version GPU version of this card.
 
Both negative characteristics also kills any chance for it to be used in a mobile card and for its chance to make a dual version GPU version of this card.

That's pretty ignorant of the fact that we have no control over voltages, as there are no custom overclocking utilities out yet.

I have run my pair of release day reference model 7970s at 1.1Ghz core/1.5Ghz mem since I got them at stock voltage (1.112v, 1.066v with vdroop) flawlessly all this time. So I'm running at faster clocks and at lower voltage than the official 7970Ghz edition. I also set a custom fan profile, during heavy gaming sessions the card maxes out at 70C with 55% fan.

If the 7970Ghz ed. is any indication it looks like AMD jacked up the voltage to get the highest number of GPUs to qualify for 290Xs.
 
Not really. Performance/mm2 can be killed if RMA return is too high from returns. Its clear that this card is being pushed closer to its limits than gk110, so I would expect higher returns.
On what basis?

Performance/watt and heat are king.

IN super computers for example, chips are often passively cooled and such a hot chip would be unusable in such a system.

Both negative characteristics also kills any chance for it to be used in a mobile card and for its chance to make a dual version GPU version of this card.
All chips are "hot" when you run them, no matter what. Hawaii, as a chip, is no different from any other. The only difference is we have changed the fan controller programming to explicitly target an operating point; that operating point can be higher or lower the only question it the performance you want and the thermal capacities of the solution.

In the example of a passive solution within a datacenter, in fact the thermal capacity can potentially be higher - there isn't any concern over acoustics in that environment so the CFM (airflow) from the case fans can be significantly higher thus the thermal solution can be of higher capacity. In that scenario you can either achieve higher performances for the same TDP (if there were headroom on the ASIC) or lower power for the same performance.
 
Not really. Performance/mm2 can be killed if RMA return is too high from returns. Its clear that this card is being pushed closer to its limits than gk110, so I would expect higher returns.

In addition, perf/mm doesn't matter at in the professional market.

Performance/watt and heat are king.

IN super computers for example, chips are often passively cooled and such a hot chip would be unusable in such a system.

Both negative characteristics also kills any chance for it to be used in a mobile card and for its chance to make a dual version GPU version of this card.


I was reading you and ask me, where he ( you ) want to go with this theory..
 
Last edited by a moderator:
Not really. Performance/mm2 can be killed if RMA return is too high from returns. Its clear that this card is being pushed closer to its limits than gk110, so I would expect higher returns.

There will be roughly 30 extra die candidates per wafer for Hawaii. The return rate would have to be astronomically high - near 30%, for them to lose out like this. It's more likely to be 3%.
 
If the component used are made to operate at this temp... no problem..

I install watercooling for make work all my system way under their max temp specification, maybe i increase their lifetime ( a little example with PWM where they are working on my 7970's at 28-32°C max, when under air, they will work at 90-100°C )... but if they are made for function normally at 100°C.. there's no problem. ofc, if they was made for work only at a maximum of 85°C, this will be different.
 
Last edited by a moderator:
Yeah tbh I think AMD can get away with the 95C temps and the explanation for why. The issue is the fan noise and throttling. If this is what we've to expect from them going forward they are going to have to provide better cooling. They've left a lot of performance on the table on some reviews, and crossfire is falling behind Titan Sli.
 
IN super computers for example, chips are often passively cooled and such a hot chip would be unusable in such a system.

Both negative characteristics also kills any chance for it to be used in a mobile card and for its chance to make a dual version GPU version of this card.

They're not really passive, but rather the fans are located on the server. It's a case of fewer larger fans being more efficient than several smaller fans.
 
Yeah tbh I think AMD can get away with the 95C temps and the explanation for why. The issue is the fan noise and throttling. If this is what we've to expect from them going forward they are going to have to provide better cooling. They've left a lot of performance on the table on some reviews, and crossfire is falling behind Titan Sli.

This will depend review, most cfx vs titan and 780 review i have see put way ahead the 290 ..

If you set temperature as a point of limitation, and you put it really high, ofc .. the temp will be really high...

I really think they was made an arch for 20nm... and well in 28nm, it is what it is.... the perf are here, the clocspeed is here...

I m a bit deceipt by the DP rate.... i dont know exactly why take this option, specially now that Adobe and Autodesk are adopting OpenCL, and this can lead to an extremely good marketing in this market, because we find Photoshop, and Autocad, Inventor, 3Dmax in all entreprise who build our life environnement ( autocad series still not integrate GPGPU and i really hope it will soon for the 3D modelisation parts ) ..
 
Last edited by a moderator:
I don't have a problem with the new powertune design aiming for a temperature mainly because they allow you to tweak it within the drivers. If it was solely something behind the scenes that you can't change it would be horrible. I fully expect cards with custom coolers to set a lower default desired temperature in the BIOS. For those into tweaking or installing their own air/water cooling, all the options are exposed aside from voltage.

Nvidia's boost still comes across as a bit less upfront in what you end up with to me. They've basically given the user a reasonably quiet and cool card with much more variable performance if used in a less ideal environment. AMD have now gone down the same path, but have given the user more choice. The quiet mode aims to do what NV are doing but with a higher temp threshold because they are forced to do this to compete on the performance front with a cheaper cooling solution (and possibly also have a more reliable ASIC wrt operating temp), and then the uber mode gives up on low noise to ensure the performance doesn't drop as much. On top of that, it's all customisable in the driver for those with a higher/lower noise tolerance or for anyone delving into after-market/custom cooling.

Boost, turbo, etc started out as bench-marketing tools in my eyes, but have evolved into something which just allows the most performance possible. The unfortunate side effect is that in a less ideal operational environment the performance will not be what was 'advertised' - but with all vendors doing the same thing, this is less of a problem. Exposing the parameters to the user gives the user a choice - perform like the reviewers benchmarks at all costs, or take a hit to suit your hotter case, warmer part of the world, or to not disrupt your quiet room. If the 290x was released in late 2011 using the same thinking as the original 7970, you can bet the clock speed would have probably been around 850MHz. It seems like boost/turbo is the only way we can have our cake and eat it too - well, some of us, those who use modern GPUs/CPUs in a warmer environment will have to make a larger sacrifice one way or the other, performance or low noise. Under the old method of having a static top end clock speed, everyone got the same performance but some experience more or less noise. I guess the new methods put more performance on the table for some while making noise less of a variable. At least now we can tweak the end result a bit more directly.
 
Last edited by a moderator:
Nvidia's boost still comes across as a bit less upfront in what you end up with to me. They've basically given the user a reasonably quiet and cool card with much more variable performance if used in a less ideal environment. AMD have now gone down the same path, but have given the user more choice.

If I'm not mistaken AMD gives you a choice of two fan profiles - uber and quiet. How is that more choice and "more upfront" than the voltage, temperature and power controls of nVidia's boost?

Am I misunderstanding how these things work?
 
If I'm not mistaken AMD gives you a choice of two fan profiles - uber and quiet. How is that more choice and "more upfront" than the voltage, temperature and power controls of nVidia's boost?

Am I misunderstanding how these things work?

Does NV have actually voltage controls in their own control panel, in addition to temp & power controls?
AMD offers temp and power in the control panel + those 2 profiles (and manual max fan speed control too)

ccc-overdrive.png


http://techreport.com/review/25509/amd-radeon-r9-290x-graphics-card-reviewed/4
 
Back
Top