HD 7970 Overclocked to 1.26 GHz: 28 nm Tech Really Stretches Its Legs

[alwayssts]

A fairly clear picture is starting to emerge about 28nm and voltage in terms to laymen and regular cooling.

At stock 1.15v, the only review I saw that didnt use CCC to overclock achieved 1165mhz. AMD mentioned to them 1200 could be possible at stock, ~1300mhz with voltage tuning.

VR-zone using 1.25v achieved an exactly linear increase.
1700mhz was 1.7v right? Also pretty linear.

That crazy 1335mhz part would then infer a voltage of around 1.31-1.32.

things to note:

1. AMD used ~1.31-1.32v stock for 4890 on 55nm...see point #2. One can always assume they would have wanted the same for 40nm had it not been leaky as hell above ~1.175v (6870/6970 stock voltage). AMD has shown a pattern of being conservative on stock voltage by about 5% on first-gen parts on a process...re 4870/5770/5870. The 3000 series is an anomaly because of the lack of decap layer...hence why yields were so good. Voltage was high...clockspeed not.

2. Intel suggests not running chips above 1.325v, as it has been shown to shorten lifespan of their cpus. AMD may agree with this given their past voltage settings, guidance on 7970 overclocking, and that Sapphire pushed-to-the-max model. This also may get VERY close to the 300W spec.

Now, TSMC aint Intel, and 28nm hkmg clearly aint 40 or 55nm...but rules of thumb are nice...and 28nm looks to put things back on course of where they should be in accordance to the AMD strategy of small dies with topped out voltages and clockspeeds while staying within tdp specs. Because it appears so well organized, I would be very surprised if the fastest and highest voltage official tahiti sku we see is 925mhz @ 1.15v.

I admire all the facets the ATi group looks at when designing a GPU. Thrifty, practical scaling...just damn smart engineering.

 

[hat]

UPDATE: Notice how the Furmark screenshot shows the card achieving a mere 40fps at 1280x720? A card in this performance bracket can obviously do way better than this. This poor framerate will be down to the power throttling circuitry protecting the card under such conditions by slowing down the performance severely. If the throttle was disengaged the card would perform spectacularly for a short while and then likely cook itself to death, especially with this overclock. Thanks to one of our eagle-eyed forum members for spotting this.

Which means the overclock was probably unstable? I wonder what would happen if the throttling was disabled...

 

[arroyo]

1.26 GHZ ... Great Scott!

That is more than 1.21 gigawatts!

 

[_JP_]

Right meme, wrong comparison

^that couldn't be more incorrect. :\

 

[TheGuruStud]

Ok if we compare new EVGA GTX580 classifide 3Gb which is also with voltage tune can hit over 1.2Ghz with 7970, how much be the different ?
10% or maybe less

about 150W difference, I imagine.

We all remember the original fermi. That is not AMD's goal. Sucking a full 300W the 7970 could run much faster.

 

[dj-electric]

A fairly clear picture is starting to emerge about 28nm and voltage in terms to laymen and regular cooling.

At stock 1.15v, the only review I saw that didnt use CCC to overclock achieved 1165mhz. AMD mentioned to them 1200 could be possible at stock, ~1300mhz with voltage tuning.

VR-zone using 1.25v achieved an exactly linear increase.
1700mhz was 1.7v right? Also pretty linear.

That crazy 1335mhz part would then infer a voltage of around 1.31-1.32.

things to note:

1. AMD used ~1.31-1.32v stock for 4890 on 55nm...see point #2. One can always assume they would have wanted the same for 40nm had it not been leaky as hell above ~1.175v (6870/6970 stock voltage). AMD has shown a pattern of being conservative on stock voltage by about 5% on first-gen parts on a process...re 4870/5770/5870. The 3000 series is an anomaly because of the lack of decap layer...hence why yields were so good. Voltage was high...clockspeed not.

2. Intel suggests not running chips above 1.325v, as it has been shown to shorten lifespan of their cpus. AMD may agree with this given their past voltage settings, guidance on 7970 overclocking, and that Sapphire pushed-to-the-max model. This also may get VERY close to the 300W spec.

Now, TSMC aint Intel, and 28nm hkmg clearly aint 40 or 55nm...but rules of thumb are nice...and 28nm looks to put things back on course of where they should be in accordance to the AMD strategy of small dies with topped out voltages and clockspeeds while staying within tdp specs. Because it appears so well organized, I would be very surprised if the fastest and highest voltage official tahiti sku we see is 925mhz @ 1.15v.

I admire all the facets the ATi group looks at when designing a GPU. Thrifty, practical scaling...just damn smart engineering.

Please do not forget the PCB and it's effects on GPU overclocking, similar to those of a motherboard on CPU overclocking. as i was mentioning on the sapphire 1335Mhz thread, at some point the voltage needed for core operation at a certain frequencies can be much differant between two PCBs with the same core. For example, with the Radeon HD5850 (had 3 reference ones) i could squeeze exactly 835Mhz core clock while with a custome sapphire PCB (witch was created for the whole HD5800 series) i managed to get 925Mhz stable and that reflected on the results using voltage tuning. With the reference cards 1Ghz was achivable under voltages ranging between 1.26 and 1.3 while with sapphire cards that was reachable with voltages ranging between 1.16 and 1.18 witch is spectacular.

That lesson tought me that sometimes with the right PCB you can get some amazing results that you just wouldn't reach with a reference one

 

[qubit]

That lesson tought me that sometimes with the right PCB you can get some amazing results that you just wouldn't reach with a reference one

Oh yeah, absolutely. The whole thing has to work together to give you good results.

A better board will improve on things like voltage regulation (more stable); signal integrity; electrical noise; signal timings around the board (these can be critical); heat generation and dissipation and a lot more subtle effects, such as parasitic board capacitance.

This is why quality CPU motherboards cost so more for example, besides the extra features they offer.