AMD 45nm process technology in Shanghai

[Raqia]

An article about AMD's current implementation of its 45nm node:

http://www.eetimes.com/news/design/showArticle.jhtml?articleID=212002243

"The transistor drive current for AMD's 45-nm devices is much lower than that of the Intel HKMG transistors. But power consumption is quickly becoming a high priority for server chips. AMD's transistors exhibit very low channel leakage. Our transistor benchmarks indicates that leakage current is less than one-third of the value measured on AMD's 65-nm process. It's also significantly lower than the Intel 45-nm HKMG process. In fact the Ion/Ioff ratio for AMD's PFET is nearly 10 times better than that for the Intel PFET."

 

[Moloch]

Now if they could only make a chip using that apparently superior process on par(i.e within 10% at the same clock speed) with Nehalem in a desktop environment as well as a server environment they would be golden, such a feat would be amazing given their (reduced) budget compared to intel.

 

[MulciberXP]

...

 

[3dilettante]

"The transistor drive current for AMD's 45-nm devices is much lower than that of the Intel HKMG transistors. But power consumption is quickly becoming a high priority for server chips. AMD's transistors exhibit very low channel leakage. Our transistor benchmarks indicates that leakage current is less than one-third of the value measured on AMD's 65-nm process. It's also significantly lower than the Intel 45-nm HKMG process. In fact the Ion/Ioff ratio for AMD's PFET is nearly 10 times better than that for the Intel PFET."

I'd like to know the exact comparison that is being made.

There are multiple grades of transistor that are sized to provide a given level of drive current for some level power consumption.
The relationship between max current and leakage is not so much linear as it is exponential.
Increase max drive current by a little, and leakage goes up by a lot.
So if your drive current lags by nearly 40%, but you tout your lower leakage, you're just hiding the fact that your transistors might explode if you tried matching on drive current. There's nothing stopping the competing "high leakage" process from scaling back to mediocre performance levels to get similar reductions in leakage.
Actually, Intel's high Vt transistors that correspond to lower performance are still better than AMD's given numbers.

http://www.realworldtech.com/page.cfm?ArticleID=RWT011608222300&p=6

Are they comparing their transistors to Intel's low Vt or high Vt transistors?
According to this chart, AMD's numbers fare poorly to either, and might be inferior to IBM's numbers for 45nm.

 

[Moloch]

It's going to have to be so much better because it will be spending most of its life time competing against intels 32nm process

How if the phenom 2 comes out in january?

 

[bearmoo]

Increase max drive current by a little, and leakage goes up by a lot.
So if your drive current lags by nearly 40%, but you tout your lower leakage, you're just hiding the fact that your transistors might explode if you tried matching on drive current. There's nothing stopping the competing "high leakage" process from scaling back to mediocre performance levels to get similar reductions in leakage.

Rumor goes overclocked PhenomIIs have modest power increase over stock and is actually better than Intel's number. We will see how that turns out soon

 

[3dilettante]

At what voltages?
Fiddling clocks at the same or nearly the same voltage leads to modest power changes.

Voltage changes are have a quadratic relationship with power.
Perhaps the idling logic is effective enough to gate portions of the chip as rapidly as possible, but the peak power draw won't be helped too much by that.

 

[bearmoo]

At what voltages?
Fiddling clocks at the same or nearly the same voltage leads to modest power changes.

Voltage changes are have a quadratic relationship with power.
Perhaps the idling logic is effective enough to gate portions of the chip as rapidly as possible, but the peak power draw won't be helped too much by that.

At stable voltages I presume, some overclocks of deneb ES samples in Taiwan does seem to show Deneb need higher Vcore to achieve clockspeed comparable to Intels so this does reflect the lower driver current story, but the story is even if overvolted the power comsumption stays tame. But its a rumor.

 

[3dilettante]

I'm not willing to give credence to the idea that overvolting leads to no power gain. Outside some intervening factor that is purposefully manipulating the clocks, voltage supply, or gating off silicon without telling the tester, the chip shouldn't be unaffected by a voltage increase.

 

[bearmoo]

"AMD plans to introduce its version of an HKMG process developed in partnership with IBM midway through 45-nm production. However, AMD's first 45-nm device is manufactured using polysilicon gates with oxy-nitride dielectric. "

looks like this is AMD's solution for 2010 ...

 

[bearmoo]

I'm not willing to give credence to the idea that overvolting leads to no power gain. Outside some intervening factor that is purposefully manipulating the clocks, voltage supply, or gating off silicon without telling the tester, the chip shouldn't be unaffected by a voltage increase.

I am not saying there are no power gains, more like power doesn't explode like the Phenom 9000s when overclocked. Besides AMD cpus, 45nm ones included seem to have higher default vcore than Intel's for a given frequence. We don't know what the voltage delta for an given overclock might look like.

 

[dizietsma]

Forget all that gobbledegook shit

Does it give loads of powaa for little wonga and overclock like a mutha?

That sentence with it's poor English is far more relevant that high tech giberish to the home user in the current economic climate.

If the answer is yes then AMD are back in the game

 

[3dilettante]

This is a thread about process technology, a topic which by its essence is technological gobbledegook.

Shanghai's initial offerings allow it to compete with a higher bracket in Intel's pricing scheme. The initial clocks don't put it in full contention across the entire range.

Intel still has some pricing play left, if it desires to use it. A mature 45nm process in volume with mature yields can cut prices more readily than one that we haven't seen out in the market.

Nehalem's additional costs in a totally new platform should blunt its advantage, but I wonder if the lowest-priced Nehalem SKU will turn out to be the highest a desktop Shanghai can be priced.


Your reference to overclocking like a mutha doesn't belong in the same argument as the resort to the "home user", and I'll wait for more widespread overclocking beyond the limited numbers of chips AMD skimmed from highest bins.