I got a few comments about yesterday’s article, most particular this quote from a linked article:
“For example, a hypothetical desktop processor using 70 watts will run an AMD bulk silicon processor with about 40 million transistors. Our 64-bit desktop processor, the AMD Athlon based on Hammer technology, will operate approximately 100 million transistors at the same 70 watts.”
This would seem to indicate tremendous improvement due to SOI.
CPUs nowadays can be divided into two parts: the “real” CPU and the L2 cache. “Real” CPU transistors chew up more power than cache CPU parts.
Does the amount (actually the proportion) of cache have anything to do with this?
How many transistors belong in each category in these CPUs?
A TBredB has 256K L2 cache. It has a total of 37.6 million transistors.
A Barton is essentially a TBredB with an extra 256K L2 cache. It has 54.3 million transistors.
Do the math: 54.3 million – 37.6 million, and you’ll see that 256K cache means about 16.7 million more transistors.
The “largest” (we’ll speak on this later) Opteron has 106 million transistors.
Knowing that, we can approximate how much of each of the three AMD processors is “real CPU” and how much is cache.
We see that while Opteron has twice the cache of Barton, it also has almost twice the “real CPU” transistors of Barton, too. So this is not a matter of low-wattage cache substantially changing the picture.
What does Opteron have that Athlon doesn’t?
Basically, three things:
You can see a picture of the memory controller here.
The memory controller is the area in the bottom left of the processor. Some quick rough measurements indicate that it takes up about 8% of the processor area.
Making the (big) assumption that transistor density is the same as the rest of the CPU, that means about 8.5 million transistors are dedicated to the memory controller. It may well be somewhat more or less, what’s clear is that it doesn’t take up most of that extra 20 million “real” transistors.
How much power does the memory controller absorb compared to other transistors? Who knows?
x86-64 actually isn’t much. About 3% of the die, which works out to about 1.5 million transistors.
This still a bit less than ten million extra transistors.
Some of these may well be improvements within the Opteron core over the Athlon, but a sizable chunk no doubt are used to make the Opteron more multiprocessor capable than Athlons.
Whatever that number is, those transistors wouldn’t be used in Opteron models running in single or dual mode, might well be even disconnected.
That would mean less power. How much less? Who knows?
All the AMD document says is that the Opteron has a maximum thermal power of 84 watts. It doesn’t say which Opteron has a maximum thermal power of 84 watts, or under what conditions.
Perhaps this fits an 8-way processor at 1.8GHz. In all likelihood, the difference between running a processor 1-way or 8-way accounts for some of the wattage difference.
What may be the rest of it is that 84 watts is the maximum power for some future Opteron running at a higher speed.
What’s The Big Secret?
Looking at just wattages, like I did yesterday, is too simplistic a way to figure out how well AMD’s SOI process is working.
On the other hand, just looking at transistor count is too simplistic, too.
The real problem is not that people are guestimating incorrectly; it’s that they find the need to guestimate in the first place.
If, for instance, an Opteron at 1.8GHz had a listed maximum thermal power of, say, 60 watts, or a typical use of 45 watts, there would be no question that SOI was doing quite a bit of good.
So what’s the big secret? Especially when it has never been one before.
Especially when the real news would look to be good??????
If AMD truly had a very cool running chip on their hands due to SOI, don’t you think they ought to let people know?
If you were out to buy an airconditioner that didn’t chew up too many watts, what you think of a manufacturer who would only tell you the maximum wattage the most powerful unit in the line would require?
If you were out to build rackmount servers, and cost of cooling counted, wouldn’t you want to know how much cooling you required for a particular processor, or bunch of them?
If Opterons really do chew up less power than equivalent Xeons, did it ever dawn on anybody at AMD that this might be a good selling point?
What’s the big secret?