PPC 970 Addendum @ Ars
Hey all.
Just wanted to let everyone know that there is an addendum to Hannibal's in-depth look at the 970. This addendum comes after he conducted a phone interview with IBM's Pete Sandon, the chief architect of the 970, and David Edelsohn, who works on optimization/compilor technology for the 970.
You can find it here:Interview
The interview clarifies some sticking points in Hannibal's original two 970 articles, and brings to light the reason for some of the design decisions for things like the VMX/Altivec unit and the chip itself
Just wanted to let everyone know that there is an addendum to Hannibal's in-depth look at the 970. This addendum comes after he conducted a phone interview with IBM's Pete Sandon, the chief architect of the 970, and David Edelsohn, who works on optimization/compilor technology for the 970.
You can find it here:Interview
The interview clarifies some sticking points in Hannibal's original two 970 articles, and brings to light the reason for some of the design decisions for things like the VMX/Altivec unit and the chip itself
Comments
The 5 instructions group where needed, when you deal with more than 200 instructions in flight simultaneously. The IBM engineer said that it's require when you have more than 100 instructions in flight. However, at the light of their experience with the power4 design, they said that they where quite happy with that.
Last points, they said that they cannot give infos about the roadmap, but that the picture will be big
A@ron
Originally posted by A@ron
I'm just glad to hear confirmation that IBM and Apple have a roadmap for the future of the PPC. After so long with MOT and the stagnation, it felt like neither company had a business plan for the future. With IBM I have more faith that they will not only succeed in their plans but surpass them (dual 2 GHz anyone?).
A@ron
Although not to rain on anyone's parade, and while I agree that I have more faith in Big Blue than I ever did with Motorola, Motorola also had a roadmap. The difference being that they failed to deliver, changed that roadmap and failed to deliver *again*[repeat].
This is *not* to imply that I think IBM will follow Mot.'s Path to Destruction, but rather a clarification of the facts.
This will also be useful in portable configurations, as lower speed == less power == less heat == longer battery life.
Some very good insights indeed
It's a really good new.
Originally posted by Powerdoc
Yes he points out two important things : the altivec unit looks like the previous G4 one (excepting the bandwitch)
Not totally. While it does look kind of similar, it has some key differences.
Contrary to what I stated in Part II of the 970 article, any vector instruction can dispatch from any of the four non-branch dispatch slots.
Peter Sandon: So you can dispatch, for example, four permute instructions on one cycle to the permute issue queue. Or for example you could dispatch four ALU instructions to the VALU issue queue. Or you can do combinations of those. So you can dispatch four of any kind of vector instruction on a given cycle to the issue queues.
This makes the 970 look a lot better for Altivec code.
This makes the 970's unit different from the first G4s unit.
While the 970's vector unit turned out to be better than I'd thought, it still didn't look as good on paper as the G4e's vector unit. I asked Sandon why they decided to go with a design that looked like the original G4, instead of something more along the lines of the G4e. Was it a time-to-market thing, I asked?
Peter Sandon: Yeah, it was, and I think what you'll find or that what application developers will find is that this will actually still perform pretty well because we have enough dispatches to it and to the other units that you need to use in vector code. You need to be able to dispatch to the vector units and the fixed point units to keep vector code going. So yes, it's both time-to-market and that there was not a great advantage to the wider dispatch to the VMX itself.
Also this image: http://www.arstechnica.com/cpu/03q2/...970-vector.png
Shows a good diagram. While it does look similar to the first G4 altivec, it is different.
--
Ed
Originally posted by Ed M.
However, I doubt that you'll hear any retractions form all those columnists that published stories attempting to debunk the results.
--
Ed
You are right, however columnists almost never retract.
It's good to be back in black.
Originally posted by Aquatic
Want to hear something really funny? I'm working out Sears now (no that's not the part)...
<OT>
I bet you are either working in either division 20 or 57... vaccuums or tvs and electronics. The worts part of Sears is the loss prevention people treat you like the criminal not the actually criminals. I bought 2 iBook batteries that were back in storage way after Sears dropped Apple (they added $100 to the cost of the comp. who would be dumb enough to buy them there) for $7.00 a piece. The loss prevention people held me in the room for probibly 30 minutes trying to figure out how I got such a good deal. That and their employee discount sucks!</OT>
I cannot wait for next quarters earnings report to see how well the G5s are selling and shipping.
A@ron