Carrell Loses His Baby: Say Goodbye to Sideport

Sitting at dinner with Eric Demers and Carrell Killebrew is honestly one of the best experiences I’ve ever had working with ATI. Before he got huge and subsequently left, I used to have annual dinners with Pat Gelsinger at Intel. They were amazing. To get to sit at the same table as someone as talented and passionate as a Gelsinger, Demers or Killebrew is one of the most fortunate and cherished parts of my job.

Eric was telling me about how they trimmed down 870 from over 400mm2 down to 334mm2 and how wonderful the end product was. I stopped him and asked for more detail here. I wanted an example of a feature that they had to throw out but they really wanted to keep in. Manufacturers rarely tell you what they threw out, marketing likes to focus on what’s in the chip and make everything sound like a well calculated move. Thankfully, marketing wasn’t allowed to speak at my dinner.

Eric turned to Carrell and said: “i know one feature we could talk about.”

“Sideport”.

Carrell responded, “OH MY GOD, that’s totally not fair.” (note that Carrell does not sound like a teenage girl, imagine that phrase just spoken more engineer-y).

When ATI first talked about the Radeon HD 4870 X2 they told us about a feature called Sideport. It was a port off each RV770 GPU that could be used for GPU-to-GPU communication.


Sideport as it was intended to be used

The whole point of doing CrossFire in alternate frame rendering mode (AFR) is that the chips never have to talk. The minute you have to start synchronizing them, you kill performance. Sideport was supposed to alleviate that.

Unfortunately, due to power reasons, Sideport was never used on the 4870 X2. ATI’s reference design had it disabled and all vendors followed suit.

Sideport was Carrell Killebrew’s favorite feature, and he had to give it up.

In early 2008 ATI realized they had to cut this chip down from 20 - 22mm on a side to 18mm, everyone had to give up something. Carrell was the big advocate for making 870 smaller, he couldn’t be a hypocrite and not give anything up.

A bunch of my conversation with Carrell at this point had to go off the record. Sideport would have been useful in RV870, but it’s unfortunately not there. Although he did tell me not to be surprised if I saw Sideport again at some point. Carrell doesn’t give up easily.

Adjusting Trajectory & Slipping Schedule What Made it All Ok: 4 GPUs in < 6 Months
Comments Locked

132 Comments

View All Comments

  • Dudler - Sunday, February 14, 2010 - link

    Whoa... Just Whoa.
    Very insightful and interesting.

    Amd moving to GF next generation and Norther Islands this year?

    Whoa.

    Thx Anand :)
  • carlhenry - Sunday, February 14, 2010 - link

    i really loved the eyefinity being so 'mission impossible'. awesome article!
  • Patrick Wolf - Sunday, February 14, 2010 - link

    The first rule of SunSpot: You don't talk about SunSpot. :D

  • mubeen - Tuesday, February 16, 2010 - link

    Such an informative article Mr. Anand. Really appreciate your efforts.
  • just4U - Tuesday, February 16, 2010 - link

    I agree it's a great article, nice to see these ones pop up from time to time. It don't really matter to me if it's nvidia, amd or intel.. just having a more insightful read is refreshing.
  • DearSX - Sunday, February 14, 2010 - link

    I really do a like article like this. I wonder how old others are cause I'm 25 and I did not even notice how much reading things like this mattered. Its more important than buying something or performance figures imo.
  • Paulman - Sunday, February 14, 2010 - link

    I, too, will second that. Really cool. :P
  • MrSpadge - Sunday, February 14, 2010 - link

    Many thanks for this interesting article and especially the in-dpeth information about TSMC, the process and how ATI deals with it. Thanks to you, Anand, for writing it and thanks to the good guys at AMD for sharing this information!

    Beyond excellence :)
  • hyvonen - Monday, February 15, 2010 - link

    And NVidia's call for "zero defects, zero variability" was jaw-dropping. Do they have any idea how silicon is manufactured?

    Unbelievable cluelessness... with this attitude, they are out of business in the next two years. They are completely unprepared to harness the potential in the latest process nodes.

    Meanwhile, it also sounds like TSMC is way over its head in this - uncharacteristic screw-ups with the chamber matching, bad leakage etc. Not to mention the choice to delay HKMG until 28nm (and having to kill 32nm because of that). GF is going to eat them for lunch, and take over the foundry business.
  • ET - Sunday, February 14, 2010 - link

    Indeed. One of the more interesting articles I've read in a while.

Log in

Don't have an account? Sign up now