Hot Chips 18: Niagara II Article Online

Article: Niagara II: The Hydra Returns
By: David Kanter (dkanter.delete@this.realworldtech.com), September 8, 2006 12:18 pm
Room: Moderated Discussions
>Excellent article.

Thank you.

>BTW, I predicted 8 months ago that Niagara II would have >two ALUs per core (and
>one FPU) and the ability to execute instructions from 2 >threads each cycle.

I'm not sure I'd characterize it in quite that manner. "The ability to execute instructions from 2 threads" implies that it is an optional sort of thing. IOW, I'd read that as a super scalar architecture that can execute from two contexts, but can also execute 2 instructions from 1 thread (which is not the case).

You certainly were quite on target with the functional units, but Niagara's threading arrangement is very peculiar.

The ALUs are more than just ALUs; pretty much every instruction flows through them. I really don't know how I would try and classify N2; it's almost a conjoined core, but not quite. Perhaps clustered multithreaded? You could sort of consider each pipe it's own cluster...

>...Anyway, it seems likely that Sun's claims of 2x >performance over the first Niagara
>are reasonable. It appears that Niagara2 will be able to >execute a theoretical maximum
>of 16 instructions (from 16 threads) per clock @ 1.4GHz. Of >course, I realize that
>MIPS (millions of instr per sec, not the SGI CPU) is a >silly index of performance,
>however in this case there's good reason to believe that >Niagara2 will achieve results
>reasonably close to the theoretical maximum much of the >time.

I have seen the following figures for per core IPC:

0.41 - OLTP
0.67 - SPECjbb2005
0.58 - SPECweb2005 eCommerce
0.65 - SPECweb2005 banking
0.56 - SPECjappServer2004

This yields an aggregate IPC of somewhere between 3.28 and 5.36. That's really compared to any x86, but it's not close to peak.

>...I've been a Niagara booster for awhile, and the more I >think about it, the more
>I become convinced that the "throughput" approach is the >best strategy that Sun could have devised.

Throughput is great for performance, when it works. The only problem is the licensing, which Sun does not have control over (except at a few large key vendors). Ironically, here is an area where IPF has an advantage. Because Intel and HP have very close relationships with most of their vendors, it is easier for them to ensure that licensing is in accord with performance.

Solaris was by far and away the most popular UNIX in the 90's, and the most widely developed for. Now, I couldn't tell you how many applications are currently being developed on Solaris (perhaps Jesper or Mas could chime in here). Let's just say 15K as a lower estimate; and I would expect this translates into about 3-5K application vendors.

The problem with a throughput approach is that you need to get all those 3-5K vendors to straighten out their pricing. Can Sun get to their top 100 ISV partners pretty easily, sure? But what about the next 500?

I think that throughput computing is definitely great for a lot of workloads, when the licensing can be worked out. Perhaps that would be a good subject for an interview...

DK
< Previous Post in Thread 
TopicPosted ByDate
Hot Chips 18: Niagara II Article OnlineDavid Kanter2006/09/04 10:01 PM
  Hot Chips 18: Core 2 tidbitsanonymous2006/09/04 10:15 PM
    Hot Chips 18: Core 2 tidbitsAndi Kleen2006/09/05 12:29 PM
      Hot Chips 18: Core 2 tidbitsanonymous2006/09/05 01:33 PM
        Hot Chips 18: Core 2 tidbitsAndi Kleen2006/09/05 01:48 PM
          Hot Chips 18: Core 2 tidbitsanonymous2006/09/05 04:40 PM
            Hot Chips 18: Core 2 tidbitsanonymous2006/09/06 07:20 AM
  Hot Chips 18: Niagara II Article OnlineGarius Bias2006/09/05 02:35 AM
    Hot Chips 18: Niagara II Article OnlineDavid Kanter2006/09/05 01:08 PM
  Hot Chips 18: Niagara II Article OnlineChuck2006/09/06 04:59 AM
    Hot Chips 18: Niagara II Article OnlineDavid Kanter2006/09/06 09:50 AM
  Hot Chips 18: Niagara II Article OnlineTom W2006/09/08 10:42 AM
    Hot Chips 18: Niagara II Article OnlineDavid Kanter2006/09/08 12:18 PM
Reply to this Topic
Name:
Email:
Topic:
Body: No Text
How do you spell green?