By: David Kanter (dkanter.delete@this.realworldtech.com), February 3, 2013 11:09 pm
Room: Moderated Discussions
[snip]
>
> This is all very true, but I'm not sure if you're in agreement with me about
> the quality of this example with respect to disruptive changes?
>
> I'd also still like to know where David got his 4x number from? That just
> seems too specific for it to not have a source (unlike the guy in the article
> who, I think, just made up the order of magnitude number on his own)
A GPU delivers around 4X more bandwidth than a CPU, and best case is around 2-3X perf/W. Clearly if you provide as much of an advantage as a GPU, some people will take you seriously.
The reason you need 10X according to Peter is to sustain a sufficient advantage over the long term. When your first gen product comes out, if you target 10X, you will perhaps deliver 4X. But then your second generation will take quite some time to deliver, while Intel will be merrily shrinking to the next node and ramping a new architecture. So you really need the 4X to:
1. Provide enough value for people to kick the tires on something unproven and different.
2. Sustain your team until the 2nd generation can be delivered with a similar performance improvement
3. You need to aim for 10X if you want to hit 4X.
I'd mostly agree with those points, but I think you can get away with a bit less.
David
>
> This is all very true, but I'm not sure if you're in agreement with me about
> the quality of this example with respect to disruptive changes?
>
> I'd also still like to know where David got his 4x number from? That just
> seems too specific for it to not have a source (unlike the guy in the article
> who, I think, just made up the order of magnitude number on his own)
A GPU delivers around 4X more bandwidth than a CPU, and best case is around 2-3X perf/W. Clearly if you provide as much of an advantage as a GPU, some people will take you seriously.
The reason you need 10X according to Peter is to sustain a sufficient advantage over the long term. When your first gen product comes out, if you target 10X, you will perhaps deliver 4X. But then your second generation will take quite some time to deliver, while Intel will be merrily shrinking to the next node and ramping a new architecture. So you really need the 4X to:
1. Provide enough value for people to kick the tires on something unproven and different.
2. Sustain your team until the 2nd generation can be delivered with a similar performance improvement
3. You need to aim for 10X if you want to hit 4X.
I'd mostly agree with those points, but I think you can get away with a bit less.
David