What was the research that you were doing and how did you get to that core thesis?
Maybe even more, what do you mean by heterogeneity here? And then why do you think it's the right way?
Yeah, so maybe I do the concrete one first, because I think it's obviously helpful for somebody who doesn't know our blocks in so much detail to know what we mean.
So when we talk about heterogeneous computing systems, it's really about building a computing system based on all the different accelerators that are out there in the world, all the different chips that are out there in the world.
And we built this at rack scale size, right? So this is not a kind of desktop PC kind of level.
And what Colossum specifically is doing in this context is building the software, design IP, and data necessary to make these systems work at scale ultimately.
And the way that this came to be is, if you take a bit of a longer arch and how we got here, is that we've really had quite a long journey on arriving at this point, partially through being academics beforehand and also just our studies.
I actually did lots of things before, cognitive science, psychology, economics, did some business, neuroscience, computing, CS.
And I think when I think of my own work, I've always been very passionate about this idea of what is an ideal systems-level architecture across these system fields.
And if you look at some of the work we've put out earlier this year, it's actually there with this one theoretical paper which shows that there's this optimization drive towards heterogeneity present across all these systems.
And lots of the empirical work that we've put out obviously really relies on now having shown that these new systems work very, very well.
It's really that all the systems we've really ever observed in nature and also our, generally speaking, human-created systems, they tend towards a geneity and the specialization subcomponents, which are brought together, made work well together.
And what you can obviously show is that the specialization creates absolute crazy benefits if you bring it together correctly.
But there's also many ways of bringing it together incorrectly, right? So...
Sometimes there's this point of like, is it just heterogeneity for the fun of it? Well, it's not.
What was the research that you were doing and how did you get to that core thesis?
Maybe even more, what do you mean by heterogeneity here? And then why do you think it's the right way?
Yeah, so maybe I do the concrete one first, because I think it's obviously helpful for somebody who doesn't know our blocks in so much detail to know what we mean.
So when we talk about heterogeneous computing systems, it's really about building a computing system based on all the different accelerators that are out there in the world, all the different chips that are out there in the world.
And we built this at rack scale size, right? So this is not a kind of desktop PC kind of level.
And what Colossum specifically is doing in this context is building the software, design IP, and data necessary to make these systems work at scale ultimately.
And the way that this came to be is, if you take a bit of a longer arch and how we got here, is that we've really had quite a long journey on arriving at this point, partially through being academics beforehand and also just our studies.
I actually did lots of things before, cognitive science, psychology, economics, did some business, neuroscience, computing, CS.
And I think when I think of my own work, I've always been very passionate about this idea of what is an ideal systems-level architecture across these system fields.
And if you look at some of the work we've put out earlier this year, it's actually there with this one theoretical paper which shows that there's this optimization drive towards heterogeneity present across all these systems.
And lots of the empirical work that we've put out obviously really relies on now having shown that these new systems work very, very well.
It's really that all the systems we've really ever observed in nature and also our, generally speaking, human-created systems, they tend towards a geneity and the specialization subcomponents, which are brought together, made work well together.
And what you can obviously show is that the specialization creates absolute crazy benefits if you bring it together correctly.
But there's also many ways of bringing it together incorrectly, right? So...
Sometimes there's this point of like, is it just heterogeneity for the fun of it? Well, it's not.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.