back
13 comments
I believe the results were presented at GTC this year, and they said the paper (this) was to follow.

edit: Here is the talk: http://on-demand.gputechconf.com/gtc/2018/video/S8122/

and slides: http://on-demand.gputechconf.com/gtc/2018/presentation/s8122...

Instead of NVidia just documenting their uarch, third parties reverse engineer it and present it at... NVidia's conference.
Papers like this have been written for older GPU architectures and it is disappointing to even have the need for them. That such a paper was presented at a manufacturer's conference is icing on the irony cake.

My underdeveloped tinfoil hat speaks that they want to acknowledge the reverse engineering methods involved and are secretly hoping that they get applied to the competition, too, so they can extract the competitor's secrets from a public source instead of ending up in a reverse engineering grey area.

I'm kinda surprised the Nvidia was totally cool with this presentation. Maybe the engineers think this info should be commonplace but are trapped in NDA hell?
Most interesting is who published this (Citadel, LLC)
And who are they?
One of the largest US hedge funds.
GPGPUs and CUDA have been used for predictive financial analysis for some time now, haven't they?
A hedge fund backs research into GPUs because they may deal in finance and GPUs can be used for financial analysis?

Not a chance. This has nothing to do with their IT department or back office requirements.

It is 100% from their "IT", Citadel is essentially a "Quant" firm.

https://www.citadel.com/careers/ultimate-guides/quantitative...

https://wei-tan.github.io/

https://www.citadel.com/careers/the-data-open/

Citadel is not a pure play quant fund, they have a ton of PMs doing conventional fundamental equities
Ok, my mistake. Not a chance was the wrong way to put it.

The point is hedge funds have no fundamental link to algorithmic trading or requiring low level GPU optimization as a core competency.

Some of them do, and maybe the number is growing, but it’s a subset, not something that defines what a hedge fund has to be.

I wonder how long it took to compile all of this. Great work!