15 replies to this topic

[gpuchallenge]

Hello All,

Just wanted to let you know about a contest my company is having for Game Developers. The contest is to develop a GPU based grid viewer. We are offering prizes of $5,000, $3,000, $1,000 and two Ipod Touches. One of the primary objectives is to find an appropriate developer or developers that our company would hire full time.

For more details checkout our website, terms and conditions, and FAQ at

http://gpuchallenge.com/


Thanks!!!

[Nils Pipenbrinck]

Do you guys really want a grid-view as described here:

http://en.wikipedia.org/wiki/Grid_view

Implemented using a GPU?

WTF?

[geon]

WTF, indeed. From the website:

Quote

THE CHALLENGE

To develop a DirectX grid viewer program which implements a FAST and auto-sorting multiple column list view user interface control.

MySQL is pretty fast. Why bother with running it on the CPU, other than mental masturbation?

Quote

One of the primary objectives is to find an appropriate developer

Maybe they want to weed out the applicants who would think it's a good idea?

[kusma]

Quote

By submitting your contributions, you are granting GPUCHALLENGE.COM and its affiliated companies the following worldwide, non-exclusive, perpetual, irrevocable, royalty-free, unconditional, fully paid-up rights: (1) to make, have made, use, copy, reproduce, modify, and create derivative works of the contributions, (2) to publicly perform or display, import, broadcast, transmit, distribute (directly and indirectly through multiple tiers), license, offer to sell and sell, rent, lease, or lend copies of the contributions (and derivative works thereof), (3) to sublicense to third parties the foregoing rights, including the right to sublicense to further third parties, and (4) to publish your name or alias in connection with this contest and your contributions.

The full rights to use your code for the possibility to win $5K? Yeah, right...

[Kenneth Gorking]

Sounds like they just want to make some razzle-dazzle to flash at their customers, and throw alot of buzzwords around...

I've been looking around trying to find out what company this is, and I have only found 1 that fits the profile, even their slogan reeks of "we have buzzwords and fancy colors!" :)

[alphadog]

Why such sensitivity in disclosure of who is sponsoring the contest?

I have no problem with the contest itself, though. If you think in MVC pattern terms, they are trying to move the "V" to the card, and leave the "M" and "C" on the CPU, all under DirectX9 for broad compatibility. They seem to be authors of trading software, and so would like to solely offload the "V" to a card such that more CPU crunching can be reserved for trading processing.

Although, I wonder how much time is spent in the "V" part (rendering and refreshing basically) versus the "C" part (where all the business rules, sorting, etc, would be). Presumably, they've run their numbers.

Also, I'm sure there's already a DirectX9 gridview out there already. I'm sure I've seen one...

[Goz]

alphadog said:

Why such sensitivity in disclosure of who is sponsoring the contest?

Cos, tbh, if they put such a task out to contract I could viably ask for 2-4 times that amount to write the code. They are blatantly trying to save money and, proper, taking the piss with their rules.

Each to their own, i guess. For $40K I'll write them a properly working and tested version. Will they pay it? I doubt it and thats where THEY win and everyone who enters the competition loses.

Btw ... It'd be better to write the solution using Cuda or OpenCL. Using DirectX is attacking a screw with a hammer (Wrong tool for the job).

[kusma]

Goz said:

Btw ... It'd be better to write the solution using Cuda or OpenCL. Using DirectX is attacking a screw with a hammer (Wrong tool for the job).

How is DirectX any more wrong than OpenGL? They're basically the same functionality, with different APIs.

Edit: Whops, sorry. I misread OpenCL for OpenGL! Well, they do answer that question in their faq.

Quote

Why not CUDA?

While we are very impressed with CUDA and may look to develop server side technology using CUDA, it is not an appropriate technology for our target audience or this challenge, so therefore we did not ask for its use. If you have suggestions or experience with CUDA please put it on the submission form or feel free to email us directly with your suggestions.

[alphadog]

Goz said:

Cos, tbh, if they put such a task out to contract I could viably ask for 2-4 times that amount to write the code.

That doesn't explain the anonymity to me. They could still be public, even play it up for PR, and many desperate or ignorant people would run in for the sheer fact that it is a contest...

Goz said:

It'd be better to write the solution using Cuda or OpenCL. Using DirectX is attacking a screw with a hammer (Wrong tool for the job).

Why would CUDA be better? There is no gain in parallelizing to GPU hardware the render of a simple 2D grid, is there? Or, do they want a Minority Report-ish grid that's all sizzle, little steak?

Heck, a good, lean GDI+ grid would be enough. Personally, I think that's the fundamental flaw with their contest. I'm assuming they want a "fast offloaded" grid for some trading app. But, without good reqs, most contestant will build a "everything-and-kitchen-sink" DirectX control, complete with a lot of overhead, when a "just right" GDI+ control would probably be just as fast for a dinky grid view.

But, given the paucity of information for their fishing expedition, I could be very wrong...

[gpuchallenge]

Hello Everybody. Thanks for your comments. Just wanted to let everyone know that we made some changes to the contest website including revealing our identity and why we are sponsoring this contest. We also opened up the contest to CUDA after many suggestions and further research. I hope the challenge details are more clear. Looking forward to your submissions.

http://gpuchallenge.com

Thanks!

[gpuchallenge]

Nils Pipenbrinck said:

Do you guys really want a grid-view as described here:

http://en.wikipedia.org/wiki/Grid_view

Implemented using a GPU?

WTF?

What we are looking for is a Market Data Grid. Please look at the updated Challenge details here http://gpuchallenge....ge-details.html

There is a PDF download with the example data and if you choose to register you will be given a link to a download the actual datagen.

Thanks,

The GPU Challenge Team!!!

[QFS]

So basically you're an equity trading firm that has created a contest to get someone to make software (which you will own and license out) and win a small amount of cash in trade.

I think who ever wins that contest is getting ripped, big time.

[.oisyn]

Quote

Winners will be notified by April 1, 2009

Nice date.
"Ha ha, fooled you, you get nothing while we still sell your application as our own. It's been nice doing business with you"

Seriously though, I agree with the rest here. Even if you win, you lose. $5,000 is a ripoff. Unless of course you patent your application first, so when they want to use your app you can ask for big-time royalties. Even thoug I'm against software patents.
.edit: gosh darn, they already thought of that.

Btw, I seriously doubt the usefulness of the application. They say a 10-20ms delay is unacceptable, and they want to speed that up. Thereby totally neglecting the fact that a computer monitor, usually operating at 60Hz, refreshes the screen at 16.667ms intervals. Also, I don't think the actual drawing is a bottleneck here, and GPU's aren't that good in sorting.

[Nils Pipenbrinck]

.oisyn said:

They say a 10-20ms delay is unacceptable, and they want to speed that up. Thereby totally neglecting the fact that a computer monitor, usually operating at 60Hz, refreshes the screen at 16.667ms intervals. Also, I don't think the actual drawing is a bottleneck here, and GPU's aren't that good in sorting.

These guys are high on my crackpot index! Bullshit-Bingo at it's best.


I had a lot of fun so far. To push it even further consider this: Nerve-impulses travel with approx. 10 feet/second (3 meters per second). Let's assume the distance between your cortex and your finger is 80 cm.

It would takes no less than 260 milliseconds between decission and finger movement. Add in perception time and good old thinking and you're up to half a second. In the meantime we've repainted the screen about 30 times, and the data is totally outdated. Oh Fuck! We're screwed!


And just to make sure: Before all the fast typers and musicans chime in and say they can move their fingers faster: Sure you can, but muscle memory is a different thing. It's connected more to the instinct/reflex than the decission. Also we humans can queue up multiple actions in the nerves before they happen, and the fingers just "replay it".

That's how we're able to play 20 notes/second on a guitar or type chars/second on the keyboard. All guiarrists that tried to increase the speed know that there is a speed barrier that can't be broken by thinking but just by repeative practice.

Back to the topic: Since the guys who started the competition say 20ms is unacceptable. This might be an explanation about the current financial crisis at last. Anyway, Let's for a moment pretend that supernatural guys are sitting behind the monitors, they can deal with such a high refresh rate and the problem is not only academic...

How would you approach it? One million updates/second is quite a lot. std::sort won't do it. Thinking out loud, my first approach would look like this:

• We may have a million of changes per second, but we only display the top 50 of the results, so it's not nessesary to sort them all.
  
  
  
• I'd use a near-sorting algorithm to break the O(n log n) barrier. For 10^6 elements the log-part of the time-complexity *is* significant. Radix-Sort may work but will be memory bound for such a large dataset. Pro for Radix-Sort: Stock-symbols are 4 chars long, so we can work them into a 32 bit dword and sort them via 4 radix8 passes in in O(n).
  
  Anyway, near-sorting will get us into the sub-linear time complexity. We don't get sorted data, but we will get "somewhat" sorted data with a known error bound.
  
  
  
  
• Since we only display - let's say - 50 entries at a time, and since we know the error bound of the near-sorting (let's say epsilon=2) we only can grab the chunk of pre-sorted data that includes - even in the worst case - all the 50 lowest elements. For a epsilon=2 that would be 100 elements.
  
  
  
  
• Now we sort these humble 100 elements and display the first 50 of them. Job done! Even bouble-sort would be fast enough for the final sorting step. Good old GDI will be more than fast enough for the display as well.

Without having made a test implementation I'd say a update rate (including sorting and display) of 5ms is realistic. If that's not enough we can run the sort/display thing on one thread and list-update on another. That'll cut the time in half. The last thing I need to do this job fast is a GPU.


Any comments?


Btw - about the basics and the idea of near-sorting wikipedia has something to say:

http://en.wikipedia.org/wiki/Soft_heap

The original paper is linked from the bottom of the page and even contains example-code... (Ain't Bernard Chazelle a genious, guys?)

[gpuchallenge]

QFS said:

So basically you're an equity trading firm that has created a contest to get someone to make software (which you will own and license out) and win a small amount of cash in trade.

I think who ever wins that contest is getting ripped, big time.

Hello everyone. This challenge is merely to prove a concept. Even an extremely polished submission would be no where near a finished product that could be sold. While the task is challenging, we believe it is possible to increase performance using the GPU. If you can do this please register, this is what we are looking for. The primary reason we put up this contest is to recruit a developer with the passion and skill set to create cutting edge trading software.

However, If you feel you would qualify without entering the $10K GPU Challenge, feel free to submit your resume and cover letter to info@gpuchallenge.com

[kusma]

gpuchallenge said:

The primary reason we put up this contest is to recruit a developer with the passion and skill set to create cutting edge trading software.

Then why are you claiming redistribution-rights to the code?