Desktop Fermi cards castrated for TESLA marketshare

[anubis]

I should pause to explain the asterisk next to the unexpectedly low estimate for the GF100's double-precision performance. By all rights, in this architecture, double-precision math should happen at half the speed of single-precision, clean and simple. However, Nvidia has made the decision to limit DP performance in the GeForce versions of the GF100 to 64 FMA ops per clock—one fourth of what the chip can do. This is presumably a product positioning decision intended to encourage serious compute customers to purchase a Tesla version of the GPU instead. Double-precision support doesn't appear to be of any use for real-time graphics, and I doubt many serious GPU-computing customers will want the peak DP rates without the ECC memory that the Tesla cards will provide. But a few poor hackers in Eastern Europe are going to be seriously bummed, and this does mean the Radeon HD 5870 will be substantially faster than any GeForce card at double-precision math, at least in terms of peak rates.


Well, this sucks...
It`s like, first youre late to the party AND you dont bring the booze you promised.

TechReport (http://www.techreport.com/articles.x/18332/5)

[W1zzard]

do you even know what double precision is, what it is used for and how that matters to you ?

[anubis]

probably not... If i get such a reaction from you. I just read reduced DP and thought it matters. Well, live, be wrong and learn i guess.

[W1zzard]

sorry i didnt mean it in an offensive way, just getting tired of people who cry about double precision, or marketing people who say it's better than sliced bread

[anubis]

none taken, thats what the forum is for :) so nothing bad in it for gpugrid?

[Strafe]

Whether double precision is important or not, I think NVIDIA is really making a mistake here. It's the enthusiasts who push technology - and by enthusiasts, I mean developers who will buy GeForce cards (not Teslas) to play with. This and Rottenmutt's post about NVIDIA limiting CUDA performance in GeForce drivers is really starting to upset me.

[W1zzard]

so as a developer you can buy geforce and develop on reduced performance for a really low cost, then when going into production you just have to pay a few k extra to gain more performance.

not so bad imo. and makes perfect sense from a business perspective of nvidia, they want to sell pro cards

[ajaidev]

do you even know what double precision is, what it is used for and how that matters to you ?

You need to take a chill pill really... I know the pain nvidia caused you lol


Anyways i use milkyway@home and it uses DP and i am not so sure about Collatz Conjecture but i think it also uses DP not sure tough.

Our 4*5850 can crunch out more than anything nvidia has at the moment and what GF100 could do in the future. Also if i remember correctly my friends 4870 reached speeds of around 140-145 Gflops DP there....

[W1zzard]

You need to take a chill pill really... I know the pain nvidia caused you lol

i don't. care to elaborate?

[Chumbucket843]

the high end part should be launching with 512 sp's. tesla is has 6GB of GDDR5 which consumes ~50 watts.
besides this is a great gpgpu. it has better accuracy than any previous card, it is fast, it has writable cache, concurrent kernels, full ieee support which no other card has.

this doesnt even matter because gpugrid uses single precision

i dont recomend posting charlies articles as he is on a personal vendetta with nvidia so any little dirt about nvidia he can find he will turn into a 1 page article.

[zalbard]

so as a developer you can buy geforce and develop on reduced performance for a really low cost, then when going into production you just have to pay a few k extra to gain more performance.

not so bad imo. and makes perfect sense from a business perspective of nvidia, they want to sell pro cards
I'm with you here.
It makes sense from the business point of view.
Otherwise, if Tesla and GeForce are exactly the same, why pay more?

[skycrane]

DP, thats not the s3xu@lyyyy one is it... lol

alls i know is that you cant use it for Milkyway, or collaz because they both use DP

but if i remember correctly, didnt they tout fermi as being better than 5870 in DP mode? and now they are castrating it

[Chumbucket843]

DP, thats not the s3xu@lyyyy one is it... lol

alls i know is that you cant use it for Milkyway, or collaz because they both use DP

but if i remember correctly, didnt they tout fermi as being better than 5870 in DP mode? and now they are castrating it

its faster, just not geforce so its going to be pricey for fastest dp. real world performance makes a big difference too. most algorithms dont need alu's as much as bandwidth. fermi has caches and shared memory to help this out.

[YukonTrooper]

Que the "Nvidia is the devil, you must buy an ATI card to save your soul" crowd. Business move and nothing more. While I do wish double-precision was more of a forethought with Fermi for a couple of my science applications, I'm hoping the gaming performance will win me over, which is my first priority.

[Strafe]

I'm with you here.
It makes sense from the business point of view.
Otherwise, if Tesla and GeForce are exactly the same, why pay more?
The Tesla's have always had significantly more RAM than their GeForce counterparts. I guess NVIDIA decided they needed to differentiate more. :shrug::(

[zalbard]

The Tesla's have always had significantly more RAM than their GeForce counterparts. I guess NVIDIA decided they needed to differentiate more. :shrug::(
RAM doesn't help in all applications, though, and not all versions have more RAM either.

[Chumbucket843]

RAM doesn't help in all applications, though, and not all versions have more RAM either.

im pretty sure the two versions are 6 and 3GB, 3GB wont be possible on desktop until H2 2010 if hynix delivers 2Gb IC's. it looks like they are going for capacity because other GDDR5 manufacturers have much faster chips.

in HPC more ram is very important. you are either going to speed up current problem sizes or increase the problem size. this is why some things just cant be done today. IBM roadrunner has 50 terabytes of ram and it can take months to solve problems. thats with a petaflop of math power! gpugrid is speeding up current WU's which is good but we need bigger proteins and longer trajectories to really understand how this stuff works.