typo, fixed, thanks
NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
http://www.slashgear.com/nvidia-fer...phics-cards-outed-plus-3d-vision-pro-2795561/
http://www.nvidia.com/docs/IO/40049/NV-DS-QUADRO-6000-Jul10-LR-final.pdf
Not as shiny as the puppy we were shown
http://www.nvidia.com/docs/IO/40049/NV-DS-QUADRO-6000-Jul10-LR-final.pdf
Not as shiny as the puppy we were shown
Last edited by a moderator:
I would not be surprised if it was even GF100.
Since 256 cores are impossible with GF104 (48 cores per SM), it has to be GF100.
For that market this might make more sense than a GF104, since the DP throughput should be higher. Makes at least way more sense than those GF100 based GTX465 in the consumer space...
Yes, it's GF100-based according to Damien:
http://www.hardware.fr/news/imprimer/11046/
And now we know how NVIDIA can claim good yields…
aaronspink
Veteran
Quadro 6000: 448 cores @1150Mhz, 6gb 384bit gddr5 @750Mhz... 225w Max board power
I like how Nvidia keeps adding parts and claiming less power.
You can afford much better cooling on a $1000 part to lower Tj. How does it compare to a Fermi? Seems similar to compute oriented cards, roughly a 25W difference...but I think that can be accounted for by the clock changes.
David
Yeah, that could explain things. The difference was huge (like 40w) for gtx480 when heating up (I think it was guru3d that did those videos), so a cooler that keeps it 20 degrees lower would help a lot (and you would be able to hear how powerful this workstation card is
caveman-jim
Regular
He used the clause of his previous employers contract that permits him to reprint his articles after 6mo.
It's actually quite easy and was my first guess: Out of each of the eight SMs, you disable one of the groups which are not DP-capable.
But it's of course not that way.
I'm not sure they have that kind of granularity when disabling units…
They had it at least in GT218, if i am not mistaken.
Unless they've been outright lying to me, you're wrong wrt GF104 at least. And their statement doesn't leave room for interpretation in this case.
edit: Inside each group of 16, they could be looping/combining though
Last edited by a moderator:
http://www.slashgear.com/nvidia-fer...phics-cards-outed-plus-3d-vision-pro-2795561/
http://www.nvidia.com/docs/IO/40049/NV-DS-QUADRO-6000-Jul10-LR-final.pdf
Seems to be doing well enough.
http://hothardware.com/Reviews/NVID...d-5000-Series-Workstation-GPUs-Review/?page=1
It will be more interesting when they release the much cheaper 1000 and 2000 series quadro cards. With this performance (i think its mainly the geometry improvement, as the SPECviewperf models are quite heavy) they could beat even V8800.
It would be interesting to know the sales ratio of under 800$ quadro cards and these several thousand dolar cards.
Well, if you think about it, for cluster GPGPU computing applications, in some instances these cards may be worth it. Consider, for instance, that you're building a cluster, and you want the most computing power for the least cost and power consumption. Naively you might think of going for the more cost-effective parts alone, but a video card isn't able to get anything done alone: it needs a system to be inserted into. And that system will need server-class CPUs, motherboard, ECC memory, etc. So we could easily be talking about $5000+ just for the system.It would be interesting to know the sales ratio of under 800$ quadro cards and these several thousand dolar cards.
Suddenly, if you're really interested in maxing out GPGPU performance, a $2500 extra cost for 40% extra performance over the Quadro 5000 might not seem so bad: you'll end up with fewer nodes in the end, but each node will be substantially more powerful. Ultimately it'll come down to a question of just how much GPGPU performance matters when building the cluster.
Similar threads
- Deleted member 2197
- Architecture and Products
