NVIDIA Fermi: Architecture discussion

How long do you think it'll take for someone to hack it so I can use a Fermi as a physX accelerator with a 5970 in for the graphics?

 

according to : Nvidia
"New Fermi-based Tesla products deliver the performance of a CPU-based cluster at one-tenth the cost and one-twentieth the power."

"The Tesla C2050 and C2070 products will retail for $2,499 and $3,999 and the Tesla S2050 and S2070 will retail for $12,995 and $18,995. Products will be available in Q2 2010."

Tesla C2050 3GB GDDR5*
Tesla C2070 6GB GDDR5*

* With ECC enabled, memory available to the user will be 2.625GB for an C2050 and 5.25GB for an C2070.
http://hpc.sprinx.cz/tesla/c2070.aspx

 

no what i mean is if the tesla version of a 285 is 500% of the price of a 285
then we can sort of extrapolate that the geforce version of the c2050 will retail for about $500

 

Why extrapolate when some of the members can write down the MSRP for you?

 

"512 cores" is not a feature...

 

Not anymore anyways.

-Charlie

 

They can I thought the price of the geforce version of fermi was unknown at this point
so how much then ?

 

Even for the high end graphics card?

 

You think they're saving some high-end bins for the graphics card over the Tesla?

 

Lol, what? How could you possibly extrapolate Geforce prices based on last generation's price differential between Geforce and Tesla? They are completely independent, not based on some fixed ratio. Btw, this color scheme is really starting to annoy me

 

But they missed the target clocks.
Fermi is capable of 512 cores. No one really cares how they will archive the performance (8x dp increase) of tesla - more cores or higher clocks.

 

For the first and last time:

CPU Flops <> GPU Flops! Its really quite simple. How much time and money are you going to spend to actually get utilization of the GPUs? The CPU code already is there.

Historically the performance improvement required to make a move is 10x.

 

That's little more than a religious argument until #1 you've seen how hard/easy it is to program Fermi and #2 you have hard performance numbers. I don't really get the philosophy that CPUs are tried and true and therefore can't be replaced by something better.

 

#1) harder than cpus. Significantly harder than just maintaining their and others existing code bases.

#2) depends on time frames and workloads.

Sure anything can be replaced, but we've been down this alley before. And before. and before. and before. At the end of the day, Nvidia might have success if they give the boards away, but the odds of actually making significantly more money are slim at best. People buying supers tend to bargain a LOT and generally aren't willing to pay 5K for something they can buy for 1/10th that.

 

Hmmm I thought the whole point of GPUs was much higher perf/$ and perf/watt? You seem to be saying the opposite.

 

This entire conversation is moot without consideration of the software. IMHO, it's just as easy to underutilise a CPU and its memory subsystem as it is a modern GPU.

 

Commercially wise, it is, actually.
'Cause it is something you can brag about on the box of the card, or in fake card press meetings.

"Capable of 512 cores"? What does that mean?

OT: 100 messages, I should have become Member, isn't it?

 

Fermi is only the architecture. They never said that tesla will have a 512 Core GPU. The only real performance number for tesla was the increase in DP over gt200(b).
BTW you can explain why 512@1224MHz is better than 448@1400Mhz.

 

Just keep in mind it "is" tesla.

 

So you don't think they're putting their best foot forward with Tesla either and are holding back the *good* chips for Fermi?