AMD Announces new HPC Microprocessor

V

Vertigo Acid

2(-log[H+])4u
Joined
May 31, 2003
Messages
12,410
For all you nerds out there itching to build a supercomputer in your mom's basement, this is the time. AMD has announced its 2nd generation high-performance computing stream processors. Dubbed the FireStream 9170, it manages to pack 500 Gflops onto a single package. Compare this to your beloved Core 2 Duo, pumping out on the order of 2-4 Gflops. With an expected retail price of $1,999, it's quite affordable by HPC standards, and will be available at the beginning of next year.

The FireStream 9170 features up to 500 GFlops, or 500 billion floating point operations per second. AMD's second-generation stream processor is built with a 55 nanometer manufacturing process and consumes less than 150 watts of power. The processor is a single-card product with 2Gbytes of onboard GDDR3 memory to compute large datasets without CPU traffic. Asynchronous direct memory access provides data flow without interrupting the stream processor or CPU.
Click to expand...
 
Wow, wonder how this would work for DC projects like Folding@Home?
 
If this is at all Credible, it would mean being Able to Render Extreemely Large Drawings at Incredible Speeds. Some of the Drawings I work with are Really Massive, and sometimes take Hours to Render, hopefully we won't have to wait too long to see some Benchmarks.

One Last thought, I do Hope it's PCIE 2.0 Compatible!
 
so this all sounds amazing, but what does it all mean? lol i tried reading through the wikipedia info on it, but don't understand enough to grasp what you all are saying. So let me ask you my question and see if you guys can put it in simple person terms :-D

Is this something you would buy to outfit a high end gaming machine and what would the noticable advantages be?
 
so thats what amd has been up to....

dosent the ati HD series use stream processors as well??

more or less amd took ATI's FPU unit used on the HD cards and strapped it onto thier FPUand processors, great move if you ask me, but why $2,000?

i know this is slated to be a server processor, and that amd wants to make thier money back on it, but putting this into the hands of the common consumer would put a big dent in intels side ( not that this tech wont trickel down... )

scrath that

and at a seconed look, this isnt a main processor package, it is an add in card ("The processor is a single-card product with 2Gbytes of onboard GDDR3 memory to compute large datasets without CPU traffic"), looks to me more like a PPU than any thing else...
 
AMD will likely incorporate a similar version into it's video cards.

I, for one, welcome our new HPC overlords. I'd like to remind them that even through the good and the bad, I never mocked or ridiculed them and always hoped this day would arrive.
 
i knew or had a feeling that AMD was up to something with their latest tech and aquisition of ATI.
This is purely a foreshadow of things to come.
i can only expect this technology to get better, and yes trickle down to the enthusiast and consumer market.

If all is marketed and planned well, AMD will make up for lost time and Revenue, and hopefully once again, dominate the CPU world with all its Glory and Magnificince!!

mark my words, i said this before and look whats happenning now................yeehaw!!!!
 
Correct me if I am wrong but doesn't nvidia has already something similar? The Tesla does 500glops as well ,though I have no idea how useful that is either at the moment.
 
The number of people who read this kind of article and automatically think "gaming" really shows that the intent of even posting such a piece is lost on the readership of this site.
 
I would love a video codec conversion utility based on this tech. That would 0wn.
 
meanmodda[H];1031647292 said:
Correct me if I am wrong but doesn't nvidia has already something similar? The Tesla does 500glops as well ,though I have no idea how useful that is either at the moment.
Click to expand...

I'm not sure that the Tesla is double precision.
 
Um it's RV670(ie those HD3800 series cards that aren't out yet) with 2GB of GDDR3 on it. It very well might not even have any display ports on the back of it. The first generation firestream was just a 1900XT with some additional ram and crazy cooling for reliability. Even that one was listed around MSRP $2000 but you could get them for $500 assuming you didn't just buy a 1900XT to use in its place.

This is essentially one of the cards that CTM is designed to run on and more than likely there will be little difference from the graphics cards that get released. It could be used as a physics processor but would be a little on the excessive side. Might as well just get another video card and dedicate it to graphics.
 
Dubbed the FireStream 9170, it manages to pack 500 Gflops onto a single package. Compare this to your beloved Core 2 Duo, pumping out on the order of 2-4 Gflops.

Uh, why would you be comparing a GPU-based stream processor (which can only do parallel computations) to a general-purpose CPU? This isn't apples to oranges, it's fruits to nuts. :rolleyes:

My car is better than your airplane? My girlfriend is better than your dog?
 
uclajd said:
Dubbed the FireStream 9170, it manages to pack 500 Gflops onto a single package. Compare this to your beloved Core 2 Duo, pumping out on the order of 2-4 Gflops.

Uh, why would you be comparing a GPU-based stream processor (which can only do parallel computations) to a general-purpose CPU? This isn't apples to oranges, it's fruits to nuts. :rolleyes:

My car is better than your airplane? My girlfriend is better than your dog?
Click to expand...

The point of the comparison was simply to illustrate the relative operation speed. 500 Gigaflops doesn't mean much to the average reader. It's like saying "My airplane flies this fast. Compared to your car that drives this fast."
 
After sifting through the pile of trash and just flat out ignorant above (Free Clue: This has NOTHING to do with GPUs. So stop saying it does.)

This should be interesting. AMD's positioning themselves interestingly here, because I didn't see any partner announcement, as there is with Cell - which is what they'll be competing with. This begs the question, who are they getting to pop out the cards? If they get in bed with IBM a little more, they beat the QS21 (Cell/BE) - math says AMD gives me 7 TFLOP in an H-chassis versus 6.4 TFLOP peak with QS21. What's confusing me more is that the MI Cell cards aren't selling worth crap, to be kind, so why is AMD going card over blade? Especially when the only AMD based system that can handle these beasts (H8152) went the way of the Dodo when Iwill got bought? To say nothing of the integration migraines I keep hearing about. Everyone's been going board/system/blade because the cards are just too much trouble.
 
Ethug said:
The number of people who read this kind of article and automatically think "gaming" really shows that the intent of even posting such a piece is lost on the readership of this site.
Click to expand...

AreEss said:
After sifting through the pile of trash and just flat out ignorant above (Free Clue: This has NOTHING to do with GPUs. So stop saying it does.)
Click to expand...

So there is absolutely no possible way that GPU's could ever benefit from improvements from this research? Never? Did AMD sign an official legal document stating they will never apply this research to GPU's?

"We will never learn from our own research and apply it elsewhere, signed AMD."

Is that what you gather when you read the following paragraphs?
The AMD FireStream 9170 uses stream processing alongside the CPU in high-performance environments. Stream processing is a relatively new paradigm to allow highly efficient parallel processing.

In building the FireStream, AMD leveraged the double-precision floating point technology found in graphics processing units. AMD acquired GPU technology in last year's purchase of ATI Technologies.
Click to expand...

The way I see it, if they do decide to mainstream the use of this they will apply it to one of their main consumer-level products, either their CPU's or their GPU's. Is this a faulty conclusion?

Also, since when are we permitted to insult each other in this fashion on this or any forum? :confused:
 
glutto said:
So there is absolutely no possible way that GPU's could ever benefit from improvements from this research? Never? Did AMD sign an official legal document stating they will never apply this research to GPU's?

"We will never learn from our own research and apply it elsewhere, signed AMD."

Is that what you gather when you read the following paragraphs?

The way I see it, if they do decide to mainstream the use of this they will apply it to one of their main consumer-level products, either their CPU's or their GPU's. Is this a faulty conclusion?
Click to expand...

No. What I gather is that the usual pack of kids all looking for the next big money drain in video cards are completely blind to the entire field of HPC - which they are. In fact, I think there are maybe a half dozen people on [H] total who actually work with/in HPC, and know what this is actually for. AMD announcing this has absolutely nothing to do with video cards, and is not likely to have anything to do with graphics cards. Instead, what I read is a bunch of "OMG THIS IS WHAT NEXT GPU WILL BE LIKE" - which is utterly untrue.
It is in fact, a faulty conclusion. A very faulty one. This is purely an HPC product. Some research might be used, but likely is limited to manufacturing process and power management features. The sole purpose of this product is to perform specific HPC tasks, which are not rendering pretty pictures - it is to take very large amounts of very complex data and churn through it very very quickly. We are talking data sets on the order of dozens to hundreds of terabytes each, which need analyzed as quickly as possible. This is not GPU work - actual interest in nVidia's "GPU as HPC" has been absolute nil, because the cost-performance sucks and the performance period sucks compared to other products such as ASIC, PFPGA, and yes, Cell.
 
AreEss said:
stuff
Click to expand...
I understand that this particular product at this point in time has limited but powerful uses in various scientific and engineering fields, but to state that the concepts in hardware will never, ever trickle down into consumer-level products and that we're all idiots seems to be rather short-sighted.

I don't feel any regret saying congrats for the target market, and congrats eventually to everyone else. I'm certain that eventually I'll gleefully make use of this tech in UT5 or UT6 or so.
 
You must log in or register to reply here.
Back
Top