NVIDIA Fermi: Architecture discussion
- Thread starter Rys
- Start date
5890 can be expected by June/July. Right.So it's second quarter for the high-end.
AMD should have speed binned mature chips for "5890" by then don't you think?
Very interesting year this gonna be. Hope the line of new products won't stop.
If i was the digitimes editor, i would be double checking the writer's sources that they didnt turn GF100 revision A4 into GF104 somehow.
Folowing up previous chiphell thread, cfcnc #92:
Replied post #95 tomsmith:
The above 2 were before the digitimes article appeared.
So if above is to be believed appears likely they are going ahead with A3 for a Telsa release, and there is some possibly an A4 for the GeForce products.
Last edited by a moderator:
Groo The Wanderer
Regular
I don't know, other than being big enough to spec their own hardware for reasons like cost and power, that is exactly what they do. I don't know about Google first hand, but I do know several others that do do exactly that. One major finance related institution that I work with on occasion has clusters of 128 2S machines, specced by the cheapest they can get at the time that meet the RFQ.
When something has an error, they pull it off line, and slap another one in to take it's place. The removed one is tested, and if it is found bad a second time, it is literally thrown out, it isn't worth the time to debug/risk another error. I suspect Google is much the same.
Yes, and that is all in software. Have you ever seen a Google hardware RFQ for mobos? I have seen a couple, and they are not anything like you are hinting. The reqs tend to be power efficiency, cost, and minimum performance levels, in that order. Last I saw one, RAM was much more important than CPU.
-Charlie
Groo The Wanderer
Regular
Why bring drivers into this, he said software, not drivers. I have seen several ECC implementations on GPUs, all of which did their job at the program level. Read this:
http://saahpc.ncsa.illinois.edu/09/sessions/day2/session2/Maruyama_presentation.pdf
Next time, try looking around a little, you will be surprised by what you will find.
-Charlie
Groo The Wanderer
Regular
I know a bit about how Digitimes works, they tend to be a mouthpiece of the Taiwan semi industry more than an investigative journalism outfit. This isn't to say they are not good at what they do, but they don't rival a long term New York Times level digging.
Most of the time, their leaks are meant to send a message in the background from an annoyed party to the annoyer. If you know how to read their tea leaves, you can glean a lot of info. Magee is exceptionally good at this.
This is the long way of saying that I think the article posted on Digitimes is not based on them reading a forum post, they either asked a well connected AIB, or they were told directly by an AIB who was pissed off.
As for an A4, I really doubt they will do one. If it went in to the oven today, they will likely get it back in WW03 (counting the first full week as WW01). Give it a week for thumbs up/down, and then production, and you would be seriously hard pressed to make Q1 at all. That and I don't see what an A4 would buy them at this point, the problems seem to be more something you would need a full respin for.
-Charlie
Sorry - lost the last sentence in the tomsmith quote when i was cutting and pasting. Have added it back now.
What i think he is saying is that doing another tape to fine tune after releasing a product(postnatal = after the birth) is a regular occurence for CPUs and not unusual for GPUs either. The final sentence reiterates his original quote that A3 is not ideal and there would thus be motivation for another revision.
A further revision would have direct costs(i think $Xmillion) and time to market costs that have to be execeeded by the final benefits received, otherwise they wont bother.
Last edited by a moderator:
If yield and clocks are really that problematic as you among others tend to indicate and if there's no new process option just around the corner, then Nvidia will have no choice but to go for A4. Firstly, to deliver a possible "Ultra" or 385 model to match a potential HD 5890 due mid-2010 at the latest. Lastly (and alternatively) to deliver on their redudancy strategy, which in contrast to AMD's relies on die revisions and process maturity to exploit a maximum number of fully functional dies from every wafer, whereas AMD has die-space sitting idle even in the fullest-featured SKUs.
I think (but am not sure) that he implies that if they would go for a respin it would be a full respin. In that case I'd guess that they'd be looking into a Bx whatever rather than "A4".
In any case I don't sense anything reliable yet in the channel whether A3 is good enough to start production or not.
Under the latest Fun Fact of the Week:
http://www.facebook.com/permalink.php?story_fbid=225030055851&id=8409118252
Didn't want to suggest that myself due to ignorance on how feasible it would be but my initial thought when reading the Digitimes piece was that GF104 would essentially be the "refresh" of Fermi/GF100.
Of course, I'd be really interested if somebody could say whether it is feasible within the stated timeframe.
GF104 is the mainstream/performance part like g84, g94.
Why?
Got any examples?
Here's what a Google server looks like - hey, it's a 2 socket box. Of all the companies you mention, Google has by far the cheapest server design philosophy, and none except Amazon (arguably, with ec2) are in the HPC business.
Maybe you should just admit already that you're basically blowing smoke
When was Nocona a state of the art server - half a decade ago?
What makes you think their current servers use the same design.
DK
I have no reason to think that at all. These are just the ones that are documented - as of April 2009, mind you.
The relevant point however is that these are DP Xeons and not desktop CPUs. Do you think that is different now?
