[OICAspork]

Originally Posted by Tim Murray

Welp. D:

I think this may be appropriate reading. Don't forget the video!

[Tim Murray]

I think I can pedant myself out of this!

[3vi1]

Originally Posted by Tim Murray

I think I can pedant myself out of this!

pedant = not a verb.

[Lux_]

FH

we have a next generation GPU core which should be faster, easier to install, and has much more accurate science (at first, comparable to the PS3, but hopefully soon, something even more accurate than the current PS3 client).

Related, I guess.

[Andrew Lauritzen]

I spoke to some AMD people at SC07 (come by the RapidMind booth if you're there! ). Some tidbits that I gleaned:

- the card is indeed R670-based - graphics parts to come soon
- the graphics parts will also support double precision
- the FireGL offering will basically be a superset of this card (i.e. R670, 2GB RAM but with display outputs naturally)
- double precision is 50% speed, but no fused MAD (seems in line with the ~40% figure quoted earlier in this thread)

So maybe some hat eating/video recording to come yet?

Anyways I can't stand 100% behind these facts since the AMD people may even have been in error, but they are probably fairly reliable, and nothing too crazy.

[BRiT]

Originally Posted by AndyTX

- the card is indeed R670-based - graphics parts to come soon
- double precision is 50% speed, but no fused MAD (seems in line with the ~40% figure quoted earlier in this thread)

Why is SP so slow then?

[Dave Baumann]

Originally Posted by AndyTX

- double precision is 50% speed, but no fused MAD (seems in line with the ~40% figure quoted earlier in this thread)

The "5'th" special function scalar is actually not invloved in DP calculations.

[Andrew Lauritzen]

Originally Posted by BRiT

Why is SP so slow then?

I don't know much about the hardware, but I think 50% DP seems to imply a fairly full utilization of the silicon in both single and double cases, no?

Originally Posted by Dave Baumann

The "5'th" special function scalar is actually not invloved in DP calculations.

Ah that makes sense The note about no MAD is interesting though as if you're doing heavy MAD stuff (say, interpolation, evaluating berstein polynomials, etc) you're getting 1/4 the speed DP rather than 1/2. Still, it's probably not critical for most code and 50% issue rates for other instructions is pretty good.

Now I'm wondering whether transcendentals are also 50% and if so, are they accurate to more bits in DP, or the same as SP?

[3dilettante]

That puts RV670's DP throughput at less than 25% of the peak FLOPS figure given for the chip in the AMD PDF.
My math earlier was based on peak marketing numbers, which I were going by the FMADD peak.

Actually, the 5th special function ALU being left out leaves my earlier math slightly optimistic when it comes to price/performance and performance/watt in the redundancy case.

[Jawed]

See this document

http://www.cs.berkeley.edu/~samw/pro...ore/sc2007.pdf

for an example of how "useless" x86 (particularly Intel) is, and why "peak" is such a meaningless concept. The applications covered, such as Proteins, FEM and circuit simulation look like reasonable evaluation candidates for GPGPU workloads...

Jawed

[3dilettante]

I've held off on discussing the difference between peak and sustained performance because it would have bloated my initial post with a discussion that is too complex to resolve for a general case.

Because at the time there was no confirmation of RV670's DP output, I only had AMD's peak SP numbers with which to compare to Intel's DP flops per cycle.

I did not think it fair to compare x86 sustained real-world flops to a GPU's unverified peak DP based off its overly optimistic SP FMADD peak.

It's not like we can't think up workloads where GPUs fall on their faces and cases where CPUs get near peak (or cases where the R600 has faltered in comparison to another GPU).
It's not something I wanted to derail the thread with.

My initial point is that a GPU is a clear price/performance winner at SP, even if it is running the same work on two cards.
I then went on to point out that at the (slightly optimistic) estimate of 25% DP throughput, that same tandem arrangement is not so clear a win.

Even if the individual card is an order of magnitude better at SP performance, cutting overall performance of the GPU by quartering its performance and then halving it again by doing tandem work removes that order of magnitude, particularly at the price point given.

Now, if it were a system of gamer RV670s, the price/performance is blasted out of the park, given it would cost an order of magnitude less.

[Geo]

Assuming the bit about DP being exposed on the consumer version as well turns out to be true, I wonder what F@H will do with an X2? Treat it like two separate cards that you have to run separate instances on?

/me starts to get moderately interested in an X2

[Tim Murray]

Originally Posted by Geo

Assuming the bit about DP being exposed on the consumer version as well turns out to be true, I wonder what F@H will do with an X2? Treat it like two separate cards that you have to run separate instances on?

At the risk of eating more hats, I'd be shocked if DP is available on consumer boards. Otherwise, why would you buy a FireStream? There may well be 1GB HD 3870s, and you'd be able to buy, what... at least 7 or 8 of those for the price of one FireStream?

[Dave Baumann]

Originally Posted by 3dilettante

That puts RV670's DP throughput at less than 25% of the peak FLOPS figure given for the chip in the AMD PDF.

Its between 20%-40% dependant on the mix of operations.

[MfA]

Originally Posted by AndyTX

- double precision is 50% speed, but no fused MAD (seems in line with the ~40% figure quoted earlier in this thread)

Dunno how you come to that conclusion ... hell I don't know how you let someone get away with saying "50% speed" to you without saying "Huh?". It's ambiguous to the point of not really meaning anything.

[Anarchist4000]

Originally Posted by Tim Murray

At the risk of eating more hats, I'd be shocked if DP is available on consumer boards. Otherwise, why would you buy a FireStream? There may well be 1GB HD 3870s, and you'd be able to buy, what... at least 7 or 8 of those for the price of one FireStream?

Keep in mind the original Firestream cards (1900XTs) had a MSRP of >$1k but were being sold around normal 1900XT prices. Stick 2GB of RAM, a 5 year warranty, and a vacuum cleaner to it and the price is a little more justifiable.

Plus with folding and other GPGPU apps, which may want to use DP, it doesn't do a whole lot of good if it's not available on any of the consumer cards. Besides HDR with FP64 is the next best thing!

[3dilettante]

Originally Posted by Dave Baumann

Its between 20%-40% dependant on the mix of operations.

With those numbers, the lower bound would make doubling up on firestream cards for redundant execution a weak proposition, while the upper bound would make the scheme incrementally price/performance-competitive at the suggested price.

I'm liking the idea of buying 8 (4 if I'm not buying another node to host the cards) HD3870s, if they are DP-capable as well.

At consumer card prices, a redundant execution scheme would be a worthwhile investment, unless Firestream comes with some very nifty order-of-magnitude exclusives.

Or I could just run 8 of them full-tilt if I could care less about data integrity.

[Dave Baumann]

You have to take efficiency for the particular task you are looking at into account as well. We already have 3rd party measurements for DP FFT calcutions that are higher throughput than that of the theoretical rates of the CPU's you are talking about.

[3dilettante]

I'm not disputing that GPUs can be more efficient.
In the case where we don't buy 2 cards to run the exact same work unit, I'd say it's an across the board win for anything GPUs don't fall on their face running.

It's only in DP with redundant cards that there is more argument.
I agree that Firestream can be a good buy, just not so amazingly good at that pricepoint that buying multiples to run redundant work units isn't a waste.

Other factors such as rack space and the cost of surrounding hardware also go up when a system doubles the number of cards to run the same amount of work, though I've tried to simplify matters to avoid bogging down on the large number of possible configurations.

[Dave Baumann]

Excuse me, why are we buying redundant cards?

[nAo]

Originally Posted by Dave Baumann

Excuse me, why are we buying redundant cards?

cosmic rays ftw.

[3dilettante]

Originally Posted by Dave Baumann

Excuse me, why are we buying redundant cards?

That was brought up earlier in the thread when dealing with very large clusters. Error rates as they are for DRAM on a system with terabytes of RAM would be on the order of one error about every five minutes.

[Dave Baumann]

I would, then, think about the applications for a single add-in card product such as been announced here. Solutions for large clusters are going to be different.

[3dilettante]

If such a demarcation exists or will exist, it would make sense.
Smaller systems would have a wider window between errors and would be less bogged down by the overhead of repeated calculation.

I did not see a statement to the effect that the FireStream 9170 was a single add-in card product.

Considering the rise of cheap clusters in HPC, I interpreted the product's being touted for HPC to indicate it would match that portion of the target market.

[Dave Baumann]

Take a look here:

http://ati.amd.com/technology/stream...-computing.pdf