AMD's new chips

[Dresdenboy]

I found someone, who will run the Prime95 25.5a benchmark on a 1.9 GHz Barcelona (maybe 2 GHz if he overclocks a bit).

[fivemack]

Has anyone seen Barcelona chips (2347 or otherwise) at an on-line retailer which actually has stock?

They don't beat the Q6600 as a cost point for pure computrons (the chips are a bit more than a Q6600, and the dual-CPU motherboard costs more than twice as much as a cheap G965 board, so two Barcelonas and one dual-CPU motherboard is more expensive than two Q6600 boards by more than the cost of one case), but given the cost of FBDIMMs a dual-Barcelona system is a bit cheaper than a dual-quad-Xeon for eight cores.

Also a lot slower for prime95 - a Barcelona 1.9GHz core is about equivalent to 0.62 2.4GHz Core2 cores. I haven't seen lasieve4I14e benchmarks on either system.

[grandpascorpion]

FireStream chips:

http://www.pcworld.com/article/id,13...d/article.html

Expensive but exciting.

[petrw1]

From the article:

The 9170 will be priced at US$1,999, making it one of AMD's most expensive products, but it offers up to 500 gigaflops of computing power, according to AMD, or about 100 times the performance of one of its dual core Opterons.

Software applications will need to be modified to take advantage of the FireStream's architecture, and AMD will release a software developer kit along with the chip package for application tuning.

====================================================

If I am understanding the Primenet stats right then 500 Gflops is 1/50 of the entire average Primenet throughput????

So does the above mean that even if I bought one Prime95 couldn't take advantage of it's potential?

[paulunderwood]

I'd take that with a pinch of salt: saying "double floating point" and "500 Gflops" in a terse advertisement or article does not mean 500 Gflops of double floating point

[jasong]

I would love to save up money to buy a Firestream processor. The problems would be (1) I would have to find somebody willing to try to get stuff to work on it, and (2) the much worse problem is when my father finds out that I'm giving someone remote access to a computer on HIS network. Edit: since I don't have the necessary expertise myself, compiling would be better achieved through remote access.

Of course, by the time I have the money saved up, there could be a TON of DC stuff out there that's ready to be run.

*drool*

[ShiningArcanine]

Originally Posted by paulunderwood

I'd take that with a pinch of salt: saying "double floating point" and "500 Gflops" in a terse advertisement or article does not mean 500 Gflops of double floating point

Why is double floating point necessary? Will long integers not suffice?

[paulunderwood]

http://en.wikipedia.org/wiki/Floating_point
http://en.wikipedia.org/wiki/Floatin...floating-point

Double precision floating point requires many more transistors than single precision, to perform the operations on the registers.

[ewmayer]

Originally Posted by ShiningArcanine

Why is double floating point necessary? Will long integers not suffice?

64-bit ints would be every bit as good as double-floats [in fact allow for a few more input bits per word] if one could do the requisite transform via modular arthmetic as fast a double-floating FFT. But due to real-world demands HW manufacturers have thrown a lot more silicon at the FPUs, so one can't [Just look e.g. at how many 64-bit integer MULs these chips can do per cycle, compared to how many DP float MULs]. And either 32-bit floats or ints don't suffice in terms of precision for the large moduli we are processing these days - the input word size would have to be too small to make the higher degree of SIMD parallelism pay off, relative to double-float.