Looking for advice for a new PC....

[Xyzzy]

…it's funny to say it as a memory pig

In most cases it appears that way because it caches all free memory, like it should. (Why waste "free" memory?)

We've seen it go as low as 2-3MB free, and we don't run any heavy applications. We also have a 4GB ReadyBoost cache as well.

We sure wish we had the 64-bit version. The 32-bit version doesn't see all of our memory.

[ewmayer]

Originally Posted by S485122

The 45nm chips are already here... But at which price ! 1100,00 USD or so.

At the moment AMD cannot deliver their quadcores because of a production bug with the Level 3 cache. I also read some reviews (anandtech I believe) that state that the AMD is not performing as it was hoped (and much slower than Intel products.) Of course there is still some type of work where AMD shines (factoring ans especially with 64 bit programs I believe.)

Nice article here about AMD’s recent woes and near-term prospects here:

Fortune Big Tech: Has Intel Crushed AMD?

Originally Posted by petrw1

2. AMD Quads are closer to 4 times than the Intel Quads.

Truth or Rumor?

Since AMD isn't shipping Barcelona in volume yet, it would have to be classified as "rumor". Also, based on some hot-of-the-presses numbers I just got from my buddies at Sun for my in-dev multithreaded Mlucas code running on 4 and 8-core Intel servers, Intel's quad-core solution looks pretty darn good. And note that I do hope AMD can fix the bugs plaguing Barcelona soon. Healthy competition - it's, well, healthy. But at least in the near-term, Intel's decision to leverage their proven dual-core into a quickie quad and keep doing what they are geniuses at - shrinking the process - appears to have been the right call. Might be a nice opportunity to pick up some AMD stock on the cheap, though. ;)

[tha]

Originally Posted by petrw1

I will probably buy it in the spring.

Requirements:
- FAST! Processor and FAST! Memory
- Probably at least a quad (or are 8's out and affordable now?)
- Prime use will be intensive computing: i.e. GIMPS and related stuff.
Advice???

- as pointed out, quads suffer from a memory bottleneck.
- fast memory lanes and fast processors, really a big increase in speed will be in 2009 with the introduction of a new Intel processor design. Like the Core 2 was a big improvement over the flawed Pentium IV design.
- So you might want to spend your money in 2009 and buy a cheap model in 2008, which you then give to your little brother a year later.

[rx7350]

The Intel NEHALEM architecture will debut in 2009, with an integrated memory controller, including an octo-core cpu.

The memory bottleneck will be somewhat relieved. Maybe the cache sizes will increase to the point where a memory bottleneck is no longer relevant to PRIME95.

[ShiningArcanine]

Originally Posted by S485122

ShiningArcanine, the "CORSAIR ValueSelect 2GB (2 x 1GB) 240-Pin DDR2 SDRAM DDR2 667 (PC2 5300) Desktop Memory" memory you recommend is a no go for a quad core : it is way to slow. (I would even go as far as saying it is to slow for a core2 duo dual core.)

Do not forget the requirement : "- FAST! Processor and FAST! Memory"

With PC2 8500 memory you will get a 60% performance boost. It is more expensive of course. 89,00 USD for the "CORSAIR XMS2 2GB (2 x 1GB) 240-Pin DDR2 SDRAM DDR2 1066 (PC2 8500) Dual Channel Kit Desktop Memory Model TWIN2X2048-8500C5" and 114,00 USD for the "CORSAIR Dominator 2GB (2 x 1GB) 240-Pin DDR2 SDRAM DDR2 1066 (PC2 8500) Dual Channel Kit Desktop Memory Model TWIN2X2048-8500C5D".

And forget everything people are saying at anandtech and so on. They do not use the CPU's like you will. They test memory throughput by loading one instance of their test program... Their conclusion is of course that PC2 6400 memory is good enough.

Jacob

As far as lucas lehmer tests go, there will be no memory fast enough until there is a computer processor with n * 8MB of L2 cache where n is its number of cores.

Then that will be fast enough until the numbers become so big that n instances of Prime95 exceed the storage capabilities of the n * 8MB of L2 cache, at which point we will need n * 16MB of L2 cache.

Originally Posted by Cruelty

Just don't go with integrated GFX as it will occupy some of the memory bandwidth needed by CPU...
Also AFAIR there were some general perfomance issues with integrated chipsets and Prime95.

The integrated graphics in Intel's chipsets have had hardware rendering since the G965 chipset. The motherboard I suggested has the latest integrated graphics chipset from Intel. He could play games like Far Cry on it if he wants and unless intel has released new drivers that offload additional processing to the CPU to raise the frame rates, all of the work is done by a dedicated graphics core.

http://blogs.intel.com/technology/20...d_graphics.php

Originally Posted by IronBits

Wait until spring, or when you are ready to buy and re-ask the question as many things will change between now and then. ;)

That would be the best thing to do.

[sdbardwick]

The primary issue with Prime95 and integrated graphics is the use of main memory for the framebuffer; it can consume a [formerly huge (think PC100 days), now less substantial] portion of the memory bandwidth. The additional CPU cycles needed to render the images remains minor in comparison.

[S485122]

Originally Posted by ShiningArcanine

As far as lucas lehmer tests go, there will be no memory fast enough until there is a computer processor with n * 8MB of L2 cache where n is its number of cores.

I agree.

But considering that :
-with quad cores the performance for LL is proportional to the memory speed (tested with 5400, 6400, 8500 memory),
- the original query was about "FAST!" memory,
the proposals of 5300 or 6400 memory did not meet the request.

8500 memory is not the fastest currently available but it is still affordable, if price was not a problem the answer to the query would have been a DDR3 board with DDR3 1800 (PC3 14400) memory but that kind of memory is priced in the 500,00 USD range.

Jacob

[axn]

Originally Posted by S485122

8500 memory is not the fastest currently available but it is still affordable, if price was not a problem the answer to the query would have been a DDR3 board with DDR3 1800 (PC3 14400) memory but that kind of memory is priced in the 500,00 USD range.

Is there any point in having a memory faster than the FSB? I think low latency might be better than high bandwidth for P95. Anybody out there with some concrete numbers?

[ShiningArcanine]

Originally Posted by sdbardwick

The primary issue with Prime95 and integrated graphics is the use of main memory for the framebuffer; it can consume a [formerly huge (think PC100 days), now less substantial] portion of the memory bandwidth. The additional CPU cycles needed to render the images remains minor in comparison.

The FSB and memory bus speeds run asyncronously on recent Intel motherboards, so the memory bandwidth requirements of the integrated graphics core should be negligible for 2D applications, unless he overclocks the FSB to match the memory bus.

Originally Posted by S485122

I agree.

But considering that :
-with quad cores the performance for LL is proportional to the memory speed (tested with 5400, 6400, 8500 memory),
- the original query was about "FAST!" memory,
the proposals of 5300 or 6400 memory did not meet the request.

8500 memory is not the fastest currently available but it is still affordable, if price was not a problem the answer to the query would have been a DDR3 board with DDR3 1800 (PC3 14400) memory but that kind of memory is priced in the 500,00 USD range.

Jacob

In that case, he should get this motherboard:

http://www.asus.com/products.aspx?l1...29&modelmenu=1


It has the lowest latencies avaliable by nature of having the ICs soldered onto the board, although avaliability will be a problem and he will need a discrete graphics accelerator as Asus does not design its products to cater to the scientific community.

Originally Posted by axn1

Is there any point in having a memory faster than the FSB? I think low latency might be better than high bandwidth for P95. Anybody out there with some concrete numbers?

All of the electrons in a computer travels at a fixed rate, roughly 0.9c. Computers use pulses, that operate at certain frequencies to send data over fixed pathways. Each cycle takes up a certain amount of time and with each cycle, a certain amount of data is sent over these pathways. If you increase the frequencies at which the computer operates, say by a factor of two, the amount of data transfered per cycle will not change, but the number of cycles per given time unit will change and since the number of cycles per given time unit in this example have increased by a factor of two, the time it takes from the moment the first unit of data is received and the last unit of data is received, from a given transmission, is cut by a factor of two. The time it takes between sending the first unit of data and receiving the first unit of data, will however, remain constant. Because of this, latencies go down with higher memory transfer rates, but there is a limit to how much it can go down.

[ewmayer]

Originally Posted by ShiningArcanine

All of the electrons in a computer travels at a fixed rate, roughly 0.9c.

Ah, no - actually it`s the electric field, i.e. the voltage-differential-induced electromagnetic current wave, which propagates at the speed of light [whose value in conducting solids is typically ~2/3 that of its vacuum value] - IIRC the actual *electrons* move [on average] exceedingly slowly - some discussion of that here. To use a macroscopic analogy, a tsunami moves across the ocean at hundreds of km/hour, even though the actual water molecules never move anywhere near that fast, and for the most part [until the wave breaks] simply do a circular motion as the wave passes and wind up where they started. [This is also what happens for AC power, which is why AC rather than DC is used for large-scale power distribution: Tesla understood this, Edison failed to]. In such cases it`s *energy* that propagates at high speed, not particles. [Admittedly, "particle" can't be divorced from "wave" for electrons, but we can still discuss the maximum amplitude of the QM wavefunction as if it were a point particle of mass.]

[ShiningArcanine]

Originally Posted by ewmayer

Ah, no - actually it`s the electric field, i.e. the voltage-differential-induced electromagnetic current wave, which propagates at the speed of light [whose value in conducting solids is typically ~2/3 that of its vacuum value] - IIRC the actual *electrons* move [on average] exceedingly slowly - some discussion of that here. To use a macroscopic analogy, a tsunami moves across the ocean at hundreds of km/hour, even though the actual water molecules never move anywhere near that fast, and for the most part [until the wave breaks] simply do a circular motion as the wave passes and wind up where they started. [This is also what happens for AC power, which is why AC rather than DC is used for large-scale power distribution: Tesla understood this, Edison failed to]. In such cases it`s *energy* that propagates at high speed, not particles. [Admittedly, "particle" can't be divorced from "wave" for electrons, but we can still discuss the maximum amplitude of the QM wavefunction as if it were a point particle of mass.]

Computer processors are DC circuits. I am not an electrical engineering major, but according to the following web page, transistors "cannot switch AC:"

http://www.kpsec.freeuk.com/trancirc.htm

Also, in a DC circuit, electrons are always moving in a current with their average position being the integral of their average velocity while in an AC circuit, their position is a sine function involving time, a frequency and a phase angle, mutipled by some amplitude and their velocity at Pi minus the phase angle plus any mutiple of two Pi being zero. In an AC circuit, talking about currents is meaningless because each electron never leaves a sphere with a radius that is the amplitude of the wave function and an origin that is the point where its velocity is zero.

Despite that, I was wrong. Not all of the electrons in a computer travel at a fixed rate (which is actually an average rate), as some of the electrons in the PSU travel back and forth, with no net change in displacement worth mentioning.

By the way, that is interesting, but it does not answer axn1's question. Getting back to his question, I would like to add some information I omitted:

Low timing memory reduces the amount of time from the memory receiving a request to the memory sending a request, but with DDR2 memory (and I believe even more profoundly with DDR3 memory), many times lower clock rates hurt latencies more than lower timings help.

Also, "All of the electrons in a computer travels at a fixed rate, roughly 0.9c" should have been "All of the electrons in a computer travel at a fixed rate, roughly 0.9c." I apologize for my misuse of the English language.