New Sandy Bridge Computer Help (Built - WOW!) - Page 3

frmky · 2011-01-11, 03:25

I wouldn't recommend a 465 for a home system. The big question is whether you want a card with 32 or 48 SP's per SM. All the higher-end cards (465, 470, 480, 570, 580) use 32 SP's per SM, while the 460 and presumably the 560 use 48 SP's per SM. Each SM has two schedulers, each of which can feed 16 SPs. For the 32-SP cards, this is perfect. On the 48-SP cards, nVidia uses superscalar execution to keep the additional 16 SP's busy if instructions aren't dependent. On the older 3.0 compiler, this wasn't very common for most code so the 460 ran more like a 224-SP card (with 112 SP's idling) than a 336-SP card. With CUDA 3.2, instruction scheduling for ILP is noticeably better but still not great.

For code development, a 48-SP card might be better since it will allow you to optimize for ILP and measure the results. This won't affect the higher-end 32-SP/SM cards.

The 560 will probably be about 15% faster than the 460 for about the same price point. There are no performance surprises expected, and it should be out sometime near the end of this month. So the bottom line ... is it worth waiting 2-3 weeks for a 15% performance boost?

Kaboom · 2011-01-11, 11:03

Quote:

Originally Posted by Prime95

I can't justify an extra $50 for a Noctua.

Do it. Noctuas are well worth the price, both more silent and better cooling.

Prime95 · 2011-01-11, 14:26

Thanks for the help everybody. All but the video card is on order!

garo · 2011-01-11, 21:45

Quote:

Originally Posted by Kaboom

Do it. Noctuas are well worth the price, both more silent and better cooling.

I'll second that recommendation!

BTW Noctua is also European. An Austrian/Taiwanese collaboration.

ixfd64 · 2011-01-12, 00:26

Quote:

Originally Posted by Prime95

2) Is the power supply good enough to power an Nvidia GPU?

3) Can I build it using the on-chip GPU and add the Nvidia GPU later (they might be cheaper by the time I'm done with AVX programming and ready to try GPU coding)?

Finally jumping on the GPGPU bandwagaon, eh George?

Flatlander · 2011-01-13, 21:25

Quote:

Originally Posted by Prime95

Sandy Bridge CPUs are now available ... I want to build a new computer with this CPU.

I am sooooo tempted to go down the same route. After all, the keyboard on this one is a bit dodgy.

Prime95 · 2011-01-15, 02:42

Newegg was fast as usual. I built it today and it is FAST!

I'm using the Intel CPU cooler now and overclocking it from 3.3 to 3.7GHz. Temps are about 80C. I'll try overclocking more when the non-Intel CPU cooler comes in a week.

Benchmarking a 4M FFT takes just 40 ms. This is compared to 53 ms on my Core i7 running at 3.5 GHz. That's nearly a 25% boost in performance without any new AVX programming. I suspect this is all due to increased memory bandwidth, but much research is required before I know for sure.

moebius · 2011-01-15, 03:49

Quote:

Originally Posted by Prime95

That's nearly a 25% boost in performance without any new AVX programming.

I look forward to the start of the new 8-core AMD 32 nm processor generation "Code Name Zambezi" in this year.

"The results varied depending on the tasks run, but, finally, the Bulldozer processor proved itself to be approximately 50% faster than the Core i7 950."

http://news.softpedia.com/news/AMD-S...s-177958.shtml

Ralf Recker · 2011-01-15, 09:05

Quote:

Originally Posted by Prime95

I suspect this is all due to increased memory bandwidth, but much research is required before I know for sure.

This could be (partially) a result of the improved L3 cache infrastructure (ring bus, cache "boxes", etc). The L3 cache latency dropped to 26 cycles from somwhere between 36 and 43 cycles (Core i7 965 and Core i7 980X).

Quote:

Originally Posted by moebius

I look forward to the start of the new 8-core AMD 32 nm processor generation "Code Name Zambezi" in this year.

"The results varied depending on the tasks run, but, finally, the Bulldozer processor proved itself to be approximately 50% faster than the Core i7 950."

http://news.softpedia.com/news/AMD-S...s-177958.shtml

If they put 8 physical cores against 8 HT cores (4 physical + 4 virtual) that result is no surprise...

henryzz · 2011-01-15, 09:32

Quote:

Originally Posted by Prime95

Newegg was fast as usual. I built it today and it is FAST!

I'm using the Intel CPU cooler now and overclocking it from 3.3 to 3.7GHz. Temps are about 80C. I'll try overclocking more when the non-Intel CPU cooler comes in a week.

Benchmarking a 4M FFT takes just 40 ms. This is compared to 53 ms on my Core i7 running at 3.5 GHz. That's nearly a 25% boost in performance without any new AVX programming. I suspect this is all due to increased memory bandwidth, but much research is required before I know for sure.

Would a same clock speed comparison be possible? Both running at 3.5 with memory at the same speed etc

Flatlander · 2011-01-16, 19:09

If overclocking is done purely by raising the multiplier, would there be any point in me getting fast DDR3? Is that overclockable too in some way?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
So Sandy Bridge Xeons are now launched	fivemack	Hardware	6	2012-03-14 11:27
Overclocking, Sandy Bridge-E : Don't	firejuggler	Hardware	6	2012-03-08 19:38
2 disabled cores in new Sandy Bridge-E :(	stars10250	Hardware	8	2011-11-16 13:55
Sandy Bridge CPU Usage only 50 percent	dmoran	Software	3	2011-06-14 21:21
Sandy Bridge benchmarks are out.	nucleon	Hardware	0	2011-01-04 11:41

2011-01-11, 03:25	#23
frmky Jul 2003 So Cal 2×1,061 Posts	I wouldn't recommend a 465 for a home system. The big question is whether you want a card with 32 or 48 SP's per SM. All the higher-end cards (465, 470, 480, 570, 580) use 32 SP's per SM, while the 460 and presumably the 560 use 48 SP's per SM. Each SM has two schedulers, each of which can feed 16 SPs. For the 32-SP cards, this is perfect. On the 48-SP cards, nVidia uses superscalar execution to keep the additional 16 SP's busy if instructions aren't dependent. On the older 3.0 compiler, this wasn't very common for most code so the 460 ran more like a 224-SP card (with 112 SP's idling) than a 336-SP card. With CUDA 3.2, instruction scheduling for ILP is noticeably better but still not great. For code development, a 48-SP card might be better since it will allow you to optimize for ILP and measure the results. This won't affect the higher-end 32-SP/SM cards. The 560 will probably be about 15% faster than the 460 for about the same price point. There are no performance surprises expected, and it should be out sometime near the end of this month. So the bottom line ... is it worth waiting 2-3 weeks for a 15% performance boost?

2011-01-11, 14:26	#25
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 16567₈ Posts	Thanks for the help everybody. All but the video card is on order!

2011-01-15, 02:42	#29
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 19·397 Posts	Newegg was fast as usual. I built it today and it is FAST! I'm using the Intel CPU cooler now and overclocking it from 3.3 to 3.7GHz. Temps are about 80C. I'll try overclocking more when the non-Intel CPU cooler comes in a week. Benchmarking a 4M FFT takes just 40 ms. This is compared to 53 ms on my Core i7 running at 3.5 GHz. That's nearly a 25% boost in performance without any new AVX programming. I suspect this is all due to increased memory bandwidth, but much research is required before I know for sure.

2011-01-16, 19:09	#33
Flatlander I quite division it "Chris" Feb 2005 England 31×67 Posts	If overclocking is done purely by raising the multiplier, would there be any point in me getting fast DDR3? Is that overclockable too in some way?