Any changes planned for the larger Prescott cache?

Digital Concepts · 2004-02-21, 02:23

Over in the hardware forum I found two threads talking about cache...

Larger Prescott Cache = Speed Improvement? and P4 Prescott - 31 Stage Pipeline ? Bad news for Prime95?.

But the question still remains. Will modding the client to use a larger cache space afforded by Prescott provide much of a gain? Without said mod, I'm not seeing much advantage to an upgrade (actually maybe a bit worse).

Notice below, the current client doesn't properly recognize the cache size for the Prescott!

3.361GHz P4E (Prescott) vs 3.000Ghz P4C

Intel(R) Pentium(R) 4 CPU 2.80GHz
CPU speed: 3361.04 MHz
CPU features: RDTSC, CMOV, PREFETCH, MMX, SSE, SSE2
L1 cache size: unknown
L2 cache size: 1024 KB
L1 cache line size: unknown
L2 cache line size: 128 bytes
TLBS: 64
Prime95 version 23.8, RdtscTiming=1
Best time for 384K FFT length: 12.593 ms.
Best time for 448K FFT length: 15.319 ms.
Best time for 512K FFT length: 17.278 ms.
Best time for 640K FFT length: 20.659 ms.
Best time for 768K FFT length: 24.788 ms.
Best time for 896K FFT length: 29.935 ms.
Best time for 1024K FFT length: 33.392 ms.
Best time for 1280K FFT length: 44.390 ms.
Best time for 1536K FFT length: 53.239 ms.
Best time for 1792K FFT length: 63.824 ms.
Best time for 2048K FFT length: 71.640 ms.

Intel(R) Pentium(R) 4 CPU 2.40GHz
CPU speed: 3000.23 MHz
CPU features: RDTSC, CMOV, PREFETCH, MMX, SSE, SSE2
L1 cache size: 8 KB
L2 cache size: 512 KB
L1 cache line size: 64 bytes
L2 cache line size: 128 bytes
TLBS: 64
Prime95 version 23.7, RdtscTiming=1
Best time for 384K FFT length: 12.224 ms.
Best time for 448K FFT length: 14.527 ms.
Best time for 512K FFT length: 16.582 ms.
Best time for 640K FFT length: 19.824 ms.
Best time for 768K FFT length: 24.127 ms.
Best time for 896K FFT length: 28.491 ms.
Best time for 1024K FFT length: 32.009 ms.
Best time for 1280K FFT length: 42.270 ms.
Best time for 1536K FFT length: 51.663 ms.
Best time for 1792K FFT length: 61.227 ms.
Best time for 2048K FFT length: 69.714 ms.

ColdFury · 2004-02-21, 03:07

Quote:

Notice below, the current client doesn't properly recognize the cache size for the Prescott!

I believe this is because Prime95 just looks up the processor type and retrieves all the cache info from a table. Since it doesn't recognize the processor, it doesn't have a table entry for it.

If it's defaulting to a code path that's non-optimal, that could be a cause for some lack of performance.

PrimeCruncher · 2004-02-21, 16:01

Quote:

Originally Posted by ColdFury

If it's defaulting to a code path that's non-optimal, that could be a cause for some lack of performance.

Some lack!

Quote:

Originally Posted by Digital Concepts

Best time for 2048K FFT length: 71.640 ms.
...
Best time for 2048K FFT length: 69.714 ms.

Prime95 is 2ms SLOWER on the Prescott, which happens to be running 300 MHz FASTER!

Prime95 · 2004-02-21, 18:14

Prime95 does not care about the L1 cache size, but I'll look into why it is not recognizing the CPUID return code.

There is code in prime95 to use a 1MB L2 cache. It is in theory a little faster for some FFT sizes, but not proven. Please post benchmarks with and without "CpuL2CacheSize=512" in local.ini. Thanks.

Dresdenboy · 2004-02-22, 17:41

I think, some of the reasons for Prescott's performance are that the latencies for many instructions (including SSE2) and cache accesses increased.

Intel's main goals in developing this CPU were different to what many expected. It was not meant to increase per clock performance but to allow higher clock frequencies. Besides this the CPU will have a positive effect on Intel due to production using 300mm wafers and the 90nm node, which allows smaller dice (even with larger cache). It will be hard to modify a heavily optimized app to reach the same IPC on Prescott if key performance factors like latencies changed that much that they cause a severe slowdown compared to Northwood.

As I wrote in another thread, one advantage for Prescott will be clock speed. Near the end of this year the fastest sold Prescotts will outperform any Northwood ever sold in Prime95.

Citing the CPU-car-analogy: Prescott is like moving into the next gear for Intel.

TauCeti · 2004-02-24, 13:56

It's only slightly off-topic here, but i strongly recommend to anyone interested in chip-design to watch this very interesting 90 minute talk from Bob Colwell, former Intel chief architect of the IA32.

GSV3MiaC · 2004-03-04, 23:58

Quote:

Originally Posted by Dresdenboy

Citing the CPU-car-analogy: Prescott is like moving into the next gear for Intel.

Yes, but it's the next =lowest= gear - i.e. when they get the engine revving to 3.6Ghz, they may actually achieve the same roadspeed as they had before.

Dresdenboy · 2004-03-05, 07:54

Quote:

Originally Posted by GSV3MiaC

Yes, but it's the next =lowest= gear - i.e. when they get the engine revving to 3.6Ghz, they may actually achieve the same roadspeed as they had before.

I thought of the amount of gas vs. acceleration relation

QuintLeo · 2004-03-06, 06:54

Quote:

Originally Posted by Dresdenboy

I thought of the amount of gas vs. acceleration relation

Oh, you've been seeing the "heat problem" reports too, eh?

2004-02-21, 02:23	#1
Digital Concepts Aug 2002 54₁₀ Posts	Any changes planned for the larger Prescott cache? Over in the hardware forum I found two threads talking about cache... Larger Prescott Cache = Speed Improvement? and P4 Prescott - 31 Stage Pipeline ? Bad news for Prime95?. But the question still remains. Will modding the client to use a larger cache space afforded by Prescott provide much of a gain? Without said mod, I'm not seeing much advantage to an upgrade (actually maybe a bit worse). Notice below, the current client doesn't properly recognize the cache size for the Prescott! 3.361GHz P4E (Prescott) vs 3.000Ghz P4C Intel(R) Pentium(R) 4 CPU 2.80GHz CPU speed: 3361.04 MHz CPU features: RDTSC, CMOV, PREFETCH, MMX, SSE, SSE2 L1 cache size: unknown L2 cache size: 1024 KB L1 cache line size: unknown L2 cache line size: 128 bytes TLBS: 64 Prime95 version 23.8, RdtscTiming=1 Best time for 384K FFT length: 12.593 ms. Best time for 448K FFT length: 15.319 ms. Best time for 512K FFT length: 17.278 ms. Best time for 640K FFT length: 20.659 ms. Best time for 768K FFT length: 24.788 ms. Best time for 896K FFT length: 29.935 ms. Best time for 1024K FFT length: 33.392 ms. Best time for 1280K FFT length: 44.390 ms. Best time for 1536K FFT length: 53.239 ms. Best time for 1792K FFT length: 63.824 ms. Best time for 2048K FFT length: 71.640 ms. Intel(R) Pentium(R) 4 CPU 2.40GHz CPU speed: 3000.23 MHz CPU features: RDTSC, CMOV, PREFETCH, MMX, SSE, SSE2 L1 cache size: 8 KB L2 cache size: 512 KB L1 cache line size: 64 bytes L2 cache line size: 128 bytes TLBS: 64 Prime95 version 23.7, RdtscTiming=1 Best time for 384K FFT length: 12.224 ms. Best time for 448K FFT length: 14.527 ms. Best time for 512K FFT length: 16.582 ms. Best time for 640K FFT length: 19.824 ms. Best time for 768K FFT length: 24.127 ms. Best time for 896K FFT length: 28.491 ms. Best time for 1024K FFT length: 32.009 ms. Best time for 1280K FFT length: 42.270 ms. Best time for 1536K FFT length: 51.663 ms. Best time for 1792K FFT length: 61.227 ms. Best time for 2048K FFT length: 69.714 ms.

2004-02-21, 18:14	#4
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 16546₈ Posts	Prime95 does not care about the L1 cache size, but I'll look into why it is not recognizing the CPUID return code. There is code in prime95 to use a 1MB L2 cache. It is in theory a little faster for some FFT sizes, but not proven. Please post benchmarks with and without "CpuL2CacheSize=512" in local.ini. Thanks. Last fiddled with by Prime95 on 2004-02-21 at 18:15

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Planned Abscesses	davar55	Game 3 - ♚♛♝♞♜♟ - Morphy's Maniacs	3	2015-02-19 09:24
Planned downtime on GB servers	mdettweiler	No Prime Left Behind	35	2009-11-22 01:50
larger L2 cache, slower iterations?	ixfd64	Hardware	3	2008-05-19 20:46
Prescott	ET_	Hardware	6	2004-07-06 02:38
Larger Prescott Cache = Speed Improvement?	ColdFury	Hardware	7	2003-10-12 16:43

2004-02-22, 17:41	#5
Dresdenboy Apr 2003 Berlin, Germany 19² Posts	I think, some of the reasons for Prescott's performance are that the latencies for many instructions (including SSE2) and cache accesses increased. Intel's main goals in developing this CPU were different to what many expected. It was not meant to increase per clock performance but to allow higher clock frequencies. Besides this the CPU will have a positive effect on Intel due to production using 300mm wafers and the 90nm node, which allows smaller dice (even with larger cache). It will be hard to modify a heavily optimized app to reach the same IPC on Prescott if key performance factors like latencies changed that much that they cause a severe slowdown compared to Northwood. As I wrote in another thread, one advantage for Prescott will be clock speed. Near the end of this year the fastest sold Prescotts will outperform any Northwood ever sold in Prime95. Citing the CPU-car-analogy: Prescott is like moving into the next gear for Intel.

2004-02-24, 13:56	#6
TauCeti Mar 2003 Braunschweig, Germany 2×113 Posts	It's only slightly off-topic here, but i strongly recommend to anyone interested in chip-design to watch this very interesting 90 minute talk from Bob Colwell, former Intel chief architect of the IA32.