TR Forums

For no good reason at all I toyed a bit with two similar C2D CPUs: an E7400 with 3MB cache, and a Pentium E6600 with 2MB (that's not the C2D E6600). I set them both to 306 x 10 = 3.06GHz, which is an overclock for the E7400's 2.8GHz and a sideclock for the E6600's default 266x11.5.

I benchmarked with WinRAR, 7zip, SuperPI, and AIDA64. They ended up practically the same. Only one specific AIDA test gained +7% on the E7400, maybe due to the cache. But there's one peculiar deficit: the E7400 is -4% in 7zip decompression. Any idea why that might be?

And ignoring RAM, any thoughts on FSB vs multiplier?

PS: Aren't mobos supposed to beep when the P4 12V is not connected? At first I thought the CPU was bad.

PS2: Did you know Intel CPUs might differ in features in the same model and stepping, just with a different sSpec? This E7400 is missing VT-x.

IIRC the Pentium is Penyrn, whereas the C2D is Wolfdale.
Pentium has SSE4.1 but smaller cache, which probably causes the differences. Like 7z is more cache sensitive whereas the AIDA64 test might be using SSE4.1.

They're both Wolfdale. The E7400 has both SSE4.1 and the larger cache, the E6600 has neither.

The 3MB cache is 12-way set associative, while the 2MB cache is 8-way set associative.

I expect this is an edge case where searching through more blocks slows the L2 cache, because there is no improvement in hit-rate to offset this with 7-zip as it's already 100%, while both that and the larger cache improve hit-rate for the AIDA test. I would probably still select the E7400 because the extra cache improves real-world esp when multitasking more than benchmarks.

In general for Core 2 you want to set the FSB as high as possible until it matches your memory speed, so 1600 if you are using DDR2-800 (PC2-6400), particularly if your BIOS has no tRd adjustment. Asynchronous can add quite a bit of latency.

I would at least set it to 400 x 9 = 3.60GHz even if it's a particularly bad example and I had to exceed the maximum safe 1.45v to get there. I've run a E5700 at 4GHz @ 1.5v for 8 years now, and as the E7400 has a sub-$5 replacement cost you aren't risking much, esp with a spare E6600 on hand (if it actually died though I'd likely upgrade to a cheap quad, even a modified S771 Xeon depending on the board). BTW actual max load wattage of a 45nm dual @4GHz is ~100w so any reasonable quad heatsink should work.

Of course this also depends on your mobo--for example the Abit IP35-PRO that was so popular here locks the multiplier if you select one lower than default, so you'd lose Speedstep. In that case you could just settle on something like this:

High FSB improves many things on Core 2 but duals don't need it as much as quads do, even though they are capable of far higher FSB due to only having two things on the bus instead of three.

Interesting. Didn't think of set associativity. Any benchmarks specifically designed to check the effects of associativity?

The RAM is DDR3-1333 (so 666 MHz bus). If an even FSB ratio matters, that would be 333. But doesn't the RAM anyway change its external and internal clocks in lockstep? And I recall something about problems with much higher than 300MHz FSBs on this chipset (G41). Don't remember the details now, maybe has to do with the PCIe clock (which I manually set to 100).

But actually, the goal here wasn't to get the highest clocks. Just thought more memory bandwidth could be useful (though these benchmarks disagree) and without side effects. Less noise and power is better. I think even with multiplier overriding the CPU did SpeedStep, though the higher FSB leads to higher idle clocks due to the lowest multiplier being 6.

G41 is indeed limited to around 340/1360 FSB without going to silly Northbridge voltages, due to the artificial lack of slower straps. 333/1333 is perfectly fine on it and would be the stock, supported FSB with any 6MB cache Wolfdale installed.

At 1333 FSB, the bandwidth between the CPU and the memory controller on the Northbridge is 10.67GB/s. A single channel of DDR3-1333 (PC3-10600) matches this perfectly by itself, and any extra bandwidth there (including adding a 2nd channel) is pretty much wasted unless you are running the Northbridge-mounted IGP (which is GMA X4500), because nothing else can use it so the bottleneck is the FSB. The RAM speed can be set to a fixed fraction faster than the FSB (mobile chipsets also allow slower), but if you force the memory controller to cross clock domains it will introduce latencies. Think of it like a store-and-forward switch which has to "translate" packets from 10/100 to gigabit--you certainly don't gain any performance from having one end gigabit when there is no other traffic, even if the conversion doesn't take much time. With this analogy the 2nd memory channel is equivalent to adding link aggregation somewhere in the middle--that also cannot be an improvement with no other traffic.

I haven't seen any G41 board that locks the multiplier like that but then they tend to not be enthusiasts' boards with severe CPU power limits and few voltage adjustments anyway. It was so low-budget after all that it even used old ICH7 which lacks AHCI, so you have to run any SSD in IDE mode. Despite this the GA-G41M-ES2L is apparently a popular board for BIOS modding--there's a Hackintosh BIOS for it as well as open-source BIOS Libreboot if you don't trust Intel's proprietary firmware blob perhaps having a backdoor for the CIA.

Now that 11 years have passed and even the most highly overclocked Core 2 Quads drawing over 260w are largely matched by dual-core 15w Pentiums and clobbered by 15w i3s, such vintage hardware is nowadays mostly just for fun but still perfectly usable as a backup spare.

The lack of AHCI support was puzzling, and caught me by surprise. There are ICH7 variants that do support it, but not the vanilla variant.

At least below 333MHz there's benefit to dual-channel. Bandwidth improved with FSB = 306 instead of 266. I guess the CPU- and RAM-side clocks are dependent.

Is there any program that can report the actual PCIe clock?

As I explained before, there's no benefit to dual-channel with DDR3 here because it has enough bandwidth in a single channel already to fully saturate the FSB. The bandwidth improvement you observed going from 266 to 306 is entirely because you increased the bottleneck of the FSB itself. Is there a reason you didn't just set it at 333?

And unlike with AGP there has never been a reason for GPU manufacturers to be very flexible with PCIe speeds, so they all start to misbehave by around 105. If you are trying to increase PCIe bandwidth it would've been much better to not choose a PCIe 1.1 chipset (artificially limited again for market segmentation reasons) as like BCLK tinkering, little gain is possible before running into instability. If you just want to know if it's an independent clock then yes, Intel chipsets allowed the Northbridge PCIe clock to be set independently of any FSB speed.
Note the PCIe x16 clock won't affect anything attached to the Southbridge PCIe (which has its own clock and is only PCIe 1.0a)--plus of course if nothing is plugged into the x16 PCIe then it can't be affected by odd PCIe clocks anyway. But then if nothing is plugged in, you could actually benefit from dual-channel DDR3.

I guess what I said up there is conflicting. Dual-channel that's only good for the northbridge IGP is a shame.

Not trying to overclock the PCIe, only make sure it stays at 100MHz. I haven't bookmarked the discussions, but I recall something about >300 FSBs doing PCIe trouble in some cases. Maybe I'm misremembering, and there were other details. Supposedly not only PCIe cards are affected (I do use a gfx card), but also the ICH7.

There's nothing wrong with dual-channel if it lets you install more RAM--if the frequency matches the FSB then there is no latency penalty, you just can't take advantage of the doubled bandwidth. Unfortunately G41 only officially supports 4GB of DDR3 (or 8GB of DDR2)

You may be thinking of those 1066-FSB chipsets such as 945G or G965 which had neither a working PCIe lock nor the correct strap for operation at 1333. Even if there was no PCIe lock on a 1333-FSB chipset, the proper divisor would exist and be set for 100MHz operation at exactly 1333. But all 1333-FSB chipsets for S775 have PCIe locks so this is academic. G41 officially supports 1333 and that wouldn't be the case if it couldn't run PCIe at 100MHz at that speed.

Many years ago, back when C2Q Q6600 first dropped below $20 on eBay I would routinely padmod OEM G41 machines such Dell, HP or Lenovo to run those at 3GHz and it was very rare to find a chip that couldn't run at that speed at default voltage. As you can adjust voltages, a $12 Q6700 or $18.65 modified X5460 (if your motherboard is on this list. The X5470 is still too expensive) would be a better choices

As an aside, running the PCIe bus asynchronously also obviously introduces a latency penalty, but GPUs are apparently insensitive to it (that's why they are tuned for bandwidth above all else). Though this may explain why some types of PCIe devices seem to have trouble on the CPU-connected (or in this case, Northbridge-connected) PCIe lanes.

Why did you have all these OEM machines?

For the same price (and assuming you have suitable mobos), would you take an X5460 over an i5-760?

I routinely repurposed business boxes for children to use, though their parents often seemed to take and use them for business purposes again once they noticed how well they ran with a SSD. Many are still actually in use today, and the Q6600s remain the only overclocked machines I built used in production environments. Of course after that, everything got locked down and OEM BIOSes didn't even let you overclock the "free" 4x multipliers of some later locked processors anyway.

The Nehalem with integrated memory controller is going to be faster in things highly dependent on memory speed, particularly games and compression (ie 7-zip). However despite the architectural changes (monolithic die, etc), the cores in Nehalem aren't any faster than Core 2 per clock so for everything else the Core 2 Xeon will be faster just due to clockspeed. I will just say that practical considerations are likely going to be more important than performance--it's not easy to find the kind of super low-density DDR3 nowadays that either platform takes, while any old DDR2 you can scrounge up will work fine with Core 2.

If the claims for Ice Lake pan out then it will be 3x faster than Core 2 in everything. To put that into perspective, going from P4 in 2005 to Core 2 in 2006 was also a 3x increase in brute power, so because of the lack of competition from AMD for much of that time it's taken 13 years to accomplish the same. I will still build a vintage system today when I can get the parts for free or nearly so (even if I have to component-level repair them), but it's just for fun as most people today aren't going to appreciate a junky old desktop and would be better served with a 6 watt tablet that is also "fast enough."

How do you figure a 3x increase between 2005 and 2006?

I don't think AMD is to blame for the slowdown in CPU progress, not materially. It's physics, and you can't always get technological breakthroughs on demand. It's not only Intel and AMD/GF, but also Samsung, TSMC, and others.

When a die shrink means going from 14nm to 10nm(or whatever the real figures are) instead of 130 to 90 you can blame that on physics. I don't know about the 3x thing, but I think it's pretty clear there's more to the story than that.

Is it really surprising nowadays that Core 2 was 3x as fast as Pentium 4 at the same clock? Look up for example the Sunspider/Kraken benchmark numbers for yourself. When Core 2 came out in 2006, it was an incredible leap in performance, especially considering it was on the same node + all of the compilers were still optimized for Pentium 4 then. It was the biggest Tock ever, and nothing like the ~5-10% per generation we've had ever since.

If you remember the jump being closer to ~30% than 300% (and A64 being ~20% faster than P4), then you may have been a gamer then. Games have other bottlenecks besides computing power most notably GPU, so expecting 3x the FPS is unreasonable. And certain benchmarks such as the ubiquitous Q3A measured mostly memory bandwidth--something Pentium 4 was actually faster at than early Core 2.

My point was that, nowadays a Core 2 or Nehalem or AMD Phenom-to-Bulldozer machine (all of which have fairly similar IPC) may be slow but is still perfectly usable as a backup machine or for checking emails, just as a P4 machine was a dozen years ago. Especially given the mentioned JavaScript benchmarks supposedly simulate browsing performance.

Big jumps in performance for specific tasks are still possible with dedicated hardware and new instructions, such as AVX-512 or AES-NI, but those require software to be aware of and written to take advantage of the new features. By the time such software becomes commonplace the original processor with those features is long obsolete. (same thing with GPUs--ray tracing, anyone?). It's a lot harder to improve general performance though, especially since clockspeed stopped scaing after the 32nm node. It's difficult for anyone to get excited about new x86 hardware when the improvements are so small, so many of the hardware enthusiasts are off following <3w phone SoC ARM hardware. Perhaps that's really the reason why TR is in such dire straits.

I wouldn't characterize CPU performance based on just one specific domain. JavaScript is more popular than AES, but even so.
Based on a wider set of performance figures (from this review) C2D seems twice as fast as its predecessor, which is still impressive and surprising.

C2D E6700 2.66GHz vs Pentium D 920 2.80GHz x0.95:

+ 84% SysMark 2004
+ 65% Winstone 2004
+100% 3dsmax 7
+ 67% DivX 6.1
+105% QT 7.1 H264
+ 86% iTunes 6 MP3

Mean: 84%

Games at 1600x1200 on two Radeon X1900 XTs. They respond surprisingly well:

+ 61% Quake 4
+112% Battlefield 2
+ 98% HL2: Ep 1
+ 74% FEAR
+228% Rise of Legends
+132% TES: Oblivion

Mean: 117%, geomean: 107%

I used to do a ton of 3DMark2001 benchmarking back in the day. It's a great single-threaded CPU test, and yeah...Core 2 is well over twice as fast as Netburst clock-for-clock.

Here's a Core 2 Quad Q6700 overclocked to 4GHz with a GTX 680:


And here's a Pentium D 935, also overclocked to 4GHz. Yeah, it's using a GTX 560, but the 3DMark01 Game Tests are completely CPU limited (the overall score doesn't take any of the synthetic GPU tests into account).


And just for fun, here's my DDR-equipped PIII @ 1575MHz with a 6800GT GPU, managing to outperform a Northwood P4 @ 2.4GHz with a 7800GS GPU!




And some other scores, all under WinXP:
Celeron 2.6GHz (Northwood-128, GeForce 7800GS): 9099 - that's no typo. Netburst with only 128K of L2 really is that slow.
Pentium M Dothan @ 2.72GHz (GTX 560): 43639
Opteron 185 (Socket 939) @ 3.13GHz (GTX 560): 47464
Core Duo (Yonah) @ 3.08GHz (GTX 560): 50329
Core i5 4670K @ 4.5GHz (GTX 680): 117447

setaG_lliB wrote:

Core 2 is well over twice as fast as Netburst clock-for-clock.

But that's quad core versus dual core.

Like I said, 3DMark01 is single-threaded. The Pentium M @ 2.72GHz is a single-core CPU and it's still beating the pants off that 4GHz Pentium D.

Right. Intriguing. I wonder what's the deal there.

meerkt wrote:

Right. Intriguing. I wonder what's the deal there.

The Pentium M is a well designed CPU, and the Pentium 4... isn't?