Intel released some Penryn numbers. What do they mean?
Well, most of what looks like standard app benchmarking actually is benchmarking with benchmarks souped up with SSE4 processing. Nice if you use the application and the out-in-the-wild app actually uses it, but certainly not representative.
As we’ll see later, there’s a reason for that.
Maybe the best way to get a ballpark figure on how much extra juice a Penryn has is to see if the Intel numbers include any pure CPU benchmarks, benchmarks that scale linearly with CPU power and aren’t affected by other factors like video cards or FSBs or cache and can be reliably adjusted for differing test speeds.
One of the Intel-provided benchmarks does that, the 3DMark ’06 V1.1.0 Pro CPU (score).
Just to prove that, let’s look at some numbers from XBit Labs. It ran the CPU test on two C2Ds, an Allandale running at 3.2Ghz and a Conroe at 3.5GHz.
The Conroe ran at a speed 9.4% faster than the Allandale. It has double the cache of the Allandale, and ran at an FSB 35% greater than the Allandale.
What was the difference in the scores between the two? 9.4%, the same difference as the difference in clock speed. (This also holds true for the AMD processors shown in the comparison.)
If we compare the CPU score from Xbit Lahs’ overclocked Conroe to that of Intel’s Wolfdale, we find that the Wolfdale does about 8.5% better than the Conroe would at the same speed. If you do the same comparison between the Intel-supplied numbers for quad-core, the Penryn does about 7.5% better.
So it’s pretty safe to say that raw CPU performance is about 8% better with 45nm than with 65nm chips.
Of course, raw CPU performance isn’t the same as real-life CPU improvement. If you plug a CPU with 8% more juice into a current system, a very rough rule is that you’ll see a real-life improvement of about 60% of the raw CPU improvment. So that 8% turns into about 5%.
On the other hand, items like additional cache and higher FSBs sometimes improve real-life performance by a handful of percentage points.
Taken all together, it’s probably fair to say as a rough estimate that a Penryn will do about 10% better than a Conroe, clock-for-clock. Obviously not a barnburner, but not bad for what is basically a tweaked process shrink.
Of course, we overclockers aren’t bound by clock limitations. We’re more interested in knowing how much faster a Penryn will be overclocked than a Conroe, and we don’t know that right now. Given that Intel is showing a product with a speed a little less than 15% faster than that of current C2Ds, it’s probably very safe to say an overclocked Penryn will be able to run 10-15% faster than an overclocked Conroe (though initial numbers may prove lower than that due to motherboard FSB limitations).
Could the Penryns end up doing somewhat better than that? Sure. As Hamlet once said, that is the question, and that will matter as much if not more than any clock-for-clock advantage.
But this isn’t just an intramural competition.
The Real Story
If you’ve read the accounts of these Intel numbers, you’re bound to notice that Intel put up an awful lot of SSE4-enhanced benchmarks.
Why would they do that?
Well, the one thing we know for sure about the K10 series is that they heavily beefed up the FPU section, and that gets reflected in the one FP benchmarks we know about.
It doesn’t take a rocket scientist to speculate that this will give the K10s a big boost, perhaps even a sizable advantage in FP over Conroe. It takes little more than an idiot to suppose that if that’s true, that’s just about all Green will be talking about in Tunisia this week.
To counter that, Intel apparently shoved in all the SSE4 instructions possible into Penryn so they could talk about big FP improvements, too.
Of course, the difference between the two is that AMD’s improvements will improve FP with current apps and games, while Intel’s improvements will only show up when those apps and games start using SSE4, which will probably, in general, take more than a little time.
If someone put a gun to your head and demanded, “What will the Intel/AMD balance become?” the safest answer looks like “For integer, Intel, for FP, for a while, AMD.”
In other words, a muddle, which will leave both sides claiming victory.
For overclockers and/or gamers, we’ll have to see what the overclocked performance ends up being and what the overclocker actually does with computer.