At what point does Athlon FSB O/Cing give diminishing returns?

[aka1nas]

It seems like even on the 400Mhz FSB, Most benchmarks give slighlty lower scores per clock than at 266 or 333Mhz chips. This at least seems so from anands big barton CPU benchmark earlier this year. So I ask all of you extreme overclockers around here: at what point does increasing the Athlon's FSB start giving diminishing returns.

 

[pspada]

Usually when smoke starts pouring out of the system. 😱

 

[Sunny129]

i think the only way for you to really find out at what point increasing the Athlon's fsb will start giving diminished returns is for you to benchmark yourself by slowly increasing the fsb on an Athlon CPU. because otherwise, the only way to do so is to make calculations based on specific multipliers and fsb's, which only results in theoretical values, but values that should hold true for the most part in reality. and of course these values will indicate that the 400mhz fsb CPUs are more efficient, and not the other way around.

until recently, the multipliers on bartons have been unlocked, so performance is increased as fsb is increased. take the 3000+ for example. a 2500+ @ 3000+ speeds is only on a 191fsb (assuming the multiplier has been left at 11), compared to the 3000+'s 200fsb. do the math and the 3000+ should outperform the OCed 2500+, though only slightly. of course the solution to this would be to bring down the multiplier and up the fsb on the 2500+.
10.5 x 200fsb = 2100mhz = 3000+ speeds. so a 2500+ w/ a multiplier of 10.5 and an fsb of 200mhz should be identical to the 3000+. a 2500+ w/ a multiplier of 10 and an fsb of 210hmz will = 2100mhz = 3000+ speeds, but it will have the edge w/ a higher fsb, and therefore perform slightly better. so maybe the difference in benchmark numbers had something to do with fsb and multiplier manipulations? because i can't think of another reason why "even on the 400Mhz FSB, Most benchmarks give slighlty lower scores per clock than at 266 or 333Mhz chips." by the way, could you link me to the article in which you noticed this?

 

[aka1nas]

The only article I found that had benchies from that wide of a range of athlons was this one anand did earlier this year when the first 400Mhz FSB bartons came out. I simply took the scores for various cpus and divided by clock speed, comparing models that I knew were 400Mhz, 333Mhz and 266Mhz FSb etc. This of course can prob be seen more easily by anand's cpu scaling graphs. The 1700+ actually is more "efficient" per clock than most of the others, including the 3000+ barton@400MhzFSB. Interestingly, the 2800+(which is think is 333Mhz at that time) was actually one of the least "efficient" chips. I would hazard a guess that that is because it is hitting the wall so to speak at its 333Mhz FSB bandwith and that the 400Mhz FSB was really neccesary for the barton to continue to scale at that point. I was interested in this because I read an article(I think it may have been from anandtech but it could have been elsewhere) a while back before barton came out where they were trying to extrapolate performance at higher FSB by overclocking the t-bred and it seemed like the athlon didn't really benefit from 400Mhz fsb and in fact performance seemed to decrease. Is there a way one could calculate the bandwith needed by an athlon at a given frequency to more efficiently balance multipliers with FSB overclocking? I would imagine that there are "sweet spots" that you could clock the chip too that might be more efficient than just cranking the FSB as high as it will go so you get the biggest number possible. Does that sound reasonable?

 

[wolverineI]

ive found sweet spots at around 236-242 depending on mobo.benches actually diminish
over this.on m7ncdpro sweet spot is about 220fsb.on soltek benches are better at 248
but applications seem to run slower.

 

[aka1nas]

That's interesting. Also, I wonder if the CPU temp is noticably affected by running at a higher FSB vs the same clock with a higher multi and lower FSB.

 

[Accord99]

Originally posted by: aka1nas
The 1700+ actually is more "efficient" per clock than most of the others, including the 3000+ barton@400MhzFSB. Interestingly, the 2800+(which is think is 333Mhz at that time) was actually one of the least "efficient" chips. I would hazard a guess that that is because it is hitting the wall so to speak at its 333Mhz FSB bandwith and that the 400Mhz FSB was really neccesary for the barton to continue to scale at that point.

Actually this has to do application's depending on several factors for performance, including CPU, memory, video card, etc. Unless the application is completely CPU limited, you'll won't see perfect scaling because the speed of memory, in terms of latency and bandwidth does not increase at the same rate. And if you use the same video card across different CPUs, then you also have the same issue. This is why some of the early comparisons of the P4 which mathematically scaled down their benchmark scores to compare with lower-clocked competitors was flawed, because it also scaled down memory and video card performane that would not have changed with a different speed P4.

 

[Jeff7181]

I doubt you'll ever reach a point at which the scores actually drop because the FSB is too high. The change in efficiency that you pointed out is so small that it may be just part of the margin of error because if you take the XP2100 or the XP2400 for example, they're both less efficient than the XP3000.
The correlation you're making between the loss of efficiency and the FSB isn't there. If you run a 400 Mhz FSB with a multiplier of 8, I can guarantee you 100% that it'll be faster than a 133 Mhz FSB and a multiplier of 12.
The relationship you see here is the result of the rising internal clock speed. If you notice, the XP2800 is a lot less efficient than the XP3000. This probably has something to do with the cache... but if you look at the XP2600 and the XP2700 and the XP2800, you'll notice as clock speed increases, efficiency decreases. That's just the way things are... a 10% increase in clock speed will RARELY increase performance by 10%. That holds true for P4's as well.

 

[aka1nas]

Great points Accord and Jeff. Another thought actually just occured to me, the original article that I read sometime ago may have had other factors that they didn't mention such as RAM timings. Considering that this was long enough ago that barton wasn't out yet and the 333Mhz FSB was brand new, it's reasonable to assume that they had to overclock the ram they used and probably increase the latency to offset that. Maybe that explains some of the performance loss as the RAM timings weren't all that great.

 

[Jeff7181]

Originally posted by: aka1nas
That's interesting. Also, I wonder if the CPU temp is noticably affected by running at a higher FSB vs the same clock with a higher multi and lower FSB.

Very slight difference. A faster FSB means more work is being done, so the CPU gets slightly hotter... but it's hardly measureable with just the sensors on the motherboard.