Athlon vs P4 - How Do They Scale?
There's been a bit of speculation, at least by a few people on this thread, about the issue, so when I had to spend a several hours on a train recently I decided to take a closer look at the issue.
The results were quite clear:
Northwood scales better with frequency than Palomino.
In games, 3D rendering, synthetic benchmarks… you name it, Northwood outscales Palomino. Sometimes it's by a huge margin, sometimes close enough to argue a tie. In order to spoil the fun in the most efficient way possible, here are the conclusions for each test now…
Quake3, NV15 demo: Northwood scales MUCH better than Palomino.
Video-Encoding MPEG-4: The Northwood core quite simply scales better than the AthlonXP.
SiSoft Sandra 2002 Pro Multimedia: Northwood consistently scales slightly better than AthlonXP.
3D Studio Max: no question about which one scales better here: Northwood.
SysMark 2002 (overall score): Northwood wins again
Unreal Performance Test 2002: Northwood looks like it scales better.
Return To Castle Wolfenstein: Northwood wins the scale-race.
The bottom line:
AMD doesn't have to just keep up in QuantiHerz - serious improvements are needed in order to stay with Intel performance.
Ok, enough with the conclusions - let's look at the data.
First, however, let's just quickly agree on what me mean by "scaling" in this context…
The increase in performance obtained by increasing the operating frequency of the CPU by a given amount. I.e. d P / d f, the slope of the performance vs frequency plot (between the relevant data point and the one before it). For AMD we should then proceed to scale this by 66/100 to get the "quantispeed scaling" (d P / d Qf), i.e. the increase in performance obtained by increasing the Quantispeed rating by a given amount.
The units for "scaling" thus become performance per (quanti)frequency.
The interesting scaling numbers are, of course, those at the top of the (quanti)frequencies currently available, but data for all applicable frequencies are given.
One last thing to note before we begin, is that it is extremely important for the benchmarks scores to be completely accurate in order to obtain meaningful slopes (d P / d f). For example, if the score at one frequency point is depressed slightly, this will increase the slope to the next (higher) frequency point. Plots of score vs frequency over a wide range (of frequency) can expose clear errors in the data (whether due to typos, poor testing methodology or whatever) and I've commented on this in the appropriate places.
The tables below contain the following entries: Qf (QuantiSpeed Rating), f (frequency), score (frames per second or per hour, or other score, as appropriate), D P / D f (average slope), d P / d f (this is the scaling value!), and d P / d Qf (this is the quanti-scaling value, with Qf measured in Quantigigaherz!).
For the more graphically minded, the numbers can easily be entered into a favorite spreadsheet and a bit of regression analysis conducted where appropriate.
A recent review at Tom's Hardware had the most data points of any reviews / articles I could find. The actual benchmark scores below are taken from: www6.tomshardware.com Quake3, NV15Demo
AthlonXP
Qf f(GHz) fps DP/Df dP/df dP/dQf
2100+ 1.733 62.9 36.30 28.36 19.00¹
2000+ 1.666 61.0 36.61 15.15 10.00
1900+ 1.600 60.0 37.50 17.91 12.00
1800+ 1.533 58.8 38.36 22.39 15.00
1700+ 1.466 57.3 39.09 24.24 16.00
1600+ 1.400 55.7 39.79 25.37 17.00
1500+ 1.333 54.0 40.51 n/a n/a ¹ The 2100+ score seems rather odd. The increase in performance for the 100 Quantivalue jump is almost twice that of the previous 100 value jump. This just doesn't make any sense. A good bet at an explanation of this can be found in the THG article where the data comes from. In the introduction, there is a bit about a BIOS upgrade required in order for the system to recognize the 2100+. The reviewer claims that this did not affect performance, but this appears to be incorrect. There are several other examples of the 2100+ scores in this review being well above expectations. Since even a minor increase in performance for a single data point renders the scaling estimate invalid, the 2100+ value will not be considered.Northwood
f(GHz) fps DP/Df dP/df
2.200 60.4 27.45 17.50
2.000 56.9 28.45 18.00
1.800 53.3 29.61 21.00
1.600 49.1 30.69 n/a
Willamette
f(GHz) fps DP/Df dP/df
2.000 52.4 26.20 17.00
1.900 50.7 26.68 13.00
1.800 49.4 27.44 22.00
1.700 47.2 27.76 18.00
1.600 45.4 28.38 20.00
1.500 43.4 28.93 16.00
1.400 41.8 29.86 n/a Conclusion: Northwood scales MUCH better than Palomino.Video-Encoding MPEG-4: Xmpeg 4.2a and Divx 4.12
AthlonXP
Qf f(GHz) fps DP/Df dP/df dP/dQf
2100+ 1.733 33.95 19.59 12.99 8.70
2000+ 1.666 33.08 19.86 12.58 8.30
1900+ 1.600 32.25 20.16 15.52 10.40
1800+ 1.533 31.21 20.36 14.63 9.80
1700+ 1.466 30.23 20.62 10.45 6.90
1600+ 1.400 29.54 21.10 16.12 10.80
1500+ 1.333 28.46 21.35 n/a n/a
Northwood
f(GHz) fps DP/Df dP/df
2.200 34.86 15.85 11.70
2.000 32.52 16.26 16.00
1.800 29.32 16.29 10.35
1.600 27.25 17.03 n/a
Willamette
f(GHz) fps DP/Df dP/df
2.000 31.72 15.86 10.10
1.900 30.71 16.16 13.90
1.800 29.32 16.29 10.70
1.700 28.25 16.62 13.70
1.600 26.88 16.80 12.50
1.500 25.63 17.09 14.40
1.400 24.19 17.28 n/a Conclusion: The 1700+ AthlonXP and 1.8GHz Northwood scores are both "strange", but the result is still quite clear: The Northwood core quite simply scales better than the AthlonXP. Willamette again included mainly for reference purposes.SiSoft Sandra 2002 Pro Multimedia Bench FLOAT
AthlonXP
Qf f(GHz) score DP/Df dP/df dP/dQf
2100+ 1.733 11123 6.42E+03 6.42E+03 4.30E+03
2000+ 1.666 10693 6.42E+03 6.47E+03 4.27E+03
1900+ 1.600 10266 6.42E+03 6.37E+03 4.27E+03
1800+ 1.533 9839 6.42E+03 6.34E+03 4.25E+03
1700+ 1.466 9414 6.42E+03 6.50E+03 4.29E+03
1600+ 1.400 8985 6.42E+03 6.43E+03 4.31E+03
1500+ 1.333 8554 6.42E+03 n/a n/a
Northwood
f(GHz) score DP/Df dP/df
2.200 10795 4.91E+03 4.89E+03
2.000 9818 4.91E+03 4.93E+03
1.800 8833 4.91E+03 4.87E+03
1.600 7859 4.91E+03 n/a Conclusion: Northwood consistently scales slightly better than AthlonXP. I mainly included this benchmark to demonstrate how utterly pathetic this SiSoft Sandra product is. While I did not include them, the Willamette scores are identical. And while I did not include them either, the INTEGER portion of this benchmark demonstrates the exact same behavior and, as could be expected, the Tbird core produces the exact same scores as the Palomino core (at a given real frequency). All in all, this self-professed "multimedia benchmark" present valid results only if you define multimedia as "doing nothing but transcendentals". There is zero performance increase for larger cache sizes, greater FSB bandwidth or lower memory latency. Pathetic. These guys should be sued for sticking the label "multimedia" on this piece of crap.3D Studio Max 4.2
AthlonXP
Qf f(GHz) fph DP/Df dP/df dP/dQf
2100+ 1.733 23.38 13.49 6.67 4.47
2000+ 1.666 22.93 13.76 6.51 4.30
1900+ 1.600 22.50 14.06 10.18 6.82
1800+ 1.533 21.82 14.23 13.25 8.88
1700+ 1.466 20.93 14.28 12.40 8.19
1600+ 1.400 20.11 14.37 12.84 8.60
1500+ 1.333 19.25 14.44 n/a n/a
Northwood
f(GHz) fph DP/Df dP/df
2.200 21.82 9.92 11.26
2.000 19.57 9.78 10.87
1.800 17.39 9.66 8.70
1.600 15.65 9.78 n/a
Willamette
f(GHz) fph DP/Df dP/df
2.000 18.09 9.05 6.99
1.900 17.39 9.15 6.47
1.800 16.74 9.30 8.15
1.700 15.93 9.37 8.03
1.600 15.13 9.45 10.08
1.500 14.12 9.41 7.35
1.400 13.38 9.56 n/a Conclusion: The last two Northwood scores and the last two AthlonXP scores seem a bit strange (XP drops too much, Northwood increases too much, heh). Still, there's certainly no question about which one scales better here: Northwood. Yet again. And very much better at that. The main factor is presumably cache size, since Willamette scales much more like the AthlonXP.Sysmark 2002 (overall score)
AthlonXP
Qf f(GHz) score DP/Df dP/df dP/dQf
2100+ 1.733 188 108.5 194.0 130.0
2000+ 1.666 175 105.0 15.2 10.0
1900+ 1.600 174 108.8 14.9 10.0
1800+ 1.533 173 112.9 89.6 60.0
1700+ 1.466 167 113.9 30.3 20.0
1600+ 1.400 165 117.9 59.7 40.0
1500+ 1.333 161 120.8 n/a n/a
Northwood
f(GHz) score DP/Df dP/df
2.200 216 98.2 55.0
2.000 205 102.5 105.0
1.800 184 102.2 45.0
1.600 175 109.4
Willamette
f(GHz) score DP/Df dP/df
2.000 183 91.5 60.0
1.900 177 93.2 70.0
1.800 170 94.4 50.0
1.700 165 97.1 50.0
1.600 160 100.0 70.0
1.500 153 102.0 60.0
1.400 147 105.0 n/a Conclusion: Uhmm… here we have one of those "odd" 2100+ scores again ;-). No increase in performance with BIOS upgrade my *ss. No question about the final verdict, however: Northwood wins again. Handily. Again…
Maybe there's better luck for AMD to be found in going with another source for benchmark scores? Anandtech recently had a rather feeble attempt at estimating "CPU Scaling", with rather poor graphs, zero detailed analysis and not a whole lot of conclusions. The benchmark scores presented below are from that article: anandtech.com Unreal Performance Test 2002
AthlonXP
Qf f(GHz) fps DP/Df dP/df dP/dQf
2100+ 1.733 91.1 52.57 19.40 13.00
2000+ 1.666 89.8 53.90 28.57 19.00
1800+ 1.533 86.0 56.10 31.58 21.00
1600+ 1.400 81.8 58.43 n/a n/a
Northwood
f(GHz) fps DP/Df dP/df
2.200 87.4 39.73 23.50
2.000 82.7 41.35 n/a
Willamette
f(GHz) fps DP/Df dP/df
2.000 75.9 37.95 25.00
1.800 70.9 39.39 27.50
1.600 65.4 40.88 n/a Conclusion: Despite the relative few data points, I'd have to say Northwood looks like it scales better again.Return To Castle Wolfenstein
AthlonXP
Qf f(GHz) fps DP/Df dP/df dP/dQf
2100+ 1.733 163.3 94.23 47.76 32.00
2000+ 1.666 160.1 96.10 56.39 37.50
1800+ 1.533 152.6 99.54 61.65 41.00
1600+ 1.400 44.4 103.14 n/a n/a
Northwood
f(GHz) fps DP/Df dP/df
2.200 166.3 75.59 48.50
2.000 156.6 78.30 n/a
Willamette
f(GHz) fps DP/Df dP/df
2.000 148.4 74.20 49.00
1.800 138.6 77.00 55.00
1.600 127.6 79.75 n/a Conclusion: Well, seems pretty clear. Northwood wins the scale-race in this final test.
All in all, I'd have to say things look ugly for AMD - unless Tbred brings serious enhancements to the table. Intel will have higher bandwidth and lower latency with dual DDR chipsets later this year. Both are going to significantly undermine AMD's QuantiSpeed rating. If AMD wants to keep it a "True Performance Initiative", they'd better either have something serious to offer, or adjust the QuantiSpeed formula for new speeds appropriately.
Looking at the difference between the Willamette and Northwood scores, it seems clear that doubling the L2 cache has a very positive effect on scalability (as well as providing a decent bump in absolute performance). Increasing bandwidth doesn't look like it will do the trick: Anandtech (http://www.anandtech.com/cpu/showdoc.html?i=1595&p=12 ) tests indicate that we can expect <5% performance increase in virtually everything - except bandwidth benchmarks, of course.
-fyo |
