Topic: Pianoteq benchmarks -- performance across various hardware
I just received my tiny/fanless computer to replace my old/noisy laptop. It uses a i5-4220Y (Haswell) processor -- dual core with hyperthreading, 1.6ghz with turbo-boost to 2.0ghz. It supposedly has a TDP of 11.5W but when I plugged in my electric meter, I see it ranges from 8W @ 600mhz idle to 22W @ 2.0ghz max load turbo. I then plugged in my Celeron 2955U laptop (15W rating) to see what kind of electric usage it has with the screen off -- 8W @ 800mhz idle to 14W @ 1.4ghz max load. Excel isn't needed to figure out the power profile -- Haswell U/Y chips use about 1W per 100mhz of operation. Hence, a 1.6ghz Y chip (TDP 11.5W) probably doesn't uses any less power than a 1.6ghz U chip (TDP 15W). It's mostly just Intel marketing with some thermal throttling thrown in in keep the average consumption lower.
Beyond the misleading Intel specifications, this box is working out pretty good. It's small and makes no sound at all which is wonderful when for working at night when everything is absolutely quiet. The entire case is an aluminum block heatsink. Ambient temperatures are 25C now -- sensor data shows the processor at 40C during light loads, 50C under max loads. When it hits 50C, the case is like a slightly hot sauna rock that you can still hold for 5-10 seconds at a time. It did overheat once when I left it running something nonstop for an hour.
As I am working in China, I was able to order this machine with 4GB RAM + 128GB SSD from Taobao (Alibaba) for a whopping total of 1190 CNY ($187 USD).
https://item.taobao.com/item.htm?spm=a2...t=5#detail
If you are in the US, you can order from AliExpress for $239 USD which isn't that bad of a premium for a product shipped from China.
http://www.aliexpress.com/store/product...46485.html
Now onto Pianoteq performance. With a few more boxes now available for testing, I re-ran my Pianoteq benchmark script and normalized it to matchup with the Pianoteq performance index number. For the benchmark, I instead held the foot pedals down and pounded the keyboard in order to generate as many simultaneous voices as possible because a regular MIDI file (real music) does not have enough polyphony to generate consistent results when you go beyond 2 cores.
Here are peak numbers for each machine I tested. Where there is a T in the ghz column, that means that CPU has the ability to turbo-boost beyond the rated speeds. In the cores column, the +HT means extra virtual cores via Intel's hyperthreading.
Score Processor Arch ghz Cores
----- ------------- ------------ ---- -----
60.8 i7-4790K Haswell 4.0T 4+HT
51.2 i5-3450S Ivy Bridge 2.8T 4
45.4 i5-2500 Sandy Bridge 3.3T 4
40.0 FX-8350 BD Vishera 4.1T 8
39.4 i7-920 Nehalem 2.7T 4+HT
37.6 A10-6700 BD Richland 3.7T 4
35.3 Core2 E8400 C2 Wolfdale 3.0 2
32.7 X2-250 Ph2 Regor 3.0 2
31.7 i5-430M Nehalem-C 2.2T 2+HT
31.2 FX-8120 BD Bulldozer 3.1T 8
31.2 Phenom2 1100T Ph2 Thuban 3.2T 6
29.5 Phenom2 940 Ph2 Deneb 3.0 4
29.4 i3-3227U Ivy Bridge 1.9 2+HT
29.2 i5-4220Y Haswell 1.6T 2+HT
27.7 Core2 E6600 C2 Conroe 2.4 2
25.8 Phenom2 940 Ph2 Deneb 3.2 4
24.3 Celeron 2955U Haswell 1.4 2
23.0 X4-645 Ph2 Propus 3.2 4
23.7 Avoton C2750 Bay Trail 2.4T 8
16.9 A4-5000 Kabini 1.5 4
14.6 Pentium T2310 C2 Merom-2M 1.5 2
5.2 E1-1200 Bobcat 1.4 2
5.0 Atom N450 Atom 1.7 1+HT
4.3 Mobile P4 Northwood 2.8 1
From using Pianoteq on a few of these machines, here's what some of these numbers roughly correlate to:
* 5 = 22khz external, 11khz internal, 16 voices, ~8ms latency
* 10 = 22/11/32/8
* 15 = 22/11/64/8
* 20 = 22/11/256/8
* 25 = 44/22/256/8
* 30 = 96/48/256/8
For the next step, I clocked every machine to a near equal frequency of 1.6ghz. With these numbers, you can then extrapolate the performance of the entire range of a processor family.
Score Arch Cores
----- -------------- -----
25.4 Haswell 4+HT
27.1 Haswell 2+HT
24.7 Haswell 2
25.0 Ivy Bridge 2+HT
23.4 Sandy Bridge 4
23.3 Nehalem-C 2+HT
23.0 Nehalem 4+HT
19.5 Core2 Wolfdate 2
18.7 Core2 Conroe 2
17.7 Kabini 4
16.0 Bay Trail 4-8
15.6 BD/Richland 4
15.3 Bay Trail 2
15.1 BD/Bulldozer 8
14.8 BD/Vishera 8
5.7 Bobcat 2
5.0 Atom 1+HT
Clock-for-clock, the top architecture tested was Intel Haswell. (I don't have any Broadwell machines yet.) The 4ghz+ Haswell downclocked ends up being slower than the chips who's native speeds are ~1.6 as high ghz chips often have internal optimizations to favor higher ghz over low. The older iCore chips (Nehalem, Sandy Bridge, Ivy Bridge) come in close behind so they still will work pretty good for Pianoteq (at the cost of power & heat).
To further expore the effects of multi-core and hyperthreading, I picked a few chips, re-ran the tests using the lowest speed possible and then turned off the cores one-by-one.
Score Cores
----- -----
Kabini / A4-5000 800mhz
9.9 4
9.7 3
9.6 2
5.6 1
Bay Trail / C2750 1.3ghz
13.2 4
13.2 3
12.7 2
8.0 1
Haswell / i5-4220Y 600mhz
10.4 2+HT
10.1 2
8.3 1+HT
5.7 1
Ivy Bridge / i3-3227U 800mhz
12.7 2+HT
12.7 2
10.6 1+HT
7.3 1
Nehalem-C / i5-430M 1.2ghz
17.9 2+HT
17.6 2
12.5 1+HT
10.3 1
Going from 1 core to 2 cores across all architectures (except for Bay Trail) adds a 70-75% improvement in performance:
Bay Trail: +60%
Kabini: +70%
Nehalem-C: +70%
Ivy Bridge: +75%
Haswell: +75%
By comparison, going from 1 core to 1 + 1HT (virtual core) is +45% which while less than a full core but it is still much better a poke in the eye:
Nehalem-C: +45%
Ivy Bridge: +45%
Haswell: +45%
After 2 cores, adding more cores (both real & virtual) adds a tiny bit of performance although the effect increases as with slower architectures/clocks.
Bay Trail: +4%
Kabini: +3%
Nehalem-C: +2%
Ivy Bridge: +0%
Haswell: +3%
Speaking of Bay Trail, these processors come up often for use with Pianoteq due to their low cost/power and ubiquity in tablets/compute sticks. For that reason, I ran a few more variations on the Avoton C2750 to simulate other Bay Trail processors running Pianoteq.
Score ghz Cores Equivalent Processor
----- --- ----- --------------------
21.9 2.4 4 Pentium N3540/J2900
21.0 2.4 2 Celeron N2820
19.3 2.0 4 Celeron N2920/J1800
18.4 2.0 2 Celeron N2810
16.0 1.6 4 Pentium N3700, Celeron N2920, Atom Z3975/Z8700
15.3 1.6 2 Celeron N3150/N3050/N2806, Atom Z3460
13.2 1.3 4 Atom Z3740/Z3530/Z8300/Z8700
12.7 1.3 2 Atom Z3680
In terms of matching up theoretical Bay Trail products, I ignored the rated turbo speeds. If you have a fanless tablet or compute stick, turbo is not going to kick in often due to thermal limitations -- especially since Pianoteq will use 2.5 cores on a slow architecture.