From Jonathan's Reference Pages
I decided to run some benchmark tests on WinUAE to calculate just how fast an emulated Amiga is. I'm using AIBB, the standard Amiga benchmarking software. The host machine is an Intel Core 2 Duo E7600 @ 3.06GHz, 4GB RAM, an NVIDIA GeForce GTX 260 graphics card, running Windows 7 64-bit. I'm using WinUAE v2.0.1.
The first test is an emulated PAL A600, 1MB chip RAM, and cycle-exact enabled. The results are very close to AIBB's "A600-NF" benchmark, from 0.99 to 1.01 in all tests except LineTest, which gave 1.04.
I wanted to run an A600 with more RAM (chip or fast) for comparison, but as the 7MHz CPU takes a long time to run the test. I decided to skip it for now.
Next I decided to test WinUAE's "fastest possible" setting. I emulated a PAL A1200 with a 68020, Kickstart 3.0, Workbench 3.1, 2MB chip RAM and 8MB fast RAM. WinUAE options "fastest possible", "24-bit addressing" and "more compatible" were enabled. JIT and cycle-exact were disabled.
The emulated A1200 runs the Dhrystone test almost 36 times faster than the benchmark A1200 and over three times faster than the A4000 040/25MHz. It manages similar results in CPU-intensive tests like BeachBall (almost 5 times better than the A4000) and Sort (3.4 times better). However, it's not so good in graphical tests.
WinUAE's emulated A1200 performs terribly in the EllipseTest, only 0.76 of the rate of a normal A1200. The LineTest result is only 1.3 of the stock A1200 and 1.2 of the A4000. The limiting factor seems to be the Amiga's graphics chipset, which is emulated accurately.
Earlier I switched off just-in-time CPU emulation to make a comparison. In the next test I disabled "24-bit addressing" and "more compatible", and enabled JIT.
The speed increase is enormous. The Dhrystone test jumps from just under 36 times the stock A1200 to 239 times faster. The result is 25 times faster than the A4000/25Mhz. On Dhrystones alone, we would estimate the emulated speed as a 68020 at 3,346MHz, or 68040 at 625MHz.
However, the EllipseTest is still rather poor, only 4.58 times the emulated A1200. TGTest is still slower than the A4000. Hypothesizing that this limitation is imposed by emulating the Amiga's graphics hardware, my next idea was to get rid of that entirely.
I began a series of tests using the Picasso96 graphics board drivers, which WinUAE supports. These should take the graphics chipset emulation out of the equation entirely.
First I installed Picasso96, emulated a 32MB graphics card and ran the benchmarks again. There was no increase, as expected since AIBB was still running the benchmarks on a standard PAL screen. In fact, there was a slight decrease, perhaps due to another Windows program using CPU time.
Test 2 ran the benchmarks on a 640x480x8 UAEGFX screen. This saw a significant increase in graphical tests: relative to the A1200 with JIT, AIBB gave 27.81 times better on EllipseTest, 74.82 times on TGTest, and 106.72 times on the LineTest.
For comparison to the A600, this configuration is 492.49 times faster on the Dhrystone test, 361.10 times faster on LineTest, and 804.15 times faster at the BeachBall rendering task.
I wanted to try and improve the result by tweaking the settings, so I ran more Picasso96 tests. Test 3 adds the new rtg.library from WinUAE. Test 6 also runs Workbench on an RTG window, and Test 7 runs the same but fullscreen. Let's not ask what happened to Tests 4 and 5.
The effects are varied and inconclusive. The new library is 3% faster at EllipseTest and 2% slower at TGTest. The RTG Workbench windows and full screen are between 0.81 and 0.88 of the P96 A1200 on the EllipseTest, and the fullscreen version is only 0.27 on the LineTest.
I wanted to see how CPUs other than the 68020 would affect the results. I ran the following emulated configurations through AIBB: 68030, 68040 (no FPU), 68060 (no FPU), 68040 (CPU internal FPU), 68060 (CPU internal FPU). For fair comparison I re-ran the 68020 also. In all cases JIT is enabled and the emulator runs both Workbench and the tests in fullscreen in 1680x1050x32.
Unfortunately, WinUAE guru'd when loading AIBB on the 030 and guru'd on boot with both configurations of the 060. This leaves the 020 and both configurations of 040. The 040 varied by less than 5% in most tests. Even the Flops test gave very close results for the 040 and the 040 with FPU. It could be that the benefit of these CPUs is outweighed by the optimizations available to JIT CPU emulation.
By using WinUAE's JIT feature and Picasso96 support, a mid-range gaming PC from 2009 can emulate an Amiga with over 1,000 times better Integer performance than an A600, 800 times better FPU performance, and 500 times better graphics performance.
We can compare the results to other AIBB benchmark tests. The Blizzard 1260 accelerator running a 68060 at 64MHz has the fastest Dhrystones performance of any real 68k Amiga listed, at 36.26 times a stock A1200. The emulated performs 674.09 times better than the stock A1200!
The highest graphical performers are 15.63 in the EllipseTest (A4000 CyberStorm PPC 68060/50), 11.15 in TGTest (A3000 CyberStorm II 68060/050), 11.51 in LineTest (A4000 CyberStorm II 68040/040) and 15.00 in WritePixel (A2000 WildFire 68060/050). Our emulated Amiga gives 207.99 in EllipseTest, 180.90 in TGTest, 236.17 in WritePixel, and 412.55 in LineTest.
However, all is not lost. The Blizzard 1260 060/64 manages 497.68 on the Savage test, which beats the emulator's score of 359.99 by a margin of about a hundred and forty A1200s.
Bear in mind that benchmarks are only so useful, unless you mainly use your computers to run benchmarking tools. Even so, it's impressive to see Amiga Workbench load in under two seconds, and respond instantly even in in 1080p resolution with a 24-bit background image - faster even than Windows runs itself.
It would be interesting to compare the results of real and emulated Amiga computers to other machines for which a common benchmark can be established.
Roy Longbottom's Dhrystone Benchmarks for PCs gives the Dhrystone results of various Intel processors, converted into MIPS by dividing by 1,757 (the speed of the DEC VAX 11/780, a 1977 computer believed to have a speed very close to 1 MIPS, or one million instructions per second. On this scale, a 7.09MHz emulated Amiga generates 1,734.88 Dhrystones/second or 0.98 MIPS. This is probably quite accurate, as the 68000 CPU is also said to run at 1 MIPS.
The fast emulated 020 Amiga gives 2,170,138.89 Dhrystones or 1,235.14 MIPS, suggesting speed equal to a hypothetical 020/6.17Ghz, 030/3.7GHz, 040/1.1GHz or 060/929Mhz. Wikipedia's benchmarks place this comparable to a Pentium III 500MHz. A high-end modern CPU, the Intel Core i7 Extreme 965EE, is rated at 76,383 MIPS, over sixty times the emulated Amiga and over 75,000 times faster at integer operations than an Amiga 600.
Note that Dhrystones aren't all that useful for measuring 3D performance, which relies on a lot of floating point calculations. 3D games released for both Amiga and another platform tend to have been optimized for the Amiga release, so it's hard to use the two as a comparison.
Page created: 27th January 2010
« Back to Jonathan's Reference Pages | Comments? | About this website