Comparison of phone CPUs and desktop CPUs?

[Ancalagon44]

Mods - not sure if Highly Technical is the right place for this, if not, please move.

Does anyone know if any SPEC benchmarks have ever been run on mobile/cellular phone CPUs? Or any similar cross platform benchmarks?

I'm just curious as to how fast a 1GHz Cortex A9 would be compared to, say, a 1GHz first generation Athlon, or a 1GHz Core i7. I mean, I'm pretty sure that the A9 would be slower, because of its design goals and architectural differences (such as its pipelining and exactly how superscalar it is), but I'm just curious as to how much slower. Could it actually beat a 1GHz Athlon I? I mean I think those Athlon 1's had cache that did not run at full bus speed, if I remember correctly.

Anyone want to hazard a guess even if no benchmarks are available?

 

[Mr. Pedantic]

AFAIK the only one is LINPACK. And there is a huge difference. According to GreeneComputing (who make the linpack for android) the top android phone CPUs are in the region of a quarter of a GFLOPS. This is pretty bad, compared to modern CPUs.

EDIT: The link I gave was pretty bad, but for the range we're talking about it's fairly accurate in my (rough) estimation - my old Athlon64 @ 2.2GHz got 780 MFLOPS on LinX.

 

[oynaz]

So, around the speed of a Pentium 3/Athlon Thunderbird at the same clock speed. Makes sense.

 

[podspi]

That's insane. I remember the first time I used a P3 @ 1ghz. Now we're talking about that kind of power in a telephone!

Tegra3 should have the throughput of a slow C2D, right? You could actually do some real work on those things!

 

[Modelworks]

That's insane. I remember the first time I used a P3 @ 1ghz. Now we're talking about that kind of power in a telephone!

Tegra3 should have the throughput of a slow C2D, right? You could actually do some real work on those things!

The real limitation with the embedded micros is not that the architecture can't do faster speeds it is that the market they target wants low power .
There are engineering designs that can already match a 2.0Ghz quad core in performance, but the problem is it consumes too much power for battery devices. There is no market right now for things like ARM desktops so the chips are not produced. If you try to put something like ARM on the desktop then by the time you add in all the development work and the smaller market you are better choosing intel or amd so embedded sticks to what it is good at. Embedded has its place, there are some dsp chips that are part of larger SOC that can do things a core i7 can't do in real time but that same chip couldn't run windows.

 

[wuliheron]

John Carmack, founder of Id software, suggested within five years handhelds should surpass the current generation of consoles in power. That's something you hold in your hand capable of playing just about any game on the market today including Crysis in HD. In the meantime India is about to begin production of the first $50.oo tablet PC and within about ten years I'd estimate every snot nosed kid will be dragging one around that doubles as their phone, internet, media, and gaming device.

The desktop may not be dying, but its definitely about to be knocked down a peg and relegated to specialty functions.

 

[Ancalagon44]

The other thing I'm curious about is why they went dual core instead of ramping up clock speeds. I mean, I'm assuming that the power draw of 2 x 1GHz cores is higher than 1 x 2 GHz, because of inter connects and extra cache etc. Anyway, I would think a faster single core would make more sense for a phone because mobile phones are still struggling with multitasking, hence they may not take proper advantage of multiple cores. I know why they did it on the desktop - they hit a wall - but does the same wall exist at even these "low" speeds on mobile phones?

I mean unless the power draw is manageable but the heat produced is too high, I dont know.

It is amazing what kind of power they have been able to put into our pockets though. Connect a bluetooth controller of some kind to your phone, and get it to output HDMI to your TV, and you have a gaming console probably more powerful than a PS2 or even a Wii. The LG Optimus 2X can output HDMI, not sure about other phones.

 

[Modelworks]

One of the reasons they went with dual core instead of increasing clock speed is that it was easy to do without much change in heat or power usage. The embedded architecture is designed kind of like lego blocks. Each thing you want in the micro is its own block. There may be a SOC that has room for 4 blocks and each developer can choose if those blocks will be 2 cpu + 1 GPU + 1 hardware java processor while another may choose 1 cpu + 1 GPU + 1 security processor + 1 java processor .
Embedded makes heavy use of dedicated hardware for different needs while desktop cpu tend to try to use the cpu to do those same task. On the desktop we run java on the cpu but on embedded chips they have processors dedicated to running java that leave the main cpu free for other task.

A good example of this is chips used in set top media players. The Sigma Designs SOC runs at 500Mhz for the cpu + a 333Mhz DSP on the same chip. Because of the dedicated hardware it can decode 2 - 1080p video streams in real time and write those to the hard drive while also decoding and playing back another 1080P stream. During all of this the cpu doesn't break 50% usage and the total power consumed by the chip is 14Watts. A desktop would need a quad core or better and would consume 100W+.

 

[ElFenix]

The other thing I'm curious about is why they went dual core instead of ramping up clock speeds. I mean, I'm assuming that the power draw of 2 x 1GHz cores is higher than 1 x 2 GHz, because of inter connects and extra cache etc.

as we learned on the desktop about 10 years ago, the power draw of 2x1ghz is actually lower than the same arch at 1x2 ghz, generally.

especially if you can shut one of the two down when you're not using it.

 

[metafor]

AFAIK the only one is LINPACK. And there is a huge difference. According to GreeneComputing (who make the linpack for android) the top android phone CPUs are in the region of a quarter of a GFLOPS. This is pretty bad, compared to modern CPUs.

EDIT: The link I gave was pretty bad, but for the range we're talking about it's fairly accurate in my (rough) estimation - my old Athlon64 @ 2.2GHz got 780 MFLOPS on LinX.

Linpack for Android (the one you linked) runs on the JVM which defaults to 64-bit DP math and has an incredibly shaitty scheduler.

A more accurate comparison (and even then, not so much) would be to something like the iPad 2 (which gets around 160 MFLOPs). But even that is not using SIMD instructions and (from the difference between iPad 1 and 2) looks to still be using a ton of divide instructions instead of multiply by reciprocal.

Realistically, if one writes optimized code (or uses an optimized math library) that takes advantage of NEON, the FP power of modern cell CPU's would be similar to early Athlons/Pentium 3s for the same frequency.

 

[metafor]

The other thing I'm curious about is why they went dual core instead of ramping up clock speeds. I mean, I'm assuming that the power draw of 2 x 1GHz cores is higher than 1 x 2 GHz, because of inter connects and extra cache etc.

Exactly the opposite. Frequency (for a given architecture) has a direct relationship with core voltage. And power consumption increases with the square of voltage. So a 2GHz core uses way more than 2x the power of a 1GHz core, assuming an architecture that can go up to 2GHz. That's assuming you'll have to increase voltage to get the core to 2GHz compared to running it at 1GHz, which is almost always the case.