How does iPad graphics compare to typical desktop Graphics?

nobond

Newcomer
Maybe a dumb question since I know it is very hard to compare two graphics card farily, just want to get a rough idea though.

Let us take the new ipad with sgx543-mp4 as an example, put it straight, how may Ipad3 put together will equivalent to a state-of-art nvidia/ati desktop discrete graphic card, let us take
nvidia gt600 as an example, here i mean rendering capability/computing capability.
 
If we're talking top of the line discrete GPUs, it's many, many, many, many ipads... AFAIR, iPad2 was like 12Gflops or somesuch of computing performance. Not absolutely sure about that, but really low anyhow compared to anything remotely recent GPU-wise. iPad3 is supposed to be double up of iPad2, so say 25Gflops to be generous.

Discrete GPUs are up to several Tflops now, so orders of magnitude higher. Of course, a discrete GPU would burn up that iPad battery in seconds instead of hours though... ;)
 
The Radeon 9700pro was 33.8GFlops, so maybe something like an 9600 might be around 25. Not sure how other parameters line up. But you may consider the mobile space to be roughly what the desktop was a decade ago.
 
Yes, that seems about right. Didn't a DF article show that in most cases, iPad 3 was about 1/5th of a 360 or so at best?

One of the things people forget is that these devices don't have very fast memory either (comparatively).
 
If only there were some kind of website on the Internet, run by people who would be experts in graphics software and hardware, some of whom might even work for important GPU design houses like AMD and IMG, a website dedicated to 3D and beyond…

I bet those guys could come up with a great article about handheld GPUs, how they're designed, how they relate to their desktop counterparts, whether they seem to be catching up to some extent, etc. :)
 
Maybe a dumb question since I know it is very hard to compare two graphics card farily, just want to get a rough idea though.

Let us take the new ipad with sgx543-mp4 as an example, put it straight, how may Ipad3 put together will equivalent to a state-of-art nvidia/ati desktop discrete graphic card, let us take
nvidia gt600 as an example, here i mean rendering capability/computing capability.

A top of the line tablet GPU eats about 100x less power than a top of the line desktop GPU.

So it is about 100x slower.
 
I'ld say around the GeForce 6800 level (though AA doesnt have as much of a hit on the ipad 3, though some shader operations do take a larger hit)
someone needs to write a benchmark that runs on both desktop/mobile and then you could compare
 
A top of the line tablet GPU eats about 100x less power than a top of the line desktop GPU.

So it is about 100x slower.

Theoretical maximum in terms of FLOPs for the SGX543MP4@250MHz is at 36 GFLOPs times 100, gives 3.6 TFLOPs, which isn't out of line for a today's GeForce 6xx desktop GPU. Only other difference would be that the Apple A5X is being manufactured under 45nm while Kepler under 28nm, the first is DX9.0 while the latter DX11.

In order to gain a more fair comparison I'd suggest an at least 28nm small form factor GPU and even better a DX11 compliant one. While a FLOP/mm2 analogy of some kind would also be helpful for something like upcoming Rogue, I'd still say that the =/>100x higher performance would shrink immediately to much more reasonable levels and of course with better metrics for an apples to apples comparison.

A quad cluster Rogue GC6430 (DX11.1) should deliver under 28nm something above 210 GFLOPs, which with a theoretical costant of 3.6 TFLOPs for a today's desktop GPU would give a difference of a factor ~17x.

033_cpu_vs_gpu_GFLOPS.png
 
One of the things people forget is that these devices don't have very fast memory either (comparatively).
iPad3 actually has a very high-performant memory subsystem for a portable device, 2x compared to its predecessor (~12GB/s).

Also, a TBDR-based GPU cuts down a lot on the bandwidth needed, compared to traditional rendering.
 
A quad cluster Rogue GC6430 (DX11.1) should deliver under 28nm something above 210 GFLOPs, which with a theoretical costant of 3.6 TFLOPs for a today's desktop GPU would give a difference of a factor ~17x.

By the time we see Rogue GC6430 in a mass produced commercial device, there will be single discrete GPU's available with far more than 3.6 TFLOPS throughput. So the ratio will probably be closer to 30x or even higher. That said, GFLOP throughput is increasing at a much faster rate on mobile devices than on the PC: http://www.2dayblog.com/images/2012/april/550x-dsc6317575px.jpg
 
iPad3 actually has a very high-performant memory subsystem for a portable device, 2x compared to its predecessor (~12GB/s).

Also, a TBDR-based GPU cuts down a lot on the bandwidth needed, compared to traditional rendering.

True it does relatively very well, I think it was partly thanks to having given the GPU a double bus to memory? I think Anandtech has an article on it.
 
By the time we see Rogue GC6430 in a mass produced commercial device, there will be single discrete GPU's available with far more than 3.6 TFLOPS throughput. So the ratio will probably be closer to 30x or even higher. That said, GFLOP throughput is increasing at a much faster rate on mobile devices than on the PC: http://www.2dayblog.com/images/2012/april/550x-dsc6317575px.jpg

That's true; however investing in DX11 vs. vanilla DX9 needs quite a bit of additional transistors which could had been otherwise invested in more units/higher performance under DX9 for a small form factor device.

From the reverse perspective if the GPU block in A5X would had been DX11 compliant, there's no chance in hell the GPU would had reached today's graphics performance but quite a bit lower.

Besides Series5XT is merely a refresh to the original Series5\SGX product family which was finalized somewhere in 2005 if memory serves well. Mobile devices are increasing by a much faster rate in terms of GFLOP throughput, but they inevitably have also quite different roadmaps since it would had been absurd to have a DX11 small form factor GPU at the same time as DX11 desktop GPUs from a power perspective to start with.

If power consumption would be less relevant in two years from now (for a purely hypothetical >small form factor device) I wouldn't bet that Rogue halts at barely 1 TFLOP. The oversimplified equasion of N TFLOPs for N Watts might be an indication but I'd personally also try to factor die area in it.
 
One of the reasons that Apple was able to dramatically increase GPU GFLOP throughput when moving from the original ipad to ipad 2/ipad 3 is due to a dramatic increase in SoC die size: http://images.anandtech.com/reviews/tablets/apple/ipad3/diesizecomparison.jpg . So a combination of larger die size and architectural improvements are largely responsible (no pun intended) for the ~ 8-16x improvement in GFLOP throughput from ipad to ipad 2/ipad 3. Of course, moving forward we will continue to see this along with improvement from using smaller fabrication processes. Discrete GPU's are more limited in the sense that increases in die size have stagnated, but they are also less limited in the sense that power consumption requirements are far less strict.
 
Last edited by a moderator:
I'm sure you didn't actually mean "8-16x", as that would be 800-1600 percent increase; clearly not in accordance with reality. ;)
 
iPad 2 is generally 2x ~ 4x faster (or 3x ~ 5x as fast) than iPad's GPU in various synthetic benchmarks. The new iPad is roughly 2x as fast as iPad 2. So it's not 8x ~ 16x, but more likely 5x ~ 10x.
 
One of the reasons that Apple was able to dramatically increase GPU GFLOP throughput when moving from the original ipad to ipad 2/ipad 3 is due to a dramatic increase in SoC die size: http://images.anandtech.com/reviews/tablets/apple/ipad3/diesizecomparison.jpg . So a combination of larger die size and architectural improvements are largely responsible (no pun intended) for the ~ 8-16x improvement in GFLOP throughput from ipad to ipad 2/ipad 3. Of course, moving forward we will continue to see this along with improvement from using smaller fabrication processes. Discrete GPU's are more limited in the sense that increases in die size have stagnated, but they are also less limited in the sense that power consumption requirements are far less strict.

That's a ~34% increase in die area under the same manufacturing process for 2x times the GFLOP throughput and not 8x-16x times which sounds more like a typo. I wouldn't call the die area increase "dramatic" either, considering the provided comparison pic verifies a 63% increase between Tegra2 and Tegra3. Apple must have been more than generous with die area when they laid out A5 since its 2*A9 CPUs (+ surrounding logic) capture roughly 1/4th of the entire die are estate of the SoC (roughly as much as the MP2). If you compare it to Tegra2, which also contains a dual A9 (minus some minor differences in implementation) NV managed to use far less die area as the entire SoC at barely 49mm2 is just a tad bigger than the CPU block on A5.

The GPU blocks should be clocked both in A5 as in A5X at 250MHz, meaning MP4 vs. MP2 gives exactly twice the theoretical peak GFLOPs.
 
When I said 8-16x improvement, I was referring specifically to GFLOP GPU throughput relative to the original ipad. The original ipad (with SGX 535) had ~ 2 GFLOPS GPU throughput (assuming that the GPU was clocked at ~ 250MHz). The ipad 2 (with SGX 543MP2) has ~ 16 GFLOPS GPU throughput. The ipad 3 (with SGX 543MP4) has ~ 32 GFLOPS GPU throughput. Of course, real world performance differences (notwithstanding the effect of vsync) can be far different than that: http://images.anandtech.com/graphs/graph4971/41966.png . Now, with respect to die size, there was indeed a dramatic increase in SoC die size when going from ipad to ipad 2. The increase in SoC die size was not quite as dramatic when going from ipad 2 to ipad 3, but still significant because the majority of the die size increase went towards the GPU.
 
Back
Top