X86 Tablets CPU's

[Medallish]

Hi, I've been reading on this page forever, and now was the time to register and write about something I feel that this board will be able to not only answer but also write some damn interesting stuff, because that's what you guys seem to do .

So I've had my eyes on the Acer Iconia W500(5W AMD C-50, Ontario Tablet), and I'm actually pretty sure I will buy one very soon, this is despite the fact that I find tablets rather useless, the W500 comes with a keyboard as well that you can dock it in so I feel this is the best way to find if there really are benifits to tablets. Anyway that's the personal stuff, what interests me are the CPU's in this interesting time.

I read this article at FudZilla, that even got Fuad to say something negative about Intel, it interested me that seemingly Intel isn't taking Tablets seriously not only because of the mediocrecy of this offering, but also the sick price.
I was actually writing a comparison and while I was looking into the GPU part of the Oak Trail Z670, I found that it was a powerVR SGX 535 GPU, and yeah this is where I find it hard to compare, I mean my instincts tell me that a mobile GPU can't keep up with the HD6250 found in C-50, but I really have nothing to back it up with, since they mostly meassure mobile GPU's performance in Triangles Per Second, and I can't find any such data on the HD 6250.
So this is basically what I want to achieve with this thread a comprehensive comparison of Tablet performance, mainly I want to focus on x86 tablets but it is hard to get around the surge of ARM based tablets popping up, and I did see nVidia's claim that Tegra 3 or something would beat even a Core 2 Duo.

Anyway I hope you can help me understand this better, I've been Googling all morning since it's a slow day here at work.

So currently what I'm asking is how does SGX 535 compare to HD 6250? And what I hope to see is more comparison between tablet CPU's and GPU's.

*Edit*Oh sorry if I posted this the wrong place, I just realized my main question was about GPU, and the reason why I posted it in the CPU board was because this was a comparison of Intels tablet CPU with AMD's Tablet APU.

[Medallish]

Well I've found something and I feel like I overlooked searching on Gflops, anyway this is from a random post in an Anandtech article, and it's quite interesting, we can see that the 6970 Gflops numbers are correct, I'm not so sure about the others, but from what I can see they seem pretty accurate.

Tegra 2 - 4.8 GFLOPS (8, 1-way ALUs @ ~300MHz)
PowerVR SGX543MP2 - 19.2 GFLOPS (8, 4-way ALUs @ ~300MHz??)
Radeon 9700 Pro - 33.8 GFLOPS (8, 4-way ALUs (pixel) + 4, 5-way ALUs (vertex) @ 325MHz)
Radeon 2400 Pro - 42 GFLOPS (8, 5-way ALUs @ 525 MHz)
Radeon 5450 - 104 GFLOPS (16, 5-way ALUs @ 650MHz)
Xenos (Xbox 360) - 240 GFLOPS (48, 5-way ALUs @ 500MHz)
RSX (PS3) - 255.2 GFLOPS (24, 2 x 4-way ALUs (pixel) + 8, 5-way ALUs (vertex) @ 550MHz)
Radeon 6970 - 2703.4 GFLOPS (384, 4-way ALUs @ 880MHz)

So from this we can pretty safely say that the SGX 535 in Z670 is way slower than the HD6250 in C-50, but from what I can find on Wikipedia the HD6250 "only" has around 45 Gflops, it's still over twice as much as the SGX543 that we see in Ipad 2, but yeah I hope we see a Die shrink soon.

I also found that AMD's C-50 is speculated to be priced as high as 60$, it was just speculation, but I'm not sure if Intel or AMD sees the need to push off the competition from ARM, I mean personally I want my Cellphone to be ARM based, but anything bigger and I prefer something I have more control over, in my case that would be the X86 based one, so for me it's a somewhat easy choice, but for new costumers I don't think AMD or Intel is really showing off what you can actually do with X86 in tablets, where nVidia, PowerVR, and Qualcomm seems to constantly show off cool looking demoes, I have a cousin who thinks Tegra 2 is the second coming :P, he doesn't know much about tech, but at some point he saw a tech demo of Tegra 2 and bought into the hype.

[265586888]

GFLOPs doesn't mean anything here, you will have to account for the efficiency (actual vs. theoretical), and how easy you can utilize it.

[Ailuros]

The C50@280MHz is comparable (wherever it is comparable, obviously beneath DX10) to the SGX543 MP4+@200MHz in terms of performance.

Frequency aside (the MP2 in the iPad2 seems to be clocked at 267MHz if that data is correct - see also PM) the MP2 in the iPad2 is half of that. No surprise either since the NGP SoC should be in the =/>4W power consumption league (and has of course a quad core A9), while tablet/smart-phone SoCs are rather in the milli-Watt direction.

If you could get me the exact frequency of the SGX535 in the Z670 I could make a rough performance estimate. If it's the 400MHz variant in fillrate bound conditions it's not all too far apart from a MP2, but all the rest obviously is.

[Exophase]

Here's some more info:

SGX535: 2 ALUs (2x FP16 MADD each), 2 TMUs, 2 ROPs
HD6250: 80 ALUs (1x FP32 MADD each), 8 TMUs, 4 ROPs

(note: the above isn't absolute, it's based on HD6250 being divided like AMD's Cedar chipset)

The difference in triangle setup is probably also pretty vast. SGX535 is capable of a triangle every 15 clocks (according to Intel documents).. I don't know what it is exactly for HD6250 but I'm assuming it's higher.

HD6250 also has the edge on feature set, being a DirectX 11 part instead of DX9 like SGX535.

$60 for C-50 wouldn't come as a big surprise.. originally Ontario was announced as strictly 9W for the dual 1GHz/280MHz GPU part. 5W is dramatically less, to the extent that I'm sure there's some heavy binning going on.

I wouldn't expect a die shrink soon. I mean, the thing just came out. Expect the 28nm version in 18 months or something.

I find your comment about ARM and control a little strange.. if it means anything to you, you can run "normal" Linux distros like Ubuntu on ARM. There's an upcoming platform from a Chinese company called Nufront, that will sport a 1.6-2GHz dual-core Cortex-A9 in a netbook that runs Ubuntu and Android. Of course, if you care about GPU performance, and this seems to be a priority for you, Brazos will still be drastically ahead of the Mali-400MP2 that chip offers.

[265586888]

AMD promises yearly updates to the GPU part of the APU, so around 12 months we will see something new (Krishna/Wichita) replacing Zacate/Ontario, and it don't have to strictly mean a die shrink.

[Exophase]

One other thing I completely forgot to mention that 265586888 alluded to..

The SGX parts are TBDR, so they can be more efficient depending on workload. HD6250 is going to be early-Z/hi-Z, so typically the difference in hidden surface removal demands will amount to HD6250 needing a Z-prepass. HD6250 will also typically have substantially higher render target/texturing bandwidth requirements (with possibly lower geometry bandwidth). But of course it's going to be paired with faster memory, too.

[Ailuros]

Exophase,

I guess we'd both agree that triangle rates are as worthless as many other theoretical values.

For IMG's architectures and their internal measurements alone a SGX535@200MHz is capable of 13.5 MTris/s (if the Z670 has the 535 clocked at 400MHz obviously twice as much.

SGX543 is 35M Tris/core at 200MHz. For a MP2 it's times 2 with 95% scaling to 66.5 MTris/s.

The SGX parts are TBDR, so they can be more efficient depending on workload. HD6250 is going to be early-Z/hi-Z, so typically the difference in hidden surface removal demands will amount to HD6250 needing a Z-prepass. HD6250 will also typically have substantially higher render target/texturing bandwidth requirements (with possibly lower geometry bandwidth). But of course it's going to be paired with faster memory, too.

Early-Z can do its part quite efficiently overall; granted there's an advantage but I'd dare to say that if an application culls as it should it won't make much of a difference anymore. You'd get bigger advantages with things like MRTs, deferred shading, post processing and the likes. Also MSAA has on a TBDR roughly the bandwidth and memory footprint cost of coverage sampling AA, it only needs a healthy amount of z/stencil units which the MP2 has already 32, unlike USSE1 SGX cores which outside of SGX545 have all only 8 z/stencil.

One other major difference should be that all other tilers (irrelevant whether early Z/IMRs or wannabe TBDRs) use bounding box tiling. Take a couple of very thin diagonally placed triangles and you'll see how it'll eat into their bandwidth.

That shouldn't of course mean that TBDRs are miracle workers or any of the sort; just a different approach with both advantages and disadvantages.

[PorscheRacer14]

Allow me to quote from my ATI Radeon 9700 Pro marketing material supplied in the box. "The RADEON 9700 is the first graphics accelerator capable of pushing one vertex and one triangle per clock cycle. With the RADEON 9700 PRO's 325MHz core clock frequency, this equates to a vertex or triangle throughput rate of 325 million vertices/second."

[Vithren]

Originally Posted by PorscheRacer14

Allow me to quote from my ATI Radeon 9700 Pro marketing material supplied in the box. "The RADEON 9700 is the first graphics accelerator capable of pushing one vertex and one triangle per clock cycle. With the RADEON 9700 PRO's 325MHz core clock frequency, this equates to a vertex or triangle throughput rate of 325 million vertices/second."

/offtopic
I wonder if the 9700 isn't a meme. Just wait 'till somebody mentions that chip and suddenly someone will write "oh, by the way, I actually remember all the 9700 fame and glory in the days. Good times, good times indeed".
Something like Deus Ex meme.

Oh, by the way, I actually remember...