[ninelven]

@ pcchen, I was being slightly conservative with the estimates.

Let's put it another way... do you think the Phenom II competes with an i7?

[Albuquerque]

Originally Posted by ninelven

I didn't say 25x to 30x faster than today, I said 25x to 30x faster than the iphone 4.

I'd expect a Denver core to offer ~2x the performance/clock as a Cortex A8. The CPU in the iphone 4 is clocked at 800 MHz, while a Denver core (in Maxwell) will probably be at least 2.4 GHz. There is only a single core in the Iphone 4, there will probably be 4 to 8 in Maxwell. I leave it to you to do the math; surely you can get that right.

According to your own words, I did the math exactly right, and you obviously ignored it because it's easier to sidestep the situation than tackle it directly.

I'm sure you'll balk, so let me remind youof my math, and you can now use your next reply to properly refute it:

You're talking about octupling(sp?) total transistor count if not more, and you're also talking about no less than doubling of clock speed. So somehow we're going to power 8x more transistors at twice their current speed on perhaps two process shrinks from today? At the very best, that kind of chip would give you i3-esque performance at roughly the same power envelope.

It seems to me that you're hoping that ARM will leap three Moore's Law's cycles in 30 months (reminder: transistor count doubles every 18 months per good ol' Moore), and at the same time, Intel will somehow stand still for that entire time. Let's not ignore the fact that Intel's fabs are pretty much a minimum of 12 months ahead of EVERYBODY right now, and have just recently stretched that by probably another 6 months with their new 22nm finfet process...

Please, elaborate on which math problem I got wrong, and then elaborate on why you think this is even plausible.

[ninelven]

Nah, I'll just wait the 2 years...

[swaaye]

Also, since when does 4-8 cores mean much to general purpose computing? Once you get past dual core the benefits diminish rapidly for most usage. Or are we considering these ARM CPUs as competitors in massively parallel tasks?

I think it would be fascinating if one of the big CPU companies put all of their knowledge to work at building a really impressive ARM core. Unfortunately all of the big companies probably have a vested interest in not doing that.

[Blazkowicz]

a quad core is way better than automagical "GPU acceleration will make it ok" (can we acknolewdge that there are pretty much no OpenCL applications?, and that they need to be written for a specific GPU's memory architecture to be worth existing)

nvidia denver is interesting and might change things, currently there's only a niche market for gpgpu, running on nvidia.
but I doubt we'll see that on small laptops.

[swaaye]

GPU GUI acceleration is something that does matter. But we are already at the point where the GPUs are adequate for that.

But yes GPGPU is definitely very niche primarily because GPUs are not generally useful. The more they become so, the more transistors they need and the less GPU they become. The PowerVR SGX cores even cut back on GPU things (they do a blocky texture filtering for example).

[Simon F]

Originally Posted by swaaye

The PowerVR SGX cores even cut back on GPU things (they do a blocky texture filtering for example).

Ehh?

[Gubbi]

http://forum.beyond3d.com/showthread.php?t=59519

Cheers

[pjbliverpool]

Cool, that's exactly what I was looking for. Cheers!

[swaaye]

Originally Posted by Simon F

Ehh?

The blocking is subtle and seems somewhat angle dependent. It reminds me of some of the really old 3D chips that simplified bilinear filtering but it's more subtle than that. I'll have to look closer at it sometime. I've seen it in the Android Quake 2 port and a few of the Android games.

These are using the older SGX 535 (I think) so maybe that's it.

Also, maybe this texture filtering analysis done on GMA 500 shows it.

http://translate.google.com/translat...6prmd%3Divnsfd

[swaaye]

Originally Posted by Gubbi

http://forum.beyond3d.com/showthread.php?t=59519

Cheers

Excellent write up. It will be interesting to see if 32nm Atom can make that small window of opportunity work.

[Simon F]

Originally Posted by swaaye

The blocking is subtle and seems somewhat angle dependent. It reminds me of some of the really old 3D chips that simplified bilinear filtering but it's more subtle than that. I'll have to look closer at it sometime. I've seen it in the Android Quake 2 port and a few of the Android games.

AFAIK the blending in the filtering unit uses relatively (compared to some) high precision for the weights.

Also, maybe this texture filtering analysis done on GMA 500 shows it.

Wasn't the driver for the GMA500 done by a third party?

[argor]

Originally Posted by Simon F

Wasn't the driver for the GMA500 done by a third party?

jeb tungsten graphics

[Blazkowicz]

Tungsten graphics writing the drivers is much decried, but it has interesting implications on the linux side. (yes, good drivers would be as much or more interesting!)

Tungsten are the ones who made Gallium3D, a platform that any "state trackers" may use, i.e. OpenGL, Open GL ES, OpenVG etc. drivers, proprietary or open source.

Gallium3D has allowed great advances for 3D opensource drivers, such as much faster R300 and up drivers, or 3D support for the nouveau driver.

they made the 2D driver for the GMA 500 open source, and the 3D part using gallium is proprietary. "anyone" can write an opensource driver if they wish (yep, an impossibly hard task that may be done some day..)

http://www.phoronix.com/scan.php?pag...item&px=NzY2Mg

[swaaye]

Well I tried capturing some Quake 2 shots off my Nook Color with Quake 2's screen capture function but they just come up as black images so it's not working right. I'm wondering if the filtering uglies are some sort of software issue though because it is not always apparent whereas sometimes a texture will be super blocky.

Most of the time the trilinear filtering is impeccably smooth and luxurious.