mali
-
What do you experts say about Marvell's modified Cortex A8 architecture (Armada 610)? I don't have links and didn't do research, but I've read that they shortened A8's pipeline.
Pandora Not Powerful Enough?
- Thread starter Kangal
- Start date
Exophase
Nothing good will ever come of Exophase.
As Laurent told us before, you can't make a modified Cortex-A8. This is a totally different processor that might resemble A8 but is built from the ground up.
Marvell has actually had superscalar ARM cores available for a while now, but I don't know of anything that has used them. Earlier, I was confused into thinking that Sheeva Plug had one, but actually it was a different model under the same processor line.
Pipeline length can be quite overstated. Oversized pipelines have gotten a bad rep because of Pentium 4, but Pentium 4 was slower than other x86 at similar clock speeds for a slew of other reasons as well. Smaller pipelines usually have lower branch misprediction and other serializing event delays (but not always), but this can also come at the expense of trading pipeline stages for multi-cycle operations, basically undoing some of what pipelining gained when originally introduced. Sometimes a bigger pipeline can even cause less stalls instead of more (such as the transition from the traditional 3 stage ARM pipeline to the 5 stage one in ARM9)
What makes a bigger difference than straight up length is how the workload is distributed through it. Cortex-A8 suffers a little bit from having address generation, shifts, and ALU operation execution all taking place in different stages. If any of these took place in the same stage then there'd be less dependency stalls. Generally speaking, a long front end involving fetching, decoding, and scheduling can be okay so long as the instruction flow is well predicted. But some branches will never be well predicted, so shortening the mispredict penalty does always help. Some architectures can actually shorten the mispredict penalty if the software can determine the branch direction well in advance. ARM11 could save a couple cycles, and PowerPC was designed with this especially in mind.
Apparently Marvell claims a much higher typical instruction per cycle ratio, but I haven't seen the numbers, much less an explanation for how they're achieved. So this is a pretty big unknown to me right now, much like Snapdrgon. If someone had some timing information that'd be nice, but some CPU manufacturers seem to be loathe to release such things. It's hard to imagine that a different dual-issue in-order implementation of the same instruction set (except WMMX2, meh) can do dramatically better than Cortex-A8. Maybe it has dual ported icache and dual LSUs or something, not that even that would help tremendously.
Marvell has actually had superscalar ARM cores available for a while now, but I don't know of anything that has used them. Earlier, I was confused into thinking that Sheeva Plug had one, but actually it was a different model under the same processor line.
Pipeline length can be quite overstated. Oversized pipelines have gotten a bad rep because of Pentium 4, but Pentium 4 was slower than other x86 at similar clock speeds for a slew of other reasons as well. Smaller pipelines usually have lower branch misprediction and other serializing event delays (but not always), but this can also come at the expense of trading pipeline stages for multi-cycle operations, basically undoing some of what pipelining gained when originally introduced. Sometimes a bigger pipeline can even cause less stalls instead of more (such as the transition from the traditional 3 stage ARM pipeline to the 5 stage one in ARM9)
What makes a bigger difference than straight up length is how the workload is distributed through it. Cortex-A8 suffers a little bit from having address generation, shifts, and ALU operation execution all taking place in different stages. If any of these took place in the same stage then there'd be less dependency stalls. Generally speaking, a long front end involving fetching, decoding, and scheduling can be okay so long as the instruction flow is well predicted. But some branches will never be well predicted, so shortening the mispredict penalty does always help. Some architectures can actually shorten the mispredict penalty if the software can determine the branch direction well in advance. ARM11 could save a couple cycles, and PowerPC was designed with this especially in mind.
Apparently Marvell claims a much higher typical instruction per cycle ratio, but I haven't seen the numbers, much less an explanation for how they're achieved. So this is a pretty big unknown to me right now, much like Snapdrgon. If someone had some timing information that'd be nice, but some CPU manufacturers seem to be loathe to release such things. It's hard to imagine that a different dual-issue in-order implementation of the same instruction set (except WMMX2, meh) can do dramatically better than Cortex-A8. Maybe it has dual ported icache and dual LSUs or something, not that even that would help tremendously.
mali
-
Ah yes, you are right, it can't be called Cortex, its just an ARM v7 compatible chip. Thanks for the explanation 
hobbyman II
Member
- Joined
- Sep 7, 2008
- Messages
- 487
47M Triangles /sec (peak) <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<
24M Drawn Triangles /sec (peak)
600M Pixels /sec (peak)
240M Textured Pixels /sec (peak)"
maybe they do some occlusion on chip, so a scene and models can have 47 million triangles, and of these the visible triangles (about half .. ie 23.5 million rounded up to 24 million cos sales dept said so
) are drawn ?disclaimer: just based on knowledge of rendering from Darkbasic usage, when it comes to C I'm a 2d coder.
Exophase
Nothing good will ever come of Exophase.
hobbyman II said:maybe they do some occlusion on chip, so a scene and models can have 47 million triangles, and of these the visible triangles (about half .. ie 23.5 million rounded up to 24 million cos sales dept said so
I'd want to defer to someone who knows better, like darkblu, but I personally doubt it. You couldn't tell if a triangle is occluded without full deferred rendering. If you only needed its geometry and didn't have to render it then you wouldn't have to do a lot of the per-vertex work though. But who knows if that was included in their numbers to begin with.
It could refer to clipping.. don't need deferred rendering to determine if a triangle is within the clip view or not.
Last edited by a moderator:
Exophase said:hobbyman II said:maybe they do some occlusion on chip, so a scene and models can have 47 million triangles, and of these the visible triangles (about half .. ie 23.5 million rounded up to 24 million cos sales dept said so
I'd want to defer to someone who knows better, like darkblu, but I personally doubt it. You couldn't tell if a triangle is occluded without full deferred rendering. If you only needed its geometry and didn't have to render it then you wouldn't have to do a lot of the per-vertex work though. But who knows if that was included in their numbers to begin with.
It could refer to clipping.. don't need deferred rendering to determine if a triangle is within the clip view or not.
Could it refer to front-to-back rendering as well?
ie, perform the Z-check but not the texture sample?
Last edited by a moderator:
Exophase
Nothing good will ever come of Exophase.
andys said:
I'm thinking not because this is pixel-level and thus a matter of fillrate, not vertex rate.
Last edited by a moderator:
howardcshaw
Still Fresh
- Joined
- Oct 10, 2009
- Messages
- 2
iprice said:The most worrying aspect here for me is that most users seem to want the Pandora for emulation as its primary purpose. That also seems such a waste and a shame for me. Emulation should be an added bonus. I really hope we get some decent homebrew and commercial apps too. Don't get me wrong, I also want to play emulated games, however I want so much more than the ability to play old games on a new machine, when I can play the old games on the old, original machines already and at full speed.
Everyone needs to raise their expectations, methinks
Heh, not all of us. Emulation is a nice sideline for me, but I'm primarily getting it for the keyboard w/clamshell combination. I had a Nokia N800, but the screen cracked on it. I've also got a sony E-Reader with a cracked screen. I'm going to be reading fanfiction on it, and writing my fanfiction on it (I really look forward to waking up in the middle of the night with an idea in one of my stories, and being able to reach over, snap it open, alt-f-whatever to get to the console, and type it in.)
I like to swap in to a game when I'm writing on my stories, or coding, and then pop back, so having gaming controls just makes that an extra plus.
Howard, the Grum
Last edited by a moderator:
GizmoTheGreen
Active Member
do i smell some visual novel?
CoLDSToRAGE
Still Fresh
My main reason for getting a Pandora is as a super portable mini-laptop for doing some on-the-road work.
I'm also looking forward to emulation of 8/16bit systems for downtime fun, rather than current devices.
The Pandora looks like it will emulate these older consoles/computers beautifully.
As for newer/current platforms, I own these and I'm happy to play the games on the 'real thing' for as
they are usually longer, chair bound gaming sessions.
I guess if money is tight and you really want a "do everything" portable emulation machine, you'd be better
off with a good spec. web-book (or small laptop) and a good USB Joystick/controller?
I'm also looking forward to emulation of 8/16bit systems for downtime fun, rather than current devices.
The Pandora looks like it will emulate these older consoles/computers beautifully.
As for newer/current platforms, I own these and I'm happy to play the games on the 'real thing' for as
they are usually longer, chair bound gaming sessions.
I guess if money is tight and you really want a "do everything" portable emulation machine, you'd be better
off with a good spec. web-book (or small laptop) and a good USB Joystick/controller?
Kangal
Member
- Joined
- Jan 6, 2010
- Messages
- 245
Here's an interesting read from XDA-Developers forums.
Tegra Vs Snapdragon.
It was concluded that theoretically Tegra has better graphics and more battery power VS Sdragon which has alot more power.
The respondent said it was a trade off, but he'd rather more power as everything in general would speed up.
http://forum.xda-developers.com/showthread.php?t=567401
Tegra Vs Snapdragon.
It was concluded that theoretically Tegra has better graphics and more battery power VS Sdragon which has alot more power.
The respondent said it was a trade off, but he'd rather more power as everything in general would speed up.
http://forum.xda-developers.com/showthread.php?t=567401
ldesnogu
Well-Known Member
This talks about the old Tegra...
Exophase
Nothing good will ever come of Exophase.
Laurent said:
Yes, he thinks we care about the Tegra 1 and/or doesn't really understand the difference between it and Tegra 2. He's the only one who has been mentioning it all thread.
Last edited by a moderator:
Similar threads
