New Soc from Marvell - Armada

Wishmaster

Newcomer
Here you have press release from Marvell
http://www.marvell.com/products/cellular/marvell_armada_application_processors_mobile/release/1341/
and here you have short article from AnandTech about new family of SoC's
http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=3662

Sounds interesting but still too many unanswered questions. What kind of video performance can we expect? How will the graphics compare to Geforce ULV, Adreno graphics from Qualcomm and PowerVR? Maybe someone who knows something can shed some light on this ;)

Things are getting more and more interesting on the handheld market. Even more interesting than in desktop segment :smile:
 
The differences in the 3D performance between the ARMADAs 510 and 610 listed in Anandtech's table are strange. The 610 has three times the triangle rate but a slower pixel fill.

Having a different custom ARMv7 core to compare against the Cortex A8 and A9 and Scorpion will be interesting. The fast clock speed is an XScale kind of touch. Learning about the GPU's architecture would be nice, too.
 
A new ARMv7 core is interesting, but I'm disappointed it's still in-order and doesn't have NEON SIMD instructions.
On that last point, I strongly disagree with Anand: OMAP3 availability in low priced dev boards such as BeagleBoard has sparkled interest of the community in NEON. To me it looks like it's much more used today than WMMX2 is.
 
Divante? Surely he misheard and it's Vivante - if so, woah, that's quite a high-profile design win for them... It's interesting to notice that they don't say which SoCs they are sampling - since the smartphone one seems most advanced, I'd assume that's the one that isn't just yet. Intriguing that they don't support 3D for the CE/TV chip too, hmm.
 
Well, of all the ARM SoCs out there, the most obvious competition/comparison would be Qualcomm's Snapdragon, since it's the only other custom ARMv7 SoC (architecture license), at least for the Smartphone/MID/Netbook market (plus it also uses a non-IMG GPU). And it already broke 1 GHz, you can even buy a smartphone with it at retail right now and a second one in a few weeks. Strange that Anandtech didn't use Qualcomm as a comparison.

The Marvell ARMADA SoC (600) itself looks quite interesting and powerful. Since it's CPU design and GPU are brand new and "unique", I'm really looking forward for some real power/performance data and release infos.
But devices with it should come out in the next half year or 55nm ARMADA will have to compete against "stock" 45nm Cortex-A8 SoCs like the OMAP36xx (maybe not as optimized but on a smaller process). And at the end of next year we might already see Dual-core out-of-order Cortex-A9 SoCs.
 
First of all: the Armada 510 uses the Vivante GC600, just see this block diagram: http://www.marvell.com/files/technologies/armada/Armada510_SoC_block_diagram.jpg

Secondly: I have no idea what the Armada 610 uses. It doesn't fit the GC800 or the discontinued GC500. It doesn't fit Mali. And it doesn't quite fit SGX either. Apparently the 45MTri/s isn't a typo since they repeat it elsewhere, which makes me even more confused. I'm also not aware of any OGL1.1 core with such a high triangle throughput so that can't be it either. Oh well, who cares!

Anyhow GC600 is pretty nice in terms of texture rate, although the shader performance is likely to be slightly less impressive: according to http://www.ocpip.org/uploads/documents/Vivante_DATE_Slides.20090416.pdf their ALUs are simply Vec4 (presumably MADD) and the GC600 only has one of them. I know we had a thread regarding Vivante's products/die size, but I never contributed to it so I thought I'd do that here ;) After reading a bit, here are a few ideas on how they achieve that die size compared to SGX or others:
1) Very small shader core; Vec4/probably-not-so-fast-branching vs Scalar/MIMD is quite a difference to say the least. I also doubt they support faster INT8/FP16 ala SGX.
2) TMUs don't support anisotropic filtering or 3D textures (obviously cubemaps though). Who knows what their subpixel accuracy is among other things.
3) Their antialiasing technique is described as "Patented anti-aliasing algorithm uses one-fourth the memory and processing power to achieve the same quality as multi-sample anti-aliasing" - aka, flee for the exits? I don't know, that kind of claim (especially on an IMR!) just scares me a bit but who knows maybe it's just overzealous marketing and the solution is actually rvery easonable.
4) "Optional dedicated 2D unit" - I guess their numbers don't include that, so if you want better API suport and/or save power consumption, add a GC200/300/350? Not clear what happens if you don't and how comparable it is to competitive GPU in native 2D support.

It's not obvious to me that Snapdragon is the best comparison because it has an integrated baseband. Certainly they will compete quite a bit, but it'll likely compete even more so against NVIDIA/Samsung/TI SoCs (and a whole bunch of others for the CE chips). It would be very interesting to compare the CPU core with Snapdragon's though, yeah. Obviously all this once again makes Marvell a very real competitor in the application processor business, but there's still one big question...

When did/will those SoCs start sampling? From what I understand, the Armada 510 is the last to sample whereas the others have actually been sampling for some time; although it's so unclear it could actually be the reverse, heh. Anyway assuming that's the case, 55nm isn't very impressive and they might struggle to compete against OMAP4/Tegra2. It is still my understanding based on what I heard that the latter started sampling in Q2 in its 40nm/single-core variant, so devices would probably be available only a little bit later than the 55nm 610 - assuming, of course, that NVIDIA (and their partners) don't screw it up again ;)
 
It's not obvious to me that Snapdragon is the best comparison because it has an integrated baseband. Certainly they will compete quite a bit, but it'll likely compete even more so against NVIDIA/Samsung/TI SoCs (and a whole bunch of others for the CE chips). It would be very interesting to compare the CPU core with Snapdragon's though, yeah. Obviously all this once again makes Marvell a very real competitor in the application processor business, but there's still one big question...
Sorry, I wrote Snapdragon but of course meant Qualcomm's Scorpion+GPU vs. Marvell's Armada CPU+GPU since they are the only custom ARMv7 CPUs plus non-IGM GPUs.

When did/will those SoCs start sampling? From what I understand, the Armada 510 is the last to sample whereas the others have actually been sampling for some time; although it's so unclear it could actually be the reverse, heh. Anyway assuming that's the case, 55nm isn't very impressive and they might struggle to compete against OMAP4/Tegra2. It is still my understanding based on what I heard that the latter started sampling in Q2 in its 40nm/single-core variant, so devices would probably be available only a little bit later than the 55nm 610 - assuming, of course, that NVIDIA (and their partners) don't screw it up again ;)
Wouldn't the OMAP36xx in 45nm be the bigger competition, at least in H1/2010? OMAP4 with it's dual-core and SGX540 seems to me to be too big and power hungry for wide adoption even in high-end Smartphones, at least in it's initial 45nm version.
 
First of all: the Armada 510 uses the Vivante GC600, just see this block diagram: http://www.marvell.com/files/technologies/armada/Armada510_SoC_block_diagram.jpg

Secondly: I have no idea what the Armada 610 uses. It doesn't fit the GC800 or the discontinued GC500. It doesn't fit Mali. And it doesn't quite fit SGX either. Apparently the 45MTri/s isn't a typo since they repeat it elsewhere, which makes me even more confused. I'm also not aware of any OGL1.1 core with such a high triangle throughput so that can't be it either. Oh well, who cares!

Anyhow GC600 is pretty nice in terms of texture rate, although the shader performance is likely to be slightly less impressive: according to http://www.ocpip.org/uploads/documents/Vivante_DATE_Slides.20090416.pdf their ALUs are simply Vec4 (presumably MADD) and the GC600 only has one of them. I know we had a thread regarding Vivante's products/die size, but I never contributed to it so I thought I'd do that here ;) After reading a bit, here are a few ideas on how they achieve that die size compared to SGX or others:
1) Very small shader core; Vec4/probably-not-so-fast-branching vs Scalar/MIMD is quite a difference to say the least. I also doubt they support faster INT8/FP16 ala SGX.
2) TMUs don't support anisotropic filtering or 3D textures (obviously cubemaps though). Who knows what their subpixel accuracy is among other things.
3) Their antialiasing technique is described as "Patented anti-aliasing algorithm uses one-fourth the memory and processing power to achieve the same quality as multi-sample anti-aliasing" - aka, flee for the exits? I don't know, that kind of claim (especially on an IMR!) just scares me a bit but who knows maybe it's just overzealous marketing and the solution is actually rvery easonable.
4) "Optional dedicated 2D unit" - I guess their numbers don't include that, so if you want better API suport and/or save power consumption, add a GC200/300/350? Not clear what happens if you don't and how comparable it is to competitive GPU in native 2D support.

It's not obvious to me that Snapdragon is the best comparison because it has an integrated baseband. Certainly they will compete quite a bit, but it'll likely compete even more so against NVIDIA/Samsung/TI SoCs (and a whole bunch of others for the CE chips). It would be very interesting to compare the CPU core with Snapdragon's though, yeah. Obviously all this once again makes Marvell a very real competitor in the application processor business, but there's still one big question...

When did/will those SoCs start sampling? From what I understand, the Armada 510 is the last to sample whereas the others have actually been sampling for some time; although it's so unclear it could actually be the reverse, heh. Anyway assuming that's the case, 55nm isn't very impressive and they might struggle to compete against OMAP4/Tegra2. It is still my understanding based on what I heard that the latter started sampling in Q2 in its 40nm/single-core variant, so devices would probably be available only a little bit later than the 55nm 610 - assuming, of course, that NVIDIA (and their partners) don't screw it up again ;)

Well it's an interesting design nonetheless. If we would have any numbers from IMG on VGX150's die area it would be an interesting comparison to Vivante's <1mm2@65nm claim. For which by the way the 8xMSAA algorithm is also patent pending in the relevant whitepaper. Quick extraction:

POWERVR VGX150 offers a novel patent-pending 8x Multi-Sampling Anti-Aliasing (MSAA) mode enabling high quality with minimal processing and bandwidth
impact. Higher performance 4x MSAA and No AA modes are supported as well as a per-object AA disable option. A multi-pass AA technique offering 16x or 24x sampling is also available. Unlike coverage-only AA approaches, the VGX150 algorithms are fully compatible with Flash and do not suffer from bleed- through artefacts.

Don't wonder though that Simon laughed last time he saw the following:

Vivante guarantees our GPU to meet or exceed stated performance in silicon.
Vivante guarantees our GPU to meet or exceed stated area in silicon
Many engineers would wish they could have the latter freedom. I don't even know why they include such a ridiculous statement. I'll guarantee you it might get bigger than we estimated and we can then recap if we still have the smallest dies in the market.....LOL
 
The SGX540 in OMAP4 should be at least as small and power efficient in 45nm as its 530 counterpart was in OMAP3 on TI's 65nm process.

Actually, I felt TI perhaps should've targeted a single-core 543, but their launch schedule was apparently too far ahead for that.
 
Back
Top