Samsungs new Soc, but who's supplying the Graphics ?

[tangey]

Samsung have just released the first specs for their new Soc that was announced in Feb.

http://www.samsung.com/global/system/business/semiconductor/product/2008/4/23/370911ptb_s3c6410_rev20.pdf

The previous gen processor is the one in the iphone, using MBX 3D from IMG.

The specs above state that the 3D hardware is OpenGles 2.0 compliant, something that MBX is not, and Samsung has not licenced IMG's OpenGles 2.0 compliant core, SGX.

So who is supplying the graphics, is it Arm's own Mali, is this Apple supplying its own SGX licence to Samsung to produce the chips (but surely they would not publish even the existence of this chip in that case).

[Arun]

•Floating-point pipeline & Object-order rendering
•4-Way SIMD vertex shader+ pixel shader
•ShaderModel 3.0: World 1st implementation
•128-bit (32-bit x 4) FP x 1 Vertex Shader
128-bit (32-bit x 4) FP x 1 Pixel Shaders
8-stage pipeline
512 Instruction Slots (configurable)
•Memory BW Optimization by Hierarchical Caching

Doesn't look like SGX at all to me, since it's SIMD and not unified. Can't be ATI's either and NVIDIA doesn't license theirs... So I'm thinking ARM/Falanx's Mali, yeah... Congrats to the guys working there if that's true. And if it's not, I am incredibly curious who might be behind that IP then.

[Simon F]

Could well be: http://www.reuters.com/article/pressRelease/idUS76869+07-Jan-2008+PRN20080107

[tangey]

I was aware of that press release, however it doesn't mention Samsung in relation to hadware cores, it mentions samsung in relation to a JSR graphics "Engine", i.e. a software layer.

Seeing that IMG are getting circa 8M royalities from iphone by end of this year, and possibly more, and various analyst are predicting between 15-30M iphones for 2009, it would be a major loss of revenue for IMG IF, and I repeat IF Apple continue with samsung as the supplier of the Soc, and use this new, non-IMG version.

A bit rich of them suggsting its the first implementation of SM3, doesn't SGX have SM3, which is implemented in Intels Centrino Atom SCH chips that have been demo-ed at Intels IDF and also CES (TI OMap3 is also SGX, its also been demoed).

[Rob Evans]

Seeing that IMG are getting circa 8M royalities from iphone by end of this year, and possibly more, and various analyst are predicting between 15-30M iphones for 2009, it would be a major loss of revenue for IMG IF, and I repeat IF Apple continue with samsung as the supplier of the Soc, and use this new, non-IMG version.

Was that 8M in $ or £?

Either way it seems too high - IMG will be sharing the royalties with ARM, as the MBX license was signed with Samsung under the joint agreement IMG had with ARM at the time.

Also the IP is for the MBX Lite, so will be at the lower end of the royalty scale.

Finally IMG have previously said that royalty levels decrease with volume, so all round I would have thought IMG are getting around 10p to 20p royalty per chip.

With Apple indicating expected sales around 10M for the first year, I would expect royalties of around £1.5m to IMG. Not massive, nor insignificant, but with the projected volumes for subsequent years, including additional models, the loss of this to ARM would be very disappointing.

Rob.

[tangey]

Was that 8M in $ or £?

Rob.
Thats was 8M royalties i.e. 8M units

HI Rob, Its Ian :)

I see the iphone as just being the first in a series, and quite likely for unit sales to ramp up going forward ala ipod. I'd dont want that to happen without IMG inside.

There seems to be no info as yet that Apple have settled on a next gen platform, I wouldn't want it to be the new samsung chip.

[Rob Evans]

Thats was 8M royalties i.e. 8M units

Ah, of course - sorry!


HI Rob, Its Ian :)

I did wonder, given the similarity in viewpoints :)

I see the iphone as just being the first in a series, and quite likely for unit sales to ramp up going forward ala ipod. I'd dont want that to happen without IMG inside.

There seems to be no info as yet that Apple have settled on a next gen platform, I wouldn't want it to be the new samsung chip.

I also see the current iPhone as the first in a series, and believe it and subsequent variations will sell in massive quantities, so I totally agree regarding the desirability of the S3C6410 not being chosen.

Rob.

[knux]

I would guess it has to be mali, cuz if it were to be an SGX it would have a unified shader architecture.

[roninja]

Imagination Technologies Group plc (LSE:IMG; "Imagination") - a leader in System-on-Chip Intellectual Property ("SoC IP") - reports that it has signed a licence agreement with Samsung Electronics Co., Ltd. (“Samsung”) with respect to certain POWERVR SGX graphics and VXD video IP cores.
This licence enables Samsung to manufacture semiconductor devices which integrate these IP cores from Imagination.
Under the terms of its licensing arrangements, Imagination receives licence fees and royalty revenues on SoCs incorporating Imagination’s IP.

[tangey]

As per my post on another thread, this tells me the recently announced Samsung Soc S3C6410 will NOT be used by Apple, and that Samsung have just been given the permission my IMG to manufacture chips containing their IP, for use by Apple.

[Arun]

Alternatively, maybe Samsung synthetized VXD for Apple, then realized 'holy shit this is a bazillion times better than our technology :(' and then they ran to PowerVR to license it and got SGX at the same time... :p

[Lazy8s]

Samsung licensed the right to only manufacture those PowerVR IPs, not to implement them into their own designs, from the wording of the PR.

As the others were saying, Samsung seems to be facilitating their ability to compete for the manufacturing of a next generation iPhone SoC.

[tangey]

Samsung licensed the right to only manufacture those PowerVR IPs, not to implement them into their own designs,


Spot on.


See my post on the PA semi discussion (attached to the news article on the front page), for some more thoughts on this.

[Arun]

Hmm, yeah, now that I read the wording again you're right. So the big question is: would they need to do this 1+ year ahead of production so they can start on the synthesis work or tape-out? Or could this confirm the iPhone 3G uses VXD? Because if that's true...

[roninja]

http://www.appleinsider.com/articles/08/04/30/apples_bionic_arm_to_muscle_advanced_gaming_graphi cs_into_iphones.html

Extensive coverage now on Appleinsider thats to "T" for sorting them out!

All this looks most likely for the 2009 version, the current version most probably will be based on the same hardware given the timing coincides closely to the official SDK release.

[Lazy8s]

Yeah: all indications from Imgtec are that Intel's System Controller Hub and Texas Instruments' OMAP3 will be the first implementations of SGX.

[Arun]

I thought I'd bump this thread because I *very briefly* talked to someone at Samsung during Mobile World Congress, and from what I could understand (his accent was even worse than mine!) their OpenGL ES 2.0 3D core is actually in-house (!!!)

This kinda makes sense given the specs, since I don't think Mali does "Object-order rendering" or "Memory BW Optimization by Hierarchical Caching". Amusingly, they are also the most open about performance in the latest version of their datasheet - yay?

􀂃Rendering performance @Max freq. (133MHz)
•Peak vertex geometry performance (transform only): 9.28M vertices/s
•Vertex geometry performance with single light condition: 7.55 vertices/s
•Shaded fill rate: 125.6M pixel/sec
•Bilinear-filtered textured fill rate with Alpha blending: 37.8M pixel/secIt looks decent, until you look at the last line. 37.8M is less than 1/3rd the peak throughput; what's going on here?! I guess that's not too bad for blending, but what scares me is that 'shaded' fillrate does not include the word 'bilinear'. I certainly hope it's full-speed for bilinear! *sigh*

I suspect that they use a highly customized version of ARM's Mali SW for their drivers; it is interesting to see that ARM's website always seems to imply that this SW is optimized for Mali, but not necessarily exclude for it (even excluding the fact some of it also works on ARM CPUs). It seems pretty damn crazy to want to do your own 3D core nowadays, but then again Samsung is also making their own LTE and WiMax single-mode chips sooo... :) (that makes sense in terms of building IP/patents to reduce license/royalty fees in a way that making a 3D core doesn't though)

At their MWC stand, Samsung was demoing a pretty large number of things (including mobile TV) but wasn't being very specific about any of them. They had a demo of their next app processor which supports 720p video playback, and they had a roadmap diagram that showed they would support 1080p in 2010. So I guess they're still want to be a serious competitor in that market.

EDIT: Only alternative I can think of is that it's a GPU from Vivante, but I don't think they really fit the bill... So I would still tend to think it's in-house, although if anyone has any other info I'd obviously be very interested.

[roninja]

Can't be SGX given that Samsung only took a "direct" licence late last year long after this SoC became available...

[Lazy8s]

An old roadmap from Samsung of their mobile 3D plans implied that, after their MBX Lite-equipped SoCs, they'd use different IP for an incremental upgrade (which they rated at a performance of 10M tri/sec, if I remember correctly.)

In-house sourcing of graphics hardware IP isn't unheard of in this space. Another major semi, Toshiba, has been doing it all along, too. Their ES 1.1 co-processors were followed up with an announcement of a "100M tri/sec" ES 2.0 accelerator, speculated to be the salvaged fruits of the R&D that went into the failed contender for the PS3's GPU, the RS.

[Lazy8s]

That old Samsung roadmap specified a 3D part rated at 5M tri/sec, actually, and scheduled for availability around 2006, if I recall, to follow up their 1M tri/sec MBX Lite part, now that I'm thinking back again.

Even at that time, Samsung appeared to be designing with multiple 3D hardware IPs.

[tangey]

Just noticed that on samsungs S3C6410, somewhere between the Advanced spec of '07
http://www.samsung.com/global/system/business/semiconductor/product/2008/4/23/370911ptb_s3c6410_rev20.pdf

And the production spec of sometime in 2008
http://www.samsung.com/global/business/semiconductor/support/brochures/downloads/systemlsi/s3c6410_datasheet_200804.pdf

The graphics performance reduced from 9.3M tri/sec to 4M.

Perhaps they decided they couldn't live with the power requirements of the original 133Mhz clocked graphics and halved the clock (the GPU clock isn't specified in the later info)

[tangey]

Samsung this week announce two new 1Ghz, 45nm A8 based Socs. S5PC110 and the S5PV210. THe following article is stating that both of them contain SGX graphics.

http://www.eeherald.com/section/new-products/np10010387.html

Its the only article I've seen to do so, others, including Samsungs own PR do not speciy the 3D IP supplier.


Update....Its now on Samsungs Website:-
http://www.samsung.com/global/business/semiconductor/newsView.do?news_id=1043

So thats another instance of SGX appearing in samples at 45nm before year end.
PR also talks about 30fps encode/decode, I wonder is that Samsungs own solution ?

[kumayu]

Samsung Electronics FIMG-3DSE v1.5
It's a Samsung's inhouse GPU in S3C6410.
They also has a 2D core, "FIMG-2D".

FIMG-3DSE v1.5
one 128bit VS unit+ one 128bit PS unit , 133Mhz.
(32bit x 4)
The performances seems equal or better to PowerVR MBX .
But base on OpenGL ES2.0.

[Lazy8s]

That's interesting to know.

Though it's a short-lived architecture, learning more about what they tried there would be nice. Like with the Zii and darkblu's testing, we're sometimes fortunate enough to get a real look at the performance/operation of a few non-mainstream solutions that pop up now and then.

[tangey]

This article extract gives us a little bit more info on the two 1Ghz processors that samsung announced with SGX........

Continuing its transformation from a low-cost vendor to a high-performance supplier, Samsung has rolled out a pair of new application processors, each containing a 1GHz Cortex-A8 CPU. The S5PC110 targets smartphones, whereas the S5PV210 is designed for netbooks. The chips follow the 65nm S5PC100, a custom version of which powers the iPhone 3GS. The new 45nm processors are due to sample in December.

Samsung worked with Intrinsity to develop a custom implementation of the Cortex-A8 CPU. Although the 1GHz CPU is no faster than the Cortex-A8 used in TI's 45nm OMAP3, Intrinsity touts that its design achieves this speed while operating at just 1.0V, minimizing power consumption. Neither TI nor Samsung has disclosed power data, however, so the extent of any power savings is unclear.

The PC110 also extends the multimedia performance of its predecessor, quadrupling the 3D performance to 40 million triangles per second and adding full HD (1080p) video encoding and decoding. To better position the processor for netbooks, Samsung created the PV210 version, which uses a less expensive (0.65mm pitch) package and adds a second 32-bit DDR2 SDRAM channel to boost software performance.

http://blog.linleygroup.com/2009/10/samsung-pushes-cortex-a8-to-1ghz.html


40M tri per sec....does this make it SGX530/535 ?

[Wishmaster]

This article extract gives us a little bit more info on the two 1Ghz processors that samsung announced with SGX........

Continuing its transformation from a low-cost vendor to a high-performance supplier, Samsung has rolled out a pair of new application processors, each containing a 1GHz Cortex-A8 CPU. The S5PC110 targets smartphones, whereas the S5PV210 is designed for netbooks. The chips follow the 65nm S5PC100, a custom version of which powers the iPhone 3GS. The new 45nm processors are due to sample in December.

Samsung worked with Intrinsity to develop a custom implementation of the Cortex-A8 CPU. Although the 1GHz CPU is no faster than the Cortex-A8 used in TI's 45nm OMAP3, Intrinsity touts that its design achieves this speed while operating at just 1.0V, minimizing power consumption. Neither TI nor Samsung has disclosed power data, however, so the extent of any power savings is unclear.

The PC110 also extends the multimedia performance of its predecessor, quadrupling the 3D performance to 40 million triangles per second and adding full HD (1080p) video encoding and decoding. To better position the processor for netbooks, Samsung created the PV210 version, which uses a less expensive (0.65mm pitch) package and adds a second 32-bit DDR2 SDRAM channel to boost software performance.

http://blog.linleygroup.com/2009/10/samsung-pushes-cortex-a8-to-1ghz.html


40M tri per sec....does this make it SGX530/535 ?

No AFAIK it is based on samsung's own in-house GPU. Their using their own GPU's even now in Omnia 2 and Omnia Pro.

[tangey]

You're wrong, samsungs own press release (a few posts up) confirms that both Socs are using SGX, but it doesn't say which variant.

Here's the link again.

http://www.samsung.com/global/business/semiconductor/newsView.do?news_id=1043

[darkblu]

Continuing its transformation from a low-cost vendor to a high-performance supplier, Samsung has rolled out a pair of new application processors, each containing a 1GHz Cortex-A8 CPU. The S5PC110 targets smartphones, whereas the S5PV210 is designed for netbooks. The chips follow the 65nm S5PC100, a custom version of which powers the iPhone 3GS. The new 45nm processors are due to sample in December.
so this is known for a fact now, or is that just speculation?

[Wishmaster]

You're wrong, samsungs own press release (a few posts up) confirms that both Socs are using SGX, but it doesn't say which variant.

Here's the link again.

http://www.samsung.com/global/business/semiconductor/newsView.do?news_id=1043

My bad. Thought that if they use in-house developed GPU on omnia 2 their new SoC would use it too but it seems that they went for 535 or 540 but I'm more inclined to 540.

[Arun]

I'd guess 540 too (both would fit the poly performance) because PowerVR claimed multiple partners were taping-out with it quite some time ago and so far we've only seen TI and NXP iirc. Plus, 540 is just a better arch if you don't need the extra PC-centric features of the 535 (and didn't need to design your SoC before the 540 was ready, I suppose).

[tangey]

I'd guess 540 too (both would fit the poly performance) because PowerVR claimed multiple partners were taping-out with it quite some time ago and so far we've only seen TI and NXP iirc. Plus, 540 is just a better arch if you don't need the extra PC-centric features of the 535 (and didn't need to design your SoC before the 540 was ready, I suppose).


Does Microsoft have any thoughts of producing an ARM version of Windows 7 ? Samsung is aiming one of those chips at netbooks, if there is an ARM windows 7 in the wings then perhaps having a DX9/10 compliant core would make sense (i.e. 535 ?)

[Wishmaster]

Does Microsoft have any thoughts of producing an ARM version of Windows 7 ? Samsung is aiming one of those chips at netbooks, if there is an ARM windows 7 in the wings then perhaps having a DX9/10 compliant core would make sense (i.e. 535 ?)

I don't think they will make ARM version of Windows 7, not now.
Remember that they plan to release WM7 next year which is designed for ARM architecture and they wouldn't want to kill it with ARM port of W7.
Eldar Murtazin stated that for now they plan to focus on getting WM7 to market but then they plan to make Windows more suitable for mobile devices but this will happen somewhere around 2012 so we still got some time.
Link (http://www.mobile-review.com/review/nokia-rx51-n900-en.shtml) to the article. Pay attention to the second paragraph.

[Lazy8s]

Both the smartphone and netbook SoC are likely to be using the same SGX core, which raises some questions about its identity considering the performance numbers stated.

Is the core clocked over 200 MHz? A 540 would need to be nearer 300 MHz in order to push around 40M tri/sec based upon Imgtec's performance guidance, unless Samsung is just quoting more of a peak guidance.

Perhaps the core in question is actually a 545, where the DirectX 10.1 functionality might have at least some usefulness to a netbook, though probably a very limited amount considering it's in an ARM design.

[A.L.M.]

Very interesting thread, I was asking myself the same question...

Samsung Electronics FIMG-3DSE v1.5
It's a Samsung's inhouse GPU in S3C6410.
They also has a 2D core, "FIMG-2D".

FIMG-3DSE v1.5
one 128bit VS unit+ one 128bit PS unit , 133Mhz.
(32bit x 4)
The performances seems equal or better to PowerVR MBX .
But base on OpenGL ES2.0.


Anyway, more detailed specs for the Samsung inhouse mobile gpu. :wink:

http://www.opengpu.org/bbs/attachments/month_0909/09091617516dd30bb526aa1070.png
http://www.opengpu.org/bbs/attachments/month_0909/090916175197728971f1214c67.png


FEATURES

• 4M triangles/s @133MHz (Transform Only)

• 75.8M pixels/s fill-rates @133MHz (shaded pixels)

• Programmable Shader Model 3.0 support

• 128-bit (32-bit x 4) Floating-point Vertex Shader

- Geometry-texture cache support

• 128-bit (32-bit x 4) Floating-point two Fragment Shaders

• Max. 4K x 4K frame-buffer (16/32-bpp)

• 32-bit depth buffer (8-bit stencil/24-bit Z)

• Texture format: 1/2/4/8/16/32-bpp RGB, YUV 422, S3TC Compressed

• Support max. 8 surfaces (max. 8 user-defined textures)

• API Support: OpenGL ES 1.1 & 2.0, D3D Mobile

• Intelligent Host Interface

- 15 input data-types, Vertex Buffer & Vertex Cache

• H/W Clipping (Near & Far)

• 8-stage five-threaded Shader architecture

• Primitive assembly & hard-wired triangle setup engine

• One pixels/cycle hard-wired rasterizer

• One texturing engine (one bilinear-filtered texel/cycle each)

- Nearest/bilinear/trilinear filtering

- 8-layered multi-texturing support

• Fragment processing: Alpha/Stencil/Z/Dither/Mask/ROP

• Memory bandwidth optimization through hierarchical caching

- L1/L2 Texture-caches, Z/Color caches

• System bus interface

- Host interface: 32-bit AHB (AMBA 2.0)

- Memory Interface: two 64-bit AXI (AMBA 3.0) channels

[darkblu]

FEATURES

• 4M triangles/s @133MHz (Transform Only)

Programmable Shader Model 3.0 support

• 128-bit (32-bit x 4) Floating-point Vertex Shader


this reads quite strangely, assuming that:
* they quote tristrips as triagles (1 vertex/tri), like everybody else in the industry
* the quoted 4-wide vector support is actually true

assuming VS work in a pipeline with the trisetup (i.e. VS and trisetup work in parallel), and assuming the trisetup is not the bottleneck (4M/s sounds way too low for a 133MHz trisetup engine anyway), that leaves us with the 133 / 4 = 33 clocks/vertex shader performance as the bottleneck.

now, a matrix/vertex multiply is 4x dp4, or 4x madd (ok, 3x madd + 1x mul) + some swizzles. that gives us either 8 clocks/dp4, or 8 clocks/madd+swizzle. i'd guess the latter. and those seem like some darn expensive swizzles to me.

[tangey]

Regarding samsungs 45nm A8 Socs with SGX graphics, Anandtech seems to be confirming what was conjectured here a view months ago, that those Socs use the 540 variant of SGX.

"I met with Imagination Technologies, the makers of the GPU in everything from the iPhone 3GS to the Motorola Droid. They were showing off a 1.1GHz Samsung Cortex A8 SoC with an integrated PowerVR SGX 540:"
http://www.anandtech.com/tradeshows/showdoc.aspx?i=3719&p=2

CPU is now clocked at 1.2Ghz.

We know from IMG press releases that Samsung has licenced two SGX cores, the 2nd one being announced in Oct '08. So is the other one the 535 implementation in Apples Soc, if not I wonder where the other SGX licence will show up

[Ailuros]

Either Samsung's own devices or a next generation iPhone just to piss the Tegra2 fans off *snicker*

[roninja]

Samsung have held two types of SGX licenses firstly a production license on behalf of Apple and more recently a license allowing it to be embedded into it's own cores. I would concur that this is a 540 license and Anand appears to be confirming this. We shall see if a varient lands in the next iPhone unless apple go fully down the pa semi route. Samsung will certainly find use for more of it's own SoC's in it's own smartphones overtime with less reliance on the likes of TI Omap family.

[Lazy8s]

With such a high clock rate for the CPU, the SGX should comfortably clear above 200 MHz, a speed which would allow a 540 to approach Samsung's announced performance of 40M tris/sec.

[Ailuros]

With such a high clock rate for the CPU, the SGX should comfortably clear above 200 MHz, a speed which would allow a 540 to approach Samsung's announced performance of 40M tris/sec.

IMG's listed triangle rates are as realistic as possible. After they've seen it with the first generation they don't have much reason to doubt them nowadays.

*** All polygon throughput figures are based on real and achievable sustained throughput in a real SoC; they are not theoretical figures that can never be achieved in any practical application.

http://www.imgtec.com/News/Release/index.asp?NewsID=497

[ADEX]

Funny, B3D regulars can usually find news in the most obscure sources imaginable, but Samsung licensing a GPU appears in an earnings release and it's apparently invisible. Go figure...

[wishiknew]

I just brought a samsung b7610 with this 6410 soc. Can't say if the 3d silicon ever gets used though.

[Ike Turner]

I just brought a samsung b7610 with this 6410 soc. Can't say if the 3d silicon ever gets used though.

check this out: http://www.modaco.com/content/i8000-verizon-sch-i920-omnia-ii-http-i8000-modaco-com/293601/fpu-enabler/

[tangey]

The latest PR newsheet from IMG contains some interesting product info.
http://www.imgtec.com/Newsletters/with_imagination/issue7_10.pdf

1) Two new Sony phones, Kurare and Kanna, contain SGX, seems Sony are still developing with the TI Omap3 platform.

2) Omap4 will be in Silicon on IMG's stand at mobile world congress next week

And finally, a nice bit regarding the the samsung S5PC110. I initially put this Soc on my radar back in Feb '09, at which time google had a total of "1" hit for it.
http://forum.beyond3d.com/showpost.php?p=1283065&postcount=106

Samsung announced it end of last year along with the S5PV210 as both containing SGX. IMG in the newsletter confirm that it is SGX540, which was speculated on this thread as being the most likely.

This must be the 1.2Ghz processor that Anandtech saw at CES, from post #36 above.
http://www.anandtech.com/tradeshows/showdoc.aspx?i=3719&p=2

[knux]

According to this site: http://pdadb.net/index.php?m=specs&id=2175&view=1&c=sony_ericsson_u8___u8i_se_kanna
The cortex-A8 in the omap 3 of kanna and kurara is clocked at 720 MHz and has 1024kiB of L2 cache. AFAIK the earlier Omap3's like Satio, N900, and i8910 had only 512kiB of L2 cache. Correct me if I'm wrong.

[roninja]

Kanna and Kurana (or U5 and U8 or Vivaz and ????) are based of the newer OMAP36xx platform most likely being the 3630 whereas the older handsets such as the Satio featured OMAP3430 with the lower L2 cache. The 36xx is also a die-shrunk 45nm implementation.

[wishiknew]

Man what chipset is in that just announced Samsung i8520. And is no way in hell it can run out of ram with 3 gigs. I have to have that projector phone.

[tangey]

Well, Samsung just announced a long term strategic deal with Mali for all future Socs, so looks like after the SGX540 based Socs, they'll be no SGX stuff from samsung.

http://eetimes.eu/uk/222900321

[roninja]

http://www.imgtec.com/corporate/newsdetail.asp?NewsID=499

"The Group signed software licenses and upgrades with a number of partners including Samsung"

I would not write-off the IMG/Samsung partnership just yet!