OMAP4 & SGX540

[Ailuros]

Dunno if this is exactly the right thread, but its close enough.

I see an Omap4 samsung has hit Glbenchmark, GT-I9108
gets a decent 4246 score for egypt standard, and about 50% improvment in the egypt of-screen score compared to the SGX540 in the galaxy S.

I'm not aware of an announced 840x480 sammy using omap4 ?

Which OMAP4 though?

***edit: I just saw the first Archos 80G9 results are up

 

[tangey]

TI put up the entire specification for Omap4470.

http://focus.ti.com/pdfs/wtbu/OMAP4470_ES1.0_PUBLIC_TRM_vC.zip

confirms SGX544-SC as the graphics core. Still working thru the various clock definitions,dividers and gates to determine what the clock for it is, anyone else want to have a go, feel free !

Casting a glance over the diagrams shows that the SGX544 is connected to the rest of the Soc via DUAL 128-bit master buses. On the 4460, SGX540 was connected via a single 128-bit master bus.

Also the introduction in the datasheet confirms that the 2D composition engine is a Vivante GC320. The entire section for it is blank however.

 

[Rys]

Clock is 384 MHz for the SGX in 4470 (TI confirmed to Anandtech a while ago).

 

[tangey]

Clock is 384 MHz for the SGX in 4470 (TI confirmed to Anandtech a while ago).

So theoretically, the core has equivalent raw performance to an iphone4 (SGX543MP2@200Mhz), plus the enhanced compliance.

 

[Ailuros]

So theoretically, the core has equivalent raw performance to an iphone4 (SGX543@200Mhz), plus the enhanced compliance.

If the 544 in the 4470 is truly clocked only at 384MHz, then it should it have some additional "surrounding" stuff (like for instance more generous caches, more bandwidth etc) compared to the SGX540@304MHz in the 4430 for the 2.5x times performance increase TI claims for the 544 against the former.

 

[tangey]

If the 544 in the 4470 is truly clocked only at 384MHz, then it should it have some additional "surrounding" stuff (like for instance more generous caches, more bandwidth etc) compared to the SGX540@304MHz in the 4430 for the 2.5x times performance increase TI claims for the 544 against the former.

I assume the doubling of the master bus to DUAL 128-bits is significant in enhancing the bandwidth.

 

[mczak]

If the 544 in the 4470 is truly clocked only at 384MHz, then it should it have some additional "surrounding" stuff (like for instance more generous caches, more bandwidth etc) compared to the SGX540@304MHz in the 4430 for the 2.5x times performance increase TI claims for the 544 against the former.

Either that (could be given it's Series 5 540 vs 5XT 544 I've no idea) or they are just going by the shader flops number...

 

[Ailuros]

Either that (could be given it's Series 5 540 vs 5XT 544 I've no idea) or they are just going by the shader flops number...

Nope. The floating point value is damn close to a 3x factor and not just 2.5x. Given that texel fill-rate increases only by roughly 26% I'll just call that 2.5x claim a pure marketing claim. They all do it why shouldn't TI after all?

 

[mczak]

Nope. The floating point value is damn close to a 3x factor and not just 2.5x.

I think that depends if you count some special ops. Pure mad rate is almost exactly a factor of 2.5 at these clocks.
It is quite possible the number is just marketing though.

 

[almighty]

Off topic but one thing I've always wondered is what console these GPU's compare to.

I know you can't do a pure and direct comparison but are we at PS2/Dreamcast level?

 

[Ailuros]

I think that depends if you count some special ops. Pure mad rate is almost exactly a factor of 2.5 at these clocks.
It is quite possible the number is just marketing though.

Peak theoretical numbers vs. peak theoretical numbers. You're going to get 2 FP32/ALU on a SGX540 also only under specific conditionals. The increase in throughput should depend on the shaders used themselves.

I doubt any embedded application is going to use FP32 throughout; there's a reason why OGL_ES splits up into highp, mediump and lowp. According to their whitepapers each USSE ALU is capable also of Vec2 FP16 and Vec3 or Vec4 INT8, while operating under FP32 as scalar units. No idea how it scales down from 4*FP32 on USEE2 ALUs.

Off topic but one thing I've always wondered is what console these GPU's compare to.

I know you can't do a pure and direct comparison but are we at PS2/Dreamcast level?

No idea about PS2 but Dreamcast level was reached even with >100MHz MBX and even there should be higher capabilities included than in the Dreamcast chip.

 

[Aeoniss]

Off topic but one thing I've always wondered is what console these GPU's compare to.

I know you can't do a pure and direct comparison but are we at PS2/Dreamcast level?

It seems to me we are well and far beyond those levels. We're also def past Xbox 1.

 

[Arun]

It seems to me we are well and far beyond those levels. We're also def past Xbox 1.

Agreed, although to be fair we're not past the XBox 1 in every way, since NV2A had 4 pixel pipelines with 2 TMUs each at 233MHz. Both Mali-400MP4 and SGX543MP2 have 4 pixel pipelines with 1 TMU each, and in their highest volume implementations (iPad 2 & SGS2) run at 250 and 266MHz so barely more than the XBox1.

Obviously in practice both are massively beyond the XBox1 in most ways (ALU performance, Z rejection, programmability, and even memory bandwidth) and in fact the NV2A's 2 TMUs per pipeline weren't easy to use in bilinear iirc (and that's if you even had enough bandwidth). The one exception is Vertex Shader where Mali-400MP4 is obviously much slower than the NV2A (which is somewhat mitigated by more computation moving to the pixel shader).

 

[Ailuros]

Both Mali-400MP4 and SGX543MP2 have 4 pixel pipelines with 1 TMU each, and in their highest volume implementations (iPad 2 & SGS2) run at 250 and 266MHz so barely more than the XBox1.

Please correct me if I'm wrong, but the Mali400MP4 should be 4Vec4 PS + 2 Vec2 VS, 1 TMU per fragment processor and the SGX543MP2 8 Vec4 USC ALUs, 4 TMUs ie 2 TMUs/core.

 

[tangey]

I wonder is this a 4470 ?

http://www.glbenchmark.com/phonedetails.jsp?D=Vizio+VTAB3010&benchmark=glpro21

Its SGX540, getting 64FPS in pro 720p, and 38fps in egypt offscreen. Both are by far the highest figures we've seen for SGX540.

 

[Rys]

4470 is 544, not 540.

 

[tangey]

4470 is 544, not 540.

Yes, of course it is. My bad.

Initially thought it was SGX540 @384Mhz (i.e. Omap4460)

But a look at the system info confirms its medfield

AS I recall medfield was announced as running @400Mhz

Reposted on the Z600 thread

 

[Lazy8s]

Archos's G9 Turbo line, released to Europe and now also to the US, clocks its OMAP4460's CPU to the full specced speed of 1.5 GHz, presumably bringing the SGX540 back up to 384 MHz.

Although always different from an OMAP4430, this puts the difference back to the intended amount between the two steppings of OMAP4.

 

[Rys]

I have an Archos 101 G9 Turbo here with 4460 that still uses the same GPU clock as 4430. Buyer beware with that model I think, since at least on the UK site it still lists 1.2 GHz.

 

[Deleted member 13524]

I have an Archos 101 G9 Turbo here with 4460 that still uses the same GPU clock as 4430. Buyer beware with that model I think, since at least on the UK site it still lists 1.2 GHz.

I think there are two "turbos". The one with the pre-loaded ICS ROM is claimed to be clocked @ 1.5GHz, the others are just 20% higher clocked than the vanilla version.

They are also saying the 1.5GHz ICS version has 1GB of RAM (the others only have 512MB), putting it confortably above any OMAP4 device in the market so far.

Quite the powerhouse, for a "low-cost" tablet. Both the 80G9 and 101 G9.