PowerVR - Eurasia

They're quoting MegaPixels aren't they? The MBX PRO was claimed in their own link to be able to reach 750MPixels/s.

I'm as sure as I can be that all models have only one pipeline.

The feature-set for MBX PRO also states dual MT.

If we'd be picking needles out of straws, they shouldn't use effective fillrates either then. They didn't at least for Series3.

32bpp.

Or you forgot to encount colour depth maybe.

Holy smokes what an awkward compilation....anyway MBX might be quite powerful for the PDA/mobile market, but it for sure hasn't fill-rate to spare. Fill-rate and/or bandwidth are amongst those things an accelerator can never have enough of.

If you manage to give me a half way reasonable explanation what on God's green earth a chip could render with 1*0, apart from Z's I'd be happy to hear it.


From www.mitrax.de:



That's on a Dell Axim X50v; I'm not sure anymore what the clockspeed of MBX in that thing is, but I doubt it's any higher than 65-70 MHz.

INTEL's 2700G comes in two flavours according to their whitepapers; 2700G3 with 384kb and 2700G5 with 705kb on-die memory. If it would have had a VGP, on-die memory would had been slightly larger since the latter gets it's own on-die memory too. Both on 130nm.

Found it:

It says up to 75MHz; I might be wrong but I think it's clocked at 67MHz in the Dell Axim X50v.

ftp://download.intel.com/design/pca/companion/datashts/30094801.pdf

Look at it that way: there are 2.5M Tris/sec claimed for MBX and 3.0M Tris/sec for MBX PRO. Now assuming MBX is at 120MHz, what would you assume does it need to reach 3.0M Tris/sec? Something like 150MHz maybe?

At 200MHz an MBX will have quite a bit higher triangle throughput than a Dreamcast itself; factor in the VGP (4 MFLOPs/s per MHz = 4*200 = 800MFLOPs/sec) and it's VS1.1 capabilities and you're quite a bit ahead of DC.

KYRO didn't have free trilinear; it did have normal trilinear (2 clocks) which did cost quite a bit in performance and "fast trilinear" which worked only with texture compression enabled (16 texel fetches from a single MIPmap).

On MBX 2xMultisampling is "free"; with 2x Supersampling it'll lose half it's fillrate like any other accelerator out there. Dreamcast supported 2xSSAA if my memory shouldn't betray me.

 

I don't think so. It should be irrelevant. Processing each bit isn't going to take a whole clock cycle.

e.g. 1024*768*32*60*2*3.39=10.2Gt/s
10.2G/32=320Mt/s, which is more like it.

I got carried away.

Sure you can have excess bandwidth. There's a limit to how much an accelerator will ever require, beyond that extra is pointless.

With fill rate, you could just run out of really useful things to do with it. You could use it for trilinear and anisotropic filtering, but adding more textures using multipass is going to hit your polygon performance.

IIRC the PS2's "graphics synthesizer" uses n*0 pipelines, I forget how many. They take 2 clock cycles to fill a texel. PS2's texel fill rate is half its pixel fill rate.

How they work I do not know.

Ah, contradictory sources. There could be differences of implementation too, so I'd take that with a pinch of salt.

Still, 75MHz -> 84 Mp/s? Perhaps 75M/2 (actual) * 2.24 (overdraw) = 84M. This would technically be texel fill rate though.

I think you're right.

I'm starting to suspect that an MBX gaming handheld won't be clocked at 200MHz - more likely 150MHz or less. Faster clocks would still give useful polygon increases though. 200MHz is being rumoured around the web though for N-Gage 2, not that that means anything.

IIRC fast trilinear was free. I may be wrong.

IIRC it was less simple than that. It rendered internally to a higher resolution, and then filtered it down - but the internal resolution was not a whole number multiple of the output.

 

amk:

16 pipes. Perhaps they're general pixel units which take on the TMU job with half of the units during texturing.

 

Trilinear takes two cycles and I'd suspect AF to take four.

Lazy8 covered it already; besides I'm not sure if those are even capable of bilinear or if they have to combine units to even reach bilinear.

What sources? It's one whitepaper from INTEL bother to read it. Mobile phones for now aren't going to exceed 70-80MHz maximum irrelevant of device or architecture implemented, due to power consumption restrictions. Whether it's 800K or 1M is the exact problem or that there are other competitors out there that claim 1M tris and their devboards have a hard time exceeding the 60k Tri rate? (yes that's 60.000 and no typo). Triangle rate for MBX Lite is measured with 80MHz in mind IMHO.

150MHz are possible in 130nm; naturally 90nm will secure the increase to 200MHz and for TI at least I consider it given.

Yes but with texture compression enabled.

LOL technically it's still a form of Supersampling, and as such it was claimed in DC specifications.

4xFSAA on >MBX Lite means 2x vertical SSAA + 2x horizontal MSAA. While you get the latter fillrate free, the first will lose inevitably half the fillrate. The SSAA part will filter textures "2xAF alike" even if SS is combined with bilinear. MSAA by itself is pretty useless IMO, if you don't have fast performing AF to combine it with.

 

Depends how much jagged polygon edges bother you I suppose.

MSAA will increase your IQ and if its free like on the MBX then of course its useful!

 

it could be that mbx's triangle setup is limited to 1M TpS but is limited to 831K because of transform. the DS is in a reversed situation. IIRC it can transform 2-3M TpS but is limited to 120K because of setup.

 

If I'd have a strict choice between either "edge" or "texture antialiasing", I'd rather pick the latter first.

Since exactly 2xMSAA is for free, 4x FSAA and thus 2xMSAA +2xSSAA shouldn't be unfeasable; that way you get at least "2xAF alike" texture filtering on the vertical axis and it'll cost way less in performance than 2xAF on MBX.

 

Using a better filtering algorithm is often a better choice though than just SS'ing everything. What I'd have liked is 4xMSAA with a decent anisotropic filter.

 

Definitely; considering the limited die space though and the according restrictions I doubt there was much headroom for a fast adaptive AF algorithm (even just limited to 2x samples).

OGSS in comparison is relatively cheap; did you ever use AF on KYROs or at least 1*2 OGSS instead?

This far only Falanx seems to have (single-cycle) 4xRGMS on it's Mali cores (their pdf mentions point, bilinear and trilinear texturing and anisotropic mapping - whatever that stands for) and Bitboys have licensed FLIPQUAD (2x sample algorithm delivering better quality than Quincunx) from Sony/Ericcson (no sign of anything anisotropic there too even for G34/G40).

I doubt anyone would expect any form of antialiasing on either Imageon or GoForce.

Considering all the above having a hybrid SS/MSAA 4x sample option (2*2 EER, supersampled textures on one axis), I'd say it's a healthy compensation based on the possibilities within die space/power consumption restrictions.

 

OGSS is cheap in terms of die area but it slows you down a helluva lot. especially on smaller hardware like the KYRO. AS for AF, there was no adaptive algorithm on the KYRO, you just AF'ed every single pixel making it take twice as long to render.

 

has it actually been confirmed that Eurasia is Series 5 or based on Series 5?


or perhaps it is a combination of Series 5 (current PowerVR tech) and
Series 6 (future PowerVR tech) in a similar way that MBX used elements of Series 3 and the cancelled Series 4.

i'm confused though.

 

IMG have clarified the position re Eurasia:

Originally posted on the iii IMG board - http://www.iii.co.uk/investment/det....L&it=le&action=detail&id=2794662

 

And what makes you think that an adaptive AF algorithm is cheap to implement in hardware, even more on a small embedded chip?

2xAA didn't cost more in performance than 2xAF did on KYRO.

 

Ever read your PM box? (watch the date of the message).

 

If Eurasia has benefited from the time after Series 5 was completed, has Sega Sammy's implementation of Series 5 been upgraded at all from that point considering the timing of its release later this year?

If Eurasia could be implemented into power sensitive segments even lower than laptops, it might find the balance between robustness and conservativeness that makers of portable gaming consoles like Nintendo need, assuming that it overlaps with MBX while not being worse for the job. Potential customers could be the aforementioned Nintendo, Sony, Microsoft, Tiger, perhaps Sega Sammy, or some other company willing to try.

Whatever's happening with Eurasia, it seems positioned, along with the rapid adoption of MBX, to really boost PowerVR's presence in the graphics market.

 

random thoughts:

*imagines a PowerVR Eurasia based GBA successor. whoa, that would be amazing and would truly blow the living snot out of PSP 8)

*does ATI have anything on the horizon that can compete with Eurasia in the power-sensitive handheld market?

*It will be interesting to learn the differences between the circa 2004 Series 5 and the 2005-2006 Eurasia implementation.

*I don't expect to see anything on Series 6 until late in this decade

 

I didn't say that;P

Now that I think about it it isn't and neither is MSAA (especially not).

2xAA cost more than the single 16 tap anisotropic mode you had on KYRO performance wise, yes - but 2xAA also looked gash on the KYRO. perhaps if they had implimented a better AA algorithm other than just rendering it twice as big and downsampling it would hve looked better. Saying that, though the 4x look schweet. Perhaps if they had given us a whole quad on the card then it would have been usable at decent resolutions (who am I kidding a 4 piped KYRO running at 175Mhz with fast enough RAM would have been a absolute killer card).

Im still waiting for the powervr pc part that has virtually free MSAA up to 4x4 (16x). They certainly could do it and combine that with an adaptive algorithm for anisotropic filtering and you've got yourself a serious advantage Iq wise. IMG have been very consistant with their pc market approach though as you are aware... LALLALA I CANT HEAR YOU!!

As for eurasia, well it sucks to wait until h1 2006 (or 2010 whichever way you look at it ) but to say its based on series 5 and has been taken a long way developed more and made more scalable...

well that would certainly give ATi or Nvidia grounds to call it a next generation design.... Maybe they don't wanna call it series 6 because they think it is too similar.

 

I doubt it.

Obviously more than just graphics considering the aggressive entry lately in the mobile TV market recently as just one example. There was a past presentation (November 2004 I think) that had 3 separate circles for current IP: Graphics/Video Display, Processor & DSP, Comms & Broadcast. Future prospects then mentioned high levels of programmability and one single circle incorporating all three. That's IMHO the level of scalability and configurability that's being mentioned here and it'll be able to address obviously more than one particular market.

 

Any bets you're willing to run?

-----------------------------------------------------------------------

DaveB(TotalVR),

Such a model actually existed during the ST Micro cooperation, codenamed STG5000. It is my understanding that the 150nm/KYRO3@~166MHz was canned because ST (alone?) felt that the price/performance ratio of a KYRO@175MHz is better. It would had obviously cost quite a bit more than the K2, but considering the double amount of pipelines, the DDR support and very fast T&L processor, it would have had not only past double the performance of K2, but also a way longer lifetime.

I believe that by the time the decision for releasing a speed bumped KYRO@175MHz fell, that it was then decided to manufacture K3@130nm (STG5500) at 250MHz and beyond. The rest of the flop is well known.

PM.

 

That would be quite amazing. A Series 5 portable using SSR enabled 48MB 1-TSRAM, new Sharp 4.3" widescreen LCD, Matrix ROMs, coupled to a dual core ARM/NEC MPCore.