PowerVR - Eurasia

[amk]

Question would be if there's a difference between MBX and MBX PRO, apart from clockspeed that I'm not aware of. If the latter should be the case, then the 90nm derivative TI has been mentioned to be fabricating right now should exceed the claimed specs for MBX PRO on PowerVR's homesite.

I just went to find what PowerVR said about MBX Pro on their website ... they don't seem to mention it anymore.

IIRC Pro had better performance per MHz than MBX, about 50% better fill rate. It was aimed at set-top-boxes and (snigger) consoles. I'd always felt that the increase of power was somewhat pointless without a corresponding increase in features on which to use that power, e.g. Kyro's multitexturing, and either modifier volumes or stencil buffers.

The high-end consumer electronics for which Pro was targeted seems slap in the middle of Eurasia's stated target markets. MBX and Lite would still provide a cheaper alternative for mobile and integrated graphics.

 

[Ailuros]

Question would be if there's a difference between MBX and MBX PRO, apart from clockspeed that I'm not aware of. If the latter should be the case, then the 90nm derivative TI has been mentioned to be fabricating right now should exceed the claimed specs for MBX PRO on PowerVR's homesite.

I just went to find what PowerVR said about MBX Pro on their website ... they don't seem to mention it anymore.

IIRC Pro had better performance per MHz than MBX, about 50% better fill rate. It was aimed at set-top-boxes and (snigger) consoles. I'd always felt that the increase of power was somewhat pointless without a corresponding increase in features on which to use that power, e.g. Kyro's multitexturing, and either modifier volumes or stencil buffers.

The high-end consumer electronics for which Pro was targeted seems slap in the middle of Eurasia's stated target markets. MBX and Lite would still provide a cheaper alternative for mobile and integrated graphics.

PRO was claimed to have 750MPixels/sec (which I believe to be effective) fillrate.

The TI MBX@90nm I'm talking about will be most likely clocked at 200MHz, which gives a 1.0 GPixels/sec effective fillrate.

Or if you'd like floating point performance 800MFLOPs/sec from the VGP; now that's more than enough for a set-top-box or handheld console for today's requirements and until next generation Eurasia will be finally enter mass-production which isn't going to happen tomorrow either. Until then forementioned MBX variants will be supplying similar devices since that kind of performance and a target resolution of 1024*768 is way overkill for any current mobile phone.

PowerVR/IMG might follow past strategies and present in the foreseeable future Eurasia in a FPGA format and from that point and on it'll still take time until it'll reach final silicon state.

As far as KYRO's multitexturing goes, why on earth would they need 8-layer MT for in OGL anyway? The OGL2.0 ICD on current high end accelerators still sees only 4 MT units, unless I'm blind. Granted dual MT isn't something special, but Quake3A didn't really need more and I can't see a competitive core right now that could even run it with decent performance.

Vendors have damn limited die space for such cores and there are going to be inevitably shortcuts in quite a few departments even in the future.

 

[Ailuros]

Really is starting to look like a graphics core that can be used anywhere fro mobile to high end arcade megapower systems. Scalability has always something the PowerVR team have strived for.

I really wonder then why SEGA just didn't slap 50 MBX's together and call it a day.

Judging from Metcalfe's statements in past interviews there might be a possibility that they'll reach final silicon level with Eurasia this time, in order to probably speed up developer support and/or SDK releases of the future. In such a case it will most likely be a core that's optimized for PDA/mobile devices just like MBX is and I can't figure how it could even come close to what SEGA has already licensed.

If SEGA would want for the future more graphics power, they could easily follow the Naomi2 route and either build a dual-S5 board or bridge them for PCI-E and use a multi-board config.

 

[amk]

As far as KYRO's multitexturing goes, why on earth would they need 8-layer MT for in OGL anyway? The OGL2.0 ICD on current high end accelerators still sees only 4 MT units, unless I'm blind. Granted dual MT isn't something special, but Quake3A didn't really need more and I can't see a competitive core right now that could even run it with decent performance.

As I understand it, based on Kristof's comments on the pcstats B3D forums, Series 3's multitexturing tech allows arbitrary numbers of texture layers per pass, and is limited only by the API. D3D then allowed a maximum of eight.

 

[Ailuros]

As far as KYRO's multitexturing goes, why on earth would they need 8-layer MT for in OGL anyway? The OGL2.0 ICD on current high end accelerators still sees only 4 MT units, unless I'm blind. Granted dual MT isn't something special, but Quake3A didn't really need more and I can't see a competitive core right now that could even run it with decent performance.

As I understand it, based on Kristof's comments on the pcstats B3D forums, Series 3's multitexturing tech allows arbitrary numbers of texture layers per pass, and is limited only by the API. D3D then allowed a maximum of eight.

How arbitrary exactly? How about 256 layers of alpha blended textures? Anything theoretically "unlimited" hits sooner or later on a threshold, especially when you've only 350MPixel/Texel/sec fill-rate and 2.8GB/sec bandwidth (KYRO2).

KYRO's 8-layer multitexturing wasn't ever fully utilized in games for it's relative lifetime. By the time quad texturing became mainstream, way too many applications were T&L optimised and/or contained cube maps.

A geometry companion was a way more wiser design decision IMHO, than unnecessary multitexturing capabilities, when target applications at first aren't going to contain more than dual MT anyway.

 

[Dave B(TotalVR)]

KYRO's 8-layer multitexturing wasn't ever fully utilized in games for it's relative lifetime. By the time quad texturing became mainstream, way too many applications were T&L optimised and/or contained cube maps.

A geometry companion was a way more wiser design decision IMHO, than unnecessary multitexturing capabilities, when target applications at first aren't going to contain more than dual MT anyway.

Don't forget how much easier 8 layer mulktitexturing actually is to impoliment with PowerVR tech.

 

[Rob Evans]

I have discovered today that Eurasia is officially known within IMG as Series 5.

I presume that means that they have cannibalised the PC Series 5 that was ready to go by the middle of 2004, in the same way as the MBX was developed from Series 3.

Rob.

 

[TEXAN]

So what is going to power Lindbergh?

 

[Lazy8s]

In such a case, the revelation of Eurasia would be that it's just the name for PowerVR's very high-end architecture which is going into the Sega Sammy board among other products.

The semiconductor companies involved in the high-end Sega Sammy board are still undisclosed. With Intel already named as a Eurasia licensee, there's some chance it's them, but multiple other semiconductors have reportedly also shown interest in Eurasia and could easily be the one involved (among a series of choices, Intel probably wouldn't be the most likely.)

Intel would be a great partner to bring PowerVR to various markets, though.

 

[Ailuros]

I have discovered today that Eurasia is officially known within IMG as Series 5.

I presume that means that they have cannibalised the PC Series 5 that was ready to go by the middle of 2004, in the same way as the MBX was developed from Series 3.

Rob.

I believe the elements in MBX from Series3 are fewer than many would think.

I'd also speculate that Eurasia has more in common with generations past Series5.

 

[Ailuros]

So what is going to power Lindbergh?

Series5.

 

[Simon F on Hols]

I have discovered today that Eurasia is officially known within IMG as Series 5.

Was that using a crystal ball, tea leaves or something of a class B nature?

 

[linthat22]

MBX?? Bitboys tech or no?

 

[Lazy8s]

MBX is from PowerVR. Been shipping for a while.
http://www.powervr.com/Products/Graphics/MBX/index.asp

G30 and G40 are from Bitboys. Not yet shipping.
http://www.bitboys.com/g40.php

 

[linthat22]

MBX is from PowerVR. Been shipping for a while.
http://www.powervr.com/Products/Graphics/MBX/index.asp

G30 and G40 are from Bitboys. Not yet shipping.
http://www.bitboys.com/g40.php

Spanks for clearing that up. Here I was hoping for some Bitboys magic in something other than a phone.

 

[amk]

The TI MBX@90nm I'm talking about will be most likely clocked at 200MHz, which gives a 1.0 GPixels/sec effective fillrate.

Duh... how did you work that out?

Sega Sammy Aurora has a 150MHz MBX and a 150Mp/s fill rate, presumably raw (not "effective").

I estimate that MBX has a single pixel pipeline with a single TMU, looped back to give 2 textures per pass presumably in a simplified version of Kyro's loopback.

I expect MBX Pro has 2 pixel pipelines, also with 1 TMU each.

Overdraw of 3 or more assumed, giving effective fill rates of 360+Mp/s and 720+Mp/s (poss rounded to 750) respectively with clock rates of 120MHz, appropriate for 0.13u.

Are you assuming Pro is designed for higher clock speeds than MBX? If so, why? I've not seen any suggestion of that anywhere. Both are given the same clock rates for the same process on the PowerVR web site.

Synthesizable up to 80 MHz clock speed in 0.18µm and 120 MHz in 0.13µm.

http://www.powervr.com/Products/Graphics/MBX/index.asp
http://www.powervr.com/Products/Graphics/MBXPro/index.asp (Still there, just not linked)

That does however leave the question of MBX Lite. Perhaps 1 pipeline with no TMU so 2 clock cycles per pixel?

 

[Lazy8s]

5x overdraw may still be within the range, at the high end, of comparative average complexity for MBX level games.

 

[Ailuros]

Duh... how did you work that out?

200MHz * 2 TMUs = 400MPixels/sec * 2.5 overdraw = 1.0GPixels/s effective.

Sega Sammy Aurora has a 150MHz MBX and a 150Mp/s fill rate, presumably raw (not "effective").

On 130nm I think.

Same equation as above:

150MHz * 2 TMUs = 300MPixels/sec * 2.5 overdraw = 750MPixels/s

I estimate that MBX has a single pixel pipeline with a single TMU, looped back to give 2 textures per pass presumably in a simplified version of Kyro's loopback.

Sounds like it.

I expect MBX Pro has 2 pixel pipelines, also with 1 TMU each.

And why not 1*2?

Overdraw of 3 or more assumed, giving effective fill rates of 360+Mp/s and 720+Mp/s (poss rounded to 750) respectively with clock rates of 120MHz, appropriate for 0.13u.

I said 200MHz@90nm remember?

Are you assuming Pro is designed for higher clock speeds than MBX? If so, why? I've not seen any suggestion of that anywhere. Both are given the same clock rates for the same process on the PowerVR web site.

Then you might want to stop concentrating on fillrates exclusively on the performance pages of MBX and MBX PRO.

I doubt the quoted 1M Tris/s for MBX Lite are with 120MHz in mind.

That does however leave the question of MBX Lite. Perhaps 1 pipeline with no TMU so 2 clock cycles per pixel?

I saw a presentation where MBX Lite (sans VGP)@66MHz was claimed to run Q3A in 320*240*32 with 30Hz (vsynced).

If they ran demo001, they claim in one of their whitepapers an average overdraw of 3.39

320*240*32*3.39*30 = ~250 MPixels/sec

Compare the feature-sets of the 3 models; the only difference you'll see is that MBX-Lite doesn't claim "FSAA4free", while it's both there for MBX and MBX PRO.

MBX Lite is capable of 2x/4x ordered grid Supersampling, while on MBX/MBX PRO 4xAA is a combination of 2x vertical OGSS + 2x horizontal OGMS. Only 2xMSAA = FSAA4free.

 

[amk]

200MHz * 2 TMUs = 400MPixels/sec * 2.5 overdraw = 1.0GPixels/s effective.

Shouldn't that be 1.0GTexels/s and 500MPixels/s?

If we are both agreed that MBX is 1*1 then a 200MHz MBX would have a 500Mt/s fill rate. A 200MHz MBX Pro (1*2 or 2*1) would have a 1Gt/s fill rate, which would be OTT without something like 4+ multitexturing layers.

2.5 overdraw seems a bit low to me, though it is annoyingly arbitrary.

I expect MBX Pro has 2 pixel pipelines, also with 1 TMU each.

And why not 1*2?

If the texel fill rate were higher than the pixel fill rate, I'd expect IMG to talk about Mt/s rather than Mp/s.

320*240*32*3.39*30 = ~250 MPixels/sec

Are you assuming 32 texture layers? :?

I make it 320[x]*240[y]*30[fps]*2[MT]*3.39[overdraw] = a hilarious ~15Mtexels/s. Without the MT, an even more hilarious ~7.5Mpixels/s.

Either I've had way too little sleep, or that demo was not fill rate limited.

For that matter, 640[x]*480[y]*60[fps]*2[MT]*4[Kyro's overdraw estimate] = ~150Mt/s. MBX has fill rate to burn. Which is good, because trilinear filtering and y-axis FSAA both halve available fill rate, and dot3 (afaics) requires two passes. So with 4x FSAA and trilinear, we would need 600Mt/s, and more if we use dot3.

Furthermore, not using overdraw or effective fill rates and using 16x:9y :

~200Mt/s = 850[x]*480[y]*60[fps]*2[MT]*2[FSAA]*2[trilinear OR dot3]

Isn't that handy? Then we just need to ensure that %age of pixels using trilinear + %age of pixels using dot3 < 100

or

~200Mt/s = 600[x]*340[y]*60[fps]*2[MT]*2[FSAA]*2[trilinear]*2[dot3]

For dot3 - tbh I have no real idea how many clock cycles it would take.

Obviously the second pass hurts our polygon throughput.

This screenshot from the Nokia presentation is 600*450. Possibly it was cropped or resized, as the concept phone shown had a 16:9 screen.

Compare the feature-sets of the 3 models; the only difference you'll see is that MBX-Lite doesn't claim "FSAA4free", while it's both there for MBX and MBX PRO.

Good point, well spotted. I had assumed the feature sets were the same.

 

[amk]

For comparison, raw pixel fill rates for previous PowerVR products are:

Dreamcast:   100Mp/s
Kyro:        230Mp/s
Kyro II:     350MP/s

200Mp/s on a mobile is starting to look generous.

AFAIK all above products are 1 TMU per pipe. Kyro has free trilinear, MBX free 2x FSAA. DC has neither.

Oh, and...

Then you might want to stop concentrating on fillrates exclusively on the performance pages of MBX and MBX PRO.

I doubt the quoted 1M Tris/s for MBX Lite are with 120MHz in mind.

I'd expect PowerVR to quote the highest figures they could. Giving performance figures for e.g. 60MHz when you've stated elsewhere the core can run at 120MHz isn't my idea of aggressive marketing.

I do expect the 1M tri/s for MBX lite is for 120MHz. And I still think its 1*0