PowerVR Athena ? ....a few thoughts

[Megadrive1988]

from everything I've gathered, which isn't very much, the Athena GPU is currently the most powerful PowerVR chip in existance, as far as what anyone on the outside / the public knows of, and not counting future PowerVR architectures in development. Athena is meant to be a desktop Series 5 chip. It is said to have "8 pipes", which could mean several different things, since what is now concidered a "pipeline" in the modern, shader-GPU era, is foggy.

Athena supposedly outperforms the current highest-end ATI (R580) and Nvidia (G71) GPUs by about 50%

I am wondering if Athena is an 8 "pipeline" SGX or if Athena has a structure that is completely seperate & different from SGX, the mobile Series 5 GPU family.

IIRC, the PowerVR MBX was a combination of Series 3 and (unreleased for desktop) Series 4 technology.

It sounds like the soonest Athena could come out is, around the middle of 2007. by then not only will the highend G80 and R600 be out, but also the entire range of DX10 GPUs.

ImgTech will probably not try to compete with the $500~$600 highend cards, but maybe with the $300~$400 "performance" cards, and on down.

of course, we've been through this type of speculation before, when Series 4 never panned out.

discuss

 

[_xxx_]

Substantial facts I have heard of until now:

1) nothing
2) nothing
3) nothing at all

 

[zsouthboy]

Wake me when I can buy a card to put in my computer.

 

[Ailuros]

Athena supposedly outperforms the current highest-end ATI (R580) and Nvidia (G71) GPUs by about 50%

The rumour said that it outperforms other "8-pipe GPUs" by about 50%. R580 as well as G71 are far and beyond any weird conception of any "8 piper".

 

[Lazy8s]

SGX is the name of the entire family that spans from portable up to desktop. The Series 5 generation has been implemented at least three times as different architectures, the latest one being called first with the code name of Eurasia and finally with the official name of SGX.

The "8 pipeline" configuration is the high-end of the portable line (SGX530) or nearing the next markets up in performance, in-car nav/info systems and ultra-mobile PCs.

 

[Rootax]

The rumour said that it outperforms other "8-pipe GPUs" by about 50%. R580 as well as G71 are far and beyond any weird conception of any "8 piper".

If I remember the source, it was not very clear.

 

[Mummy]

Elvis Presley

lives.

 

[Ailuros]

SGX is the name of the entire family that spans from portable up to desktop. The Series 5 generation has been implemented at least three times as different architectures, the latest one being called first with the code name of Eurasia and finally with the official name of SGX.

The roadmap paints in their last AGM presentation shows clearly a separation between PDA/mobile cores (SGX 510/520/530) and other higher ranged cores for other markets (Thalia-L, Muse, Athena).

The "8 pipeline" configuration is the high-end of the portable line (SGX530) or nearing the next markets up in performance, in-car nav/info systems and ultra-mobile PCs.

SGX 530 is according to their own claims sized 8 mm^2 and yields 1.2 MPixels/s effective fillrate and 13M Tris/s at 200MHz. That's in my interpretation 2 TMUs * 200MHz * 3.0 Overdraw. That's not only extremely underwhelming for a PC core (yes even ultra-mobile), but doesn't also follow the roadmap's reasoning. Athena is two whole steps higher than SGX530, which is already two steps higher than SGX510 and I wouldn't be surprised if the latter follows a similar reasoning as MBX Lite compared to MBX does.

Of course will it come down to what anyone means with "pipeline" in a unified shader core, yet I have a hard imagining how any unit could count "8" in such a core, apart from Z/stencil units which are IMHO too few being at 8 even on SGX530. However through some speculative math I could under circumstances calculate 8x times the floating point throughput compared to SGX510; not necessarily 8x times the units though.

Athena should have no less than 8 TMUs IMHO, if it should yield to compete with today's mainstream offerings, but that's of course just me.

 

[Lazy8s]

The other target markets -- Athena, Muse, Thalia-L -- lack the "SGX 5xx" header probably because they're just working titles for products which haven't been announced. They seem to be sub-families of the overall brand, SGX.

The higher performance markets would obviously have a lot more "pipelines" than the wireless graphics accelerator family was quoted to offer.

The size of SGX530 at 90nm might turn out to be just under the guideline at about 7.8 mm^2.

 

[Megadrive1988]

SGX is the name of the entire family that spans from portable up to desktop. The Series 5 generation has been implemented at least three times as different architectures, the latest one being called first with the code name of Eurasia and finally with the official name of SGX.

The "8 pipeline" configuration is the high-end of the portable line (SGX530) or nearing the next markets up in performance, in-car nav/info systems and ultra-mobile PCs.

I'm kinda confused by your 2nd sentence.

It seems that Athena is not part of the SGX mobile family, even though they're all Series 5.

edit: well maybe not, going by that roadmap, SGX could be the official name for the entire Series 5 family or families of GPUs. but certainly a desktop graphics processor is not going to have the same architectural, power(electricity) and heat restictions as even the highest-end mobile graphics processor.

 

[Dave B(TotalVR)]

All I can say is they had better give it some decent raw power this time around. Sounds to me like Athena is a more powerful chip than the 530, and muse seems like the mobile version, for laptops and shit.

8 pipes in the SGX 530 almost certainly refers to 8 shader units IMO, configurable as either pixel or vertex shaders. Not unreasonable for something that's gonna be in a PDA and places. Not good enough for the desktop though, clearly.

They need to get off their rear end and release something with similar raw specs as ATi(AMD) and Nvidia, if they do that they will wipe the floor with them performance wise for the same price or mayve even less. Yeah I know it's been said a million times, but maybe if I say it just one more time they'll listen

 

[_xxx_]

I think if you're ready to finance that, they'll listen definitely

 

[Ailuros]

8 pipes in the SGX 530 almost certainly refers to 8 shader units IMO, configurable as either pixel or vertex shaders. Not unreasonable for something that's gonna be in a PDA and places. Not good enough for the desktop though, clearly.

8 ALUs in a high end chip for PDA/mobiles? Sounds way too much IMO.

According to IMG's claims the SGX510 yields 200MPixels (call them MegaTexels for accuracy's sake heh) effective fillrate at 200MHz, while the SGX530 has 1200MPixels@200MHz.

530 sounds here to me like a chip with 2 TMUs * 3.0 Overdraw * 200MHz = 1200. SGX510 following that logic an only be a single TMU chip, whereby it's of course not a "half" TMU, but most likely the chip can handle less pixels (just as MBX Lite compared to MBX) at maximum overdraw. On the SGX530 you obviously not only have higher capabilities, yet also twice the functional units and with some backwards speculative math it shouldn't be too hard to find a factor of 8x times difference between 510 and 530.

Say for example (and yes it's pure speculation) you have for SGX510 one ALU capable of 4 FLOPs per cycle, that gives at 200MHz = 800 FLOPs. If you now count for SGX530 2 ALUs capable of 16 FLOPs per cycle you get at 200MHz = 6.4 GFLOPs. Numbers could of course be higher than that; it's as I said just an example. But you can get easily a factor of "8x" even there.

I expect the ballpark to be a lot higher for something like Athena. If it's something targeting the mainstream market it cannot have less than 8 TMUs IMHO. And that's already a factor of 4x times higher than SGX530 in terms of TMUs alone.

 

[mboeller]

According to IMG's claims the SGX510 yields 200MPixels (call them MegaTexels for accuracy's sake heh) effective fillrate at 200MHz, while the SGX530 has 1200MPixels@200MHz.

530 sounds here to me like a chip with 2 TMUs * 3.0 Overdraw * 200MHz = 1200.

Why the factor of 3? PVR itself quoted only a factor of 1.25 to 1.33 back then in ancient times ( PVRSG, Kyro ).

Also the SGX530 is more than 4times bigger than the SGX510. therefore the SGX530 could very well have more than 4times the TMUs and PS/VS power. I even expect that the SGX530 has 8times the PS and VS power of the SGX510 because the VS numbers quoted below show a improvement of 6.75 times.

PowerVR SGX core architecture is fully scaleable and the Wireless Graphics Accelerator family includes 3 variants: SGX510, SGX520, SGX530 with sizes ranging from less than 2mm2 to 8mm2 in a 90nm process.

Performance
Maximum effective pixel fillrate performance from 200Mpix/sec to 1200Mpix/sec @ 200MHz with even higher Z and stencil fill rate and polygon throughput from 2Mpoly/sec to 13.5Mpoly/sec @ 200MHz. Performance depends on core and configuration selected.

So I expect that the SGX 530 has 4 TMUs and at least 8 PS/VS - ALUs @ 200MHz.

 

[chavvdarrr]

I wonder how many connector pins one can put on 8mm2 die ...

 

[_xxx_]

Zero. You mean pads, not pins

 

[amk]

Why the factor of 3? PVR itself quoted only a factor of 1.25 to 1.33 back then in ancient times ( PVRSG, Kyro ).

IIRC IMG claimed that Quake 3 demo001 had an average overdraw of 3.8. I think the marketing bumph for Kyro assumed an overdraw factor of at least 3.0. Obviously it was in their interests to err on the high side.

 

[Ailuros]

Why the factor of 3? PVR itself quoted only a factor of 1.25 to 1.33 back then in ancient times ( PVRSG, Kyro ).

KYRO was rarely marketed with effective fillrates from what I recall, but when it definitely wasn't as low as just 1.33. MBX fillrates were probably also following a something around 3.0 trend and it's not an unrealistic factor either.

Also the SGX530 is more than 4times bigger than the SGX510. therefore the SGX530 could very well have more than 4times the TMUs and PS/VS power. I even expect that the SGX530 has 8times the PS and VS power of the SGX510 because the VS numbers quoted below show a improvement of 6.75 times.

See above; I'm judging things from MBX/MBX Lite differences. I'm fairly confident that the trend here is the same with the 530 being a way too large MBX PRO equivalent, which by the way never found it's way into any device probably because it was too large and partners didn't have any particular interest. IMHO MBX Lite/MBX are 1*1 designs and the canned MBX PRO a 2*2 design. Any particular reason why it should be different on SGX? (disregarding the way more advanced USC design of SGX of course).

As for higher triangle throughputs, look closer and again at my speculative math above an think USC, USC, USC

So I expect that the SGX 530 has 4 TMUs and at least 8 PS/VS - ALUs @ 200MHz.

ROFL you optimist

I'd say it's stencil fillrate is 6.4 GPixels/s though

 

[Blitzkrieg]

Why do they bother developing high end cores(even if just on paper) if they are never going to release one? Seems like pissing money up against a wall to me.

 

[bloodbob]

Lets do some really bad extrapolation.

R580=352mm^2

352/8=44

44*1200mpixels=52800 mpixels mm that would be some pretty powerful raw fillrate. Roughly 5 times more. Now however there are 3 shader units per pipe on the 580 if we assume the SGX had a 1:1 ratio and we assumed roughly the same efficiency the souped up SGX would be about 66% faster.