ltcommander.data
Regular
Revisionist history?
Series6, Series6XT and Series7XT GPUs are all based on the Rogue architecture. Series6 and Series6XT go up to 10, while Series7XT can handle 11 with the feature pack. Sounds okay to me?
ltcommander.data
Regular
http://blog.imgtec.com/powervr/powervr-g6630-go-fast-or-go-home
It's more the earlier Series6 releases where it reads like certain Series6 family members support DX11 rather than some members of the Rogue architecture in general support DX11. But I guess the separation between Series6 and Series7 might not even have been fully defined back then.
Splitting up "Rogue" as an architecture into more than one generations is clearly a marketing decision. I wouldn't be at all surprised if it reaches to Series8 also. If you add things up the difference between Series6 and Series7XT cores can go up to >+100%. Either perspective you take there seems to be a LOT of headroom to scale performance for Rogue. What marketing then decides to call each is another story. If you backflash and compare a SGX530@250MHz and a SGX554MP4@250MHz you get a difference of 32x times in GFLOPs as one singled out value.
By the way I assume from Ryan's text:
http://www.anandtech.com/show/8706/imagination-announces-powervr-series7-gpus-series7xt-series7xe
....that if FP64 is requested, it gets added as dedicated FP64 units and it's not that existing ALUs get used for 1:2. If you'd tell me it's the latter it would build quite a design philosophy oxymoron in my mind compared to the FP32/FP16 ALUs story.
By the way I assume from Ryan's text:
http://www.anandtech.com/show/8706/imagination-announces-powervr-series7-gpus-series7xt-series7xe
....that if FP64 is requested, it gets added as dedicated FP64 units and it's not that existing ALUs get used for 1:2. If you'd tell me it's the latter it would build quite a design philosophy oxymoron in my mind compared to the FP32/FP16 ALUs story.
Last edited:
The Anandtech article added some nice tidbits on the implementation, relevance, and reasoning for inclusion of Series7's new features. I'm glad it got coverage.
The priority on improving efficiency while adding some additional scalability has the potential to make Series7 a solid update.
The priority on improving efficiency while adding some additional scalability has the potential to make Series7 a solid update.
SA has also a writeup for it: http://semiaccurate.com/2014/11/11/long-look-new-imagination-7-series-gpus/
ok!
I am always happy about compute improvements. And 16 bit integers are very useful. For example 2d buffer lookups never require more than 16 bit coordinates. I wish AMD and Nvidia follow the suit. Double rate 16 bit integers would be fun (and only take half a GPR each).
IIRC SGX didn't support Vec2 INT16, only Vec2 FP16 and Vec4 INT8. Either way it's a new pipeline
There are a few things slightly wrong in the article but I'm not sure what I can say publicly. One thing I can definitely say because it's obvious from the diagram: the "Image Processing Data Master" uses the USC and texture units for processing. This means it's actually *less* fixed-function than the original Series6 which had a dedicated 2D accelerator. It is lower area and its performance scales 'for free' with the number of USCs, while being more efficient than using the pixel data master *and* being able to run in parallel to an unrelated 3D task.
So 7XTP can run multiple programs of 4 different kinds (vertex, pixel, compute, image) at the same time on the same USC.
Arun & sebbi,
Nothing against that at all, au contraire. I was (and am probably not the only one....) expecting a few details about Series8 mostly. Unless of course marketing has changed its mind and skipped the per year "generation" numbeing scheme after all.
Anyway I know it's probably hard if not impossible but on average how much more efficient could one rate a 7400+ compared to a 7400 f.e. and why have there been released only two IP variants of it?
Nothing against that at all, au contraire. I was (and am probably not the only one....) expecting a few details about Series8 mostly. Unless of course marketing has changed its mind and skipped the per year "generation" numbeing scheme after all.
Anyway I know it's probably hard if not impossible but on average how much more efficient could one rate a 7400+ compared to a 7400 f.e. and why have there been released only two IP variants of it?
Similar threads
