Have you ever wondered how a mobile GPU is born? Or how the architecture of a mobile GPU like ARM Mali differs from the technology in your discrete PC graphics card? Perhaps you just want to know if ideas like HBM (high bandwidth memory) are going to find their way into the mobile ecosystem any time soon?
Josh and I sat down (virtually) with ARM's VP of Technology and Fellow, Jem Davies, to answer these questions and quite a bit more. The resulting interview will shed light on the design process of a mobile GPU, how you get the most out of an SoC that measures power by the milliwatt, what the world of mobile benchmarking needs to do to clean up its act and quite a bit more.
You'd be hard pressed to find a better way to spend the next hour of your day as you will without a doubt walk away more informed about the world of smartphones, tablets and GPUs.
You almost have to ask the
You almost have to ask the questions to ARM Holdings’ licensees to find out if they are going to be using HBM, as ARM holdings may or may not have the resources that some of their licensees have. It’s funny when you ask either Arm Holdings, or Imagination Technologies, any questions about any technologies that may be utilized in their licensees’ products, they have to be careful not to reveal too much about any future use of new technologies that their customers are even thinking about using.
I would guess that companies like Apple, AMD, Samsung, exc. would be the ones that would be taking the HBM technology to their custom ARM ISA based products first, especially AMD for some K12 based server variant, and AMD will be the first major user of HBM technology in their GPU product line. HBM with its extra wide data BUS, and lower bus clock speeds should do fine in mobile products that are designed to save power and space, so stacking HBM on an interposer directly adjacent to the SOC definitely would save space, epically for mobile where every bit of space saved allows more space for the battery. The phone OEMs will probably be the first mobile market users that utilize SOCs connected to HBM, if just to free up the area occupied by the standard RAM memory chip. Moving all the memory traces to the interposer would also allow the mobile device’s motherboard to be made smaller and simplified.
I hope PCPer makes it to the Hot Chips symposium in August, both AMD and Intel have some interesting presentations, and there is also a 64 core ARM processor presentation from Phytium for HPC workloads.
Ryan ,u really need to do
Ryan ,u really need to do more videos like this .
this video and the “Learn about Power Supplies with Lee” are amazing and super useful .
thanks
Thanks for the compliments!
Thanks for the compliments!
Thanks man, appreciated. We
Thanks man, appreciated. We are already pursuing some other people to interview!
That was great and an
That was great and an enjoyable watch. I love listening to smart people. Thanks!
Just have to agree with the
Just have to agree with the above comments, very good and interesting video.
just finished my first time
just finished my first time through, SOOOO OVER MY HEAD, but sorta understandable, gona watch it several more times, learn what I can, PLEASE DO MORE LIKE THIS!
Have to agree, didn’t
Have to agree, didn’t understand everything but I really enjoyed the talk and learned some new things! Thanks.
Just here to say this is a
Just here to say this is a great video! – Great questions Ryan and Josh, and great detail from Jem Davies. Thank you all!
I was a little curious about the ability of these mobile GPUs to scale up to desktop/”big console” class performance; I think I heard a little discussion in that direction but may have missed it. I’m sure the entry to market isn’t that easy though given drivers, compatibility testing, etc..
I’m also curious how the Mali GPU compares to Maxwell for efficiency at sub-1W power conditions.
i know it wont happen, but
i know it wont happen, but single slot PCIe Graphics cards featuring the same cores as the Mali-T880 (is that Midgard? i cant remember)would be interesting, MiniDP ports and possibly a MiniHDMI, no DVI and no VGA
The big problem here is what
The big problem here is what market do they target? Intel includes a GPU on every one of their consumer GPUs, and AMD does the same. Their performance per watt is great, but they would need to compete in the market of AMD and Nvidia monster GPUs.
AMD pretty much has the low end taken care of, and that is where most of the buying happens. But that is also where AMD and Nvidia really compete.
They could focus on cheap compute, and that may be a really good market for them. In fact, I would be much more excited about a PCI-E
card with say four of those 8 core arm chips with those GPUs. I think that could be more useful in the compute world, and the PC form factor can handle the power and heat requirements.