Apple To Develop Own GPU, Drop Imagination's GPUs From SoCs
by Ryan Smith on April 3, 2017 6:30 AM ESTThe Future: Competition, Secrecy, & the Unexpected
Finally, while Apple developing their own GPU is not unexpected given their interests and resources, the ramifications of it may very well be. There hasn’t been a new, major GPU vendor in almost a decade – technically Qualcomm’s team would count as the youngest, though it’s a spin-off of what’s now AMD’s Radeon Technologies Group – and in fact like the overall SoC market itself, the market for GPU vendors has been contracting as costs go up and SoC designers settle around fewer, more powerful GPU vendors. So for someone as flush with cash as Apple to join the GPU race is a very big deal; just by virtue of starting development of their own GPU, they are now the richest GPU designer.
Of course, once they start shipping their custom GPU, this will also open them up to patent challenges from those other players. While it has largely been on the backburner of public attention, this decade has seen a few GPU vendors take SoC vendors to court. This includes NVIDIA with Samsung and Qualcomm (a case that they lost), and still ongoing is AMD’s case against LG/MediaTek/Sigma/Vizio.
GPU development is a lot more competitive due to the fact that developers and compiled programs aren’t tied to a specific architecture – the abstraction of the APIs insulates against individual architectures – however it also means that there a lot of companies developing novel technologies, and all of those companies are moving in the same general direction with their designs. This potentially makes it very difficult to develop an efficient GPU, as the best means of achieving that efficiency have often already been patented.
What exists then is an uneasy balance between GPU vendors, and a whole lot of secrets. AMD and NVIDIA keep each other in check with their significant patent holdings, Intel licenses NVIDIA patents, etc. And on the flip side of the coin, some vendors like Qualcomm simply don’t talk about their GPUs, and while this has never been stated by the company, the running assumption has long been that they don’t want to expose themselves to patent suits. So as the new kid on the block, Apple is walking straight into a potential legal quagmire.
Unfortunately, I suspect this means that we’ll be lucky to get any kind of technical details out of Apple on how their GPUs work. They can’t fully hide how their CPUs work due to how program compilation works (which is why we know as much as we do), but the abstraction provided by graphics APIs makes it very easy to hide the inner-workings of a GPU and make it a black box. Even when we know how something works, features and implementation details can be hidden right under our noses.
Ultimately today’s press release is a bit bitter-sweet for all involved in the industry. On the one hand it absolutely puts Imagination, a long-time GPU developer, on the back foot. Which is not to spell doom and gloom, but the company will have to work very hard to make up for losing Apple. On the other hand, with a new competitor in the GPU space – albeit one we’ve been expecting – it’s a sign that things are about to get very interesting. If nothing else, Apple enjoys throwing curveballs, so expect the unexpected.
Facebook
Twitter
RSS
144 Comments
View All Comments
lilmoe - Monday, April 3, 2017 - link
In the short term? No. Evidently? Highly possible. Nothing's stopping them.psychobriggsy - Monday, April 3, 2017 - link
RISC-V would be a potential free-to-license ISA that has had a lot of thought put into it.But maybe for now ARM is worth the license costs for Apple.
vFunct - Monday, April 3, 2017 - link
Thing is, Arm is already Apple originated, being funded by Apple for their Newton.But, given the rumors of Apple buying Toshiba's NAND flash fabs, it seems more likely that Apple is going all in on in-house manufacturing and development of everything, including ISA and fabs.
vladx - Monday, April 3, 2017 - link
Apple owning their own fabs? Seriously doubt it, the investment is not worth it for just in-house manufacturing.Lolimaster - Monday, April 3, 2017 - link
And if your sales kind of plummet, the fab costs will make you sink.FunBunny2 - Monday, April 3, 2017 - link
-- That's a moderately large undertaking.that's kind of an understatement. the logic of the ALU, for instance, has been known for decades. ain't no one suggested an alternative. back in the good old days of IBM and the Seven Dwarves, there were different architectures (if one counts the RCA un-licenced 360 clone as "different") which amounted to stack vs. register vs. direct memory. not to mention all of the various mini designs from the Departed. logic is a universal thing, like maths: eventually, there's only one best way to do X. thus, the evil of patents on ideas.
Alexvrb - Monday, April 3, 2017 - link
The underlying design and the ISA don't have to be tightly coupled. Look at modern x86, they don't look much like oldschool CISC designs. If they're using a completely in-house design, there's no reason they couldn't start transitioning to MIPS64 or whatever at some point.Anyway I'm sad to see Apple transitioning away from PowerVR designs. That was the main reason their GPUs were always good. Now there might not be a high-volume product with a Furian GPU. :(
FunBunny2 - Tuesday, April 4, 2017 - link
-- Look at modern x86, they don't look much like oldschool CISC designs.don't conflate the RISC-on-the-hardware implementation with the ISA. except for 64 bit and some very CISC extended instructions, current Intel cpu isn't RISC or anything else but CISC to the C-level coder.
willis936 - Wednesday, April 5, 2017 - link
"Let's talk about the hardware. Now ignore the hardware."name99 - Monday, April 3, 2017 - link
I think it's perhaps too soon to analyze THAT possibility (apple-specific ISA). Before that, we need to see how the GPU plays out. Specifically:The various childish arguments being put forth about this are obviously a waste of time. This is not about Apple saving 30c per chip, and it's not about some ridiculous Apple plot to do something nefarious. What this IS about, is the same thing as the A4 and A5, then the custom cores --- not exactly *control* so much as Apple having a certain vision and desire for where they want to go, and a willingness to pay for that, whereas their partners are unwilling to be that ambitious.
So what would ambition in the space of GPUs look like? A number of (not necessarily incompatible) possibilities spring to mind. One possibility is much tighter integration between the CPU and the GPU. Obviously computation can be shipped from the CPU to the GPU today, but it's slower than it should be because of getting the OS involved, having to copy data a long distance (even if HSA provides a common memory map and coherency). A model of the GPU as something like a sea of small, latency tolerant, AArch64 cores (ie the Larrabee model) is an interesting option. Obviously Intel could not make that work, but does that mean that the model is bad, that Intel is incompetent, that OpenGL (but not Metal) was a horrible target, that back then transistors weren't yet small enough?
With such a model Apple starts to play in a very different space, essentially offering not a CPU and a GPU but latency cores (the "CPU" cores, high power and low power) and throughput cores (the sea of small cores). This sort of model allows for moving code from one type of core to another as rapidly as code moves from one CPU to another on a multi-core SoC. It also allows for the latency tolerant core to perhaps be more general purpose than current GPUs, and so able to act as more generic "accelerators" (neuro, crypto, compression --- though perhaps dedicated HW remains a better choice for those?)
Point is, by seeing how Apple structure their GPU, we get a feeling for how large scale their ambitions are. Because if their goal is just to create a really good "standard" OoO CPU, plus standard GPU, then AArch64 is really about as good as it gets. I see absolutely nothing in RISC-V (or any other competitor) that justifies a switch.
But suppose they are willing to go way beyond a "standard" ISA? Possibilities could be VLIW done right (different model for the split between compiler and HW as to who tracks which dependencies) or use of relative rather than absolute register IDs (ie something like the Mill's "belt" concept). In THAT case a new instruction set would obviously be necessary.
I guess we don't need to start thinking about this until Apple makes bitcode submission mandatory for all App store submissions --- and we're not even yet at banning 32-bit code completely, so that'll be a few years. Before then, just how radical Apple are in their GPU design (ie apparently standard GPU vs sea of latency tolerant AArch-64-lite cores) will tell us something about how radical their longterm plans are.
And remember always, of course, this is NOT just about phones. Don't you think Apple desktop is as pissed off with the slow pace and lack of innovation of Intel? Don't you think their data-center guys are well aware of all that experimentation inside Google and MS with FPGAs and alternative cores and are designing their own optimized SoCs? At least one reason to bypass IMG is if IMG's architecture maxes out at a kick-ass iPad, whereas Apple wants an on-SoC GPU that, for some of their chips at least, is appropriate to a new ARM-based iMac 5K and Mac Pro.