Apple A8x

[III-V]

Honestly, I'd like synthesis to be used in Intel's designs as well. I'd imagine that we'll reach a point where computers will be able to optimize for performance better than humans can. In the meantime, we stand to gain a lot on cost and power.

Certainly Atom should be a synthesized design. Quark would be a definite candidate, if it isn't synthesized already. Core would probably not benefit too much on an inter-core level, seeing how packed it is, but I'd love to be proven wrong. At a die-level though, there is certainly work to be done. Broadwell seems to have helped with that.

Finally, Intel getting into synthesis would be great for everyone else.

 

[Nothingness]

As far as I know parts of Intel larger chips already use synthesis and Quark is fully synthesizable. For the latter given that it's a 486 micro-arch they had to redo it anyway so they took the wiser path

 

[TuxDave]

Honestly, I'd like synthesis to be used in Intel's designs as well. I'd imagine that we'll reach a point where computers will be able to optimize for performance better than humans can. In the meantime, we stand to gain a lot on cost and power.

Certainly Atom should be a synthesized design. Quark would be a definite candidate, if it isn't synthesized already. Core would probably not benefit too much on an inter-core level, seeing how packed it is, but I'd love to be proven wrong. At a die-level though, there is certainly work to be done. Broadwell seems to have helped with that.

Finally, Intel getting into synthesis would be great for everyone else.

Intel has been "into it" for more than a decade. There are some things that synthesis is clearly better than humans at dealing with just as there are some things synthesis still sucks horribly at (even today). So you pick and choose depending on which one does a better job and the general trend is to pull % of the die that's synthesized up while pushing the % of the die that's custom built down.

 

[Arachnotronic]

As far as I know parts of Intel larger chips already use synthesis and Quark is fully synthesizable. For the latter given that it's a 486 micro-arch they had to redo it anyway so they took the wiser path

You know what's hilarious? If you look at the i486 datasheet and the Quark datasheet, the core block diagrams for the two are virtually identical. In fact, they may actually be identical...will have to look again.

 

[Thala]

SoFIA uses fully synthesized Silvermont cores as apparently they are on 28nm TSMC.

 

[witeken]

Honestly, I'd like synthesis to be used in Intel's designs as well. I'd imagine that we'll reach a point where computers will be able to optimize for performance better than humans can. In the meantime, we stand to gain a lot on cost and power.

Certainly Atom should be a synthesized design. Quark would be a definite candidate, if it isn't synthesized already. Core would probably not benefit too much on an inter-core level, seeing how packed it is, but I'd love to be proven wrong. At a die-level though, there is certainly work to be done. Broadwell seems to have helped with that.

Finally, Intel getting into synthesis would be great for everyone else.

What does that synthesis thing mean, exactly?

 

[Thala]

What does that synthesis thing mean, exactly?

Are you asking what synthesis is? It is the (automatic) translation of a high level design description in RTL to gate level.
Typically Intel has the man-power (and also the profit margin) to go the full-custom route. The advantage is, that you are getting a better performance or performance/watt with full-custom designs. However you are restricted to a specific cell library and have to do all the work again, when switching cell libraries.

You can think of it as High Level Language vs. Assembly.

 

[witeken]

Are you asking what synthesis is? It is the (automatic) translation of a high level design description in RTL to gate level.
Typically Intel has the man-power (and also the profit margin) to go the full-custom route. The advantage is, that you are getting a better performance or performance/watt with full-custom designs. However you are restricted to a specific cell library and have to do all the work again, when switching cell libraries.

You can think of it as High Level Language vs. Assembly.

Okay. It remember a question and answer at Reddit about this topic.

How many transistors/gates are hand-laid vs computer synthesized on a modern processor? Old Pentium Pro era videos shown at colleges indicated a heavy reliance on manual design for large sections, although I assume this isn't the case anymore.

I can't say. But the most important, performance and power sensitive parts are still hand-drawn. Otherwise you can't get past around 1.8GHz on Intel 22nm without losing too much perf from overhead.

http://www.reddit.com/r/IAmA/comments/15iaet/iama_cpu_architect_and_designer_at_intel_ama/c7mqk3l

 

[III-V]

Intel has been "into it" for more than a decade. There are some things that synthesis is clearly better than humans at dealing with just as there are some things synthesis still sucks horribly at (even today). So you pick and choose depending on which one does a better job and the general trend is to pull % of the die that's synthesized up while pushing the % of the die that's custom built down.

Yeah, I guess I should clarify. I mean on a higher level. You've got all kinds of macros and units that perform certain functions. Inside those, they are obviously very synthesized. But on a core floorplan level, those units generally aren't "meshed" together to save area. Well, I'm sure they are in many circumstances, but the level of "meshing" on a Core uarch floorplan is low, relative to something like the A8, where individual units are indistiguishable from each other on a die shot.

 

[TuxDave]

Yeah, I guess I should clarify. I mean on a higher level. You've got all kinds of macros and units that perform certain functions. Inside those, they are obviously very synthesized. But on a core floorplan level, those units generally aren't "meshed" together to save area. Well, I'm sure they are in many circumstances, but the level of "meshing" on a Core uarch floorplan is low, relative to something like the A8, where individual units are indistiguishable from each other on a die shot.

You seem to believe the primary benefit of synthesis is to save area. I can probably list out several areas on a core where if you flipped it to synthesis, your area would blow up because it's so heavily wire and transistor utilized (we tried).

In my opinion, the benefit of synthesis isn't to save area. What you DO get is a window to iterate on a broader scope of architecture for a longer duration of the development which is pretty damn useful (at the dismay of physical designers). Honestly, that's pretty good and you probably win more than you lose from a worse frequency/power/area design.

Anyways... enjoy this:
http://techreport.com/r.x/p4-prescott/die-shot.jpg

 

[III-V]

I do understand its primary purpose (computers are much cheaper and quicker labor than humans!), but there certainly are areal gains to be had. I don't object that synthesis can provide suboptimal layouts in some circumstances, but that is a problem that should be solved with time as better algorithms are developed. Regardless, I can't help but think that a design like the A8 CPU or Bobcat has significantly benefited in area by such high levels of synthesis.

 

[Nothingness]

http://www.reddit.com/r/IAmA/comments/15iaet/iama_cpu_architect_and_designer_at_intel_ama/c7mqk3l

I can't say. But the most important, performance and power sensitive parts are still hand-drawn. Otherwise you can't get past around 1.8GHz on Intel 22nm without losing too much perf from overhead

That quote doesn't mean a lot alone: it sounds as if Intel 22nm requires custom parts to reach 1.8GHz. It's also a property of the design itself; roughly speaking the more job you need to be done in a single cycle, the more time it takes to flow through a cycle, which reduces frequency. These parts that take a lot of time ("critical paths") can be custom redone to reduce time and thus increase frequency.

Not sure I was clear...

Anyways... enjoy this:
http://techreport.com/r.x/p4-prescott/die-shot.jpg

Frightening :biggrin:

By the way, synthesis doesn't necessarily imply automatic layout. I have seen flows using RTL synthesis followed by custom flows for place and route.

 

[Idontcare]

I do understand its primary purpose (computers are much cheaper and quicker labor than humans!), but there certainly are areal gains to be had. I don't object that synthesis can provide suboptimal layouts in some circumstances, but that is a problem that should be solved with time as better algorithms are developed. Regardless, I can't help but think that a design like the A8 CPU or Bobcat has significantly benefited in area by such high levels of synthesis.

Synthesis versus Manual is all about ROI.

Be it time to market release, cost to develop, cost to manufacture, design for functional yields, design for parametric yields, etc.

Everything comes down to maximizing the end goals (emphasis on plurality) with respect to the inputs (time, money, capability).

 

[Thala]

By the way, synthesis doesn't necessarily imply automatic layout. I have seen flows using RTL synthesis followed by custom flows for place and route.

Precisely, synthesis and layout are two different things. Top level floorplan is usually done manually. The whole layout process might be iterative. First synthesis and layout on component level, then feedback area information to top level floor planning, then go back to component level for timing closure but also RTL changes are possible within the boundaries of the top level floor plan.
This is assuming a reasonably complex design.

 

[Nothingness]

then go back to component level for timing closure but also RTL changes are possible within the boundaries of the top level floor plan.

That's why backend/implementation engineers sometimes really hate RTL designers :biggrin:

 

[Eug]

So not only does FaceTime on the A8 support H.265 HEVC, A8 apparently also supports playing it at 4K using software like WALTR.

How many other SoCs include H.265 encode/decode in hardware?

This makes me think the next Apple TV will have an A8 (not A8X), or perhaps a single core variant of A8.

 

[NTMBK]

How many other SoCs include H.265 encode/decode in hardware?

Another advantage to Apple's vertical integration. Since the h.265 specification wasn't finalized, other hardware vendors could not provide acceleration- but Apple doesn't need to care about the standard. They just need their own applications to use the same "h.265" codecs as their own hardware, even if they don't 100% match the finalized standard.

 

[Eug]

I'm thinking the draft was close enough to final for companies to integrate hardware H.265. After some Googling I came across this Mediatek smartphone SoC with hardware H.265 for example.

http://www.mediatek.com/en/news-eve...m-cortex-a17-and-ultra-hd-h265-codec-support/

That's from February 2014.

Just as important for me though is when Intel or 3rd party GPU makers will include this. I'm due for both a desktop and laptop upgrade in the next couple of years. I did see much mention of OpenCL decode acceleration though even if the GPUs and CPUs aren't specficially built for H.265.

 

[NTMBK]

Just as important for me though is when Intel or 3rd party GPU makers will include this. I'm due for both a desktop and laptop upgrade in the next couple of years. I did see much mention of OpenCL decode acceleration though even if the GPUs and CPUs aren't specficially built for H.265.

Not seen any confirmation, but I'd be surprised if Skylake doesn't get h.265 acceleration.

 

[Eug]

Ahhh, perfect. A 14 nm Skylake uber-thin Retina MacBook Pro in late 2015 / early 2016 would last me a long, long time. I'm still using a 2009 Penryn non-Retina MacBook Pro.

 

[witeken]

How many other SoCs include H.265 encode/decode in hardware?

Core M: GPU-accelerated. Tegra M1: partially.

 

[Arachnotronic]

Not seen any confirmation, but I'd be surprised if Skylake doesn't get h.265 acceleration.

Skylake does indeed support H.265 encode/decode.

http://www.cpu-world.com/news_2014/2014060901_Some_features_of_Skylake_graphics_architecture.html

 

[ViRGE]

Core M: GPU-accelerated. Tegra M1: partially.

Encode or decode? I haven't heard anything about Core M's encoder, but it's definitely a hybrid decoder.

 

[krumme]

Ahhh, perfect. A 14 nm Skylake uber-thin Retina MacBook Pro in late 2015 / early 2016 would last me a long, long time. I'm still using a 2009 Penryn non-Retina MacBook Pro.

Man you are lucky for such a huge upgrade - and patience, your next one will probably last 15 years. Lol. Penryn was a good speedy mobile processor. For normal office use a mobile 2.5GHz penryn with a good bunch of cache feels more or less as fast imo as a modern ultrabook so dont expect to much if you have a good fast sdd already. The core duo was a super nice mobile processor too. So lean for its time and vastly overlooked. My favorite processor.

 

[witeken]

Encode or decode? I haven't heard anything about Core M's encoder, but it's definitely a hybrid decoder.

https://intel.activeevents.com/sf14...6F6352FA512AF9CC26B9993/SF14_SPCS002_104f.pdf

Page 66: GPU-accelerated HEVC decoding and encoding