How the SoC Is Displacing the CPU | Hacker News Reader

How the SoC Is Displacing the CPU

medium.com

38 pointsingve51 comments

This seems like a silly title. I interpret it as 'How the CPU, when packaged with a bunch of other stuff, is displacing the CPU.'

Or am I missing something?

You're not missing anything -- the article is dumb. Of course we're eventually going to integrate everything, for the same reason that no one uses separate interrupt controllers or separate Northbridges. We have too many transistors we don't know what to do with.

Edit: I think the interesting question is what won't be integrated? RAM and Flash because they use different processes. GPU? Intel and ARM integrate it, but there is a large company selling GPUs.

Makes me wonder if we will see a return of backplanes.

Unless things have really changed, there is a significant performance hit for going off chip. Rather than backplanes, I'd predict that they figure out how to put ram right on the chip. I'm pretty sure embedded CPUs already do that.

PoP "Package on Package" which puts the memory directly on top of the CPU has been around for a while.

https://en.wikipedia.org/wiki/Package_on_package

PoP is nice, but what's really hot is SRAM right on the die. I've been playing with a Renesas part (RZ/A1H) that has 10MiB snuggled up to the A9 core. Doesn't solve everything, but when your app heap and stack are in there it's a great jump in performance for the price. It can also mmap external QSPI flash, you can create a system with no external SDRAM at all if you wanted.

Scratchpads have all kinds of benefits plus are easier to implement than caches. Unfortunately, the market rejects them mostly in Intel's sector. So, we have things like IBM using a ton of onboard cache in high-end chips that's not quite SRAM speed. Worse is better. (shrugs)

Also I would like to know how much die space is wasted on IBM PC backwards compatibility crap in Intel systems (think A20, cascading interrupt controllers, 8086 support, legacy boot, etc)

Not much, I guess. The original 8086 had only around 30000 transistors. The die of a modern quad-core Intel chip accommodates more than a billion transistors, about half of which are dedicated to the GPU and maybe around 10% are taken by a single core. That would roughly translate to 100 million transistors per core (including their L1 and L2 caches), so the footprint of the legacy cruft should be negligible.

An intel 8088 processor had a 22mm^2 die size at a 3000nm process node size. A modern x86 processor has a ~200mm^2 die size at a 14nm node size, a die size that is 10x larger with features 200x smaller.

Even if a modern x86 chip had to put an entire 8088 on die for legacy compatability (which it definitely would not need to) we are talking less than 0.1% of the die space being taken up by legacy features.

The idea that x86 is somehow doomed to be inefficent because of legacy baggage is quite simply massively overblown.

While you're right that it's not a huge deal, architectural issues are not solved by putting "an entire 8088 on die for legacy compatability" unless you'd want to run the old software at very low speed. Intel solved it with lots of R&D. Their solution involves a layer that translates x86 instructions to native ops[1], plus many little hacks and improvements.

[1] http://www.realworldtech.com/nehalem/5/

"Apple A9X SoC offers 64 bit desktop-class computing enabling a handheld tablet to go toe-to-toe with a state-of-the-art laptop CPU from Intel"

Yes and no. Most of the innovation in top-end x86 CPUs goes towards multicore and multi-socket scalability. My desktop is almost 10x faster than an A9X -- and nobody manufactures an ARM SoC that has 100 W thermal envelope but delivers the same speed as a Xeon v3.

Intel manufactures different state-of-the-art laptop CPUs. Apple SoCs can compete with the ultra-low voltage Intels, but not with high-end i7 HQ/MQ series.

nobody manufactures an ARM SoC that has 100 W thermal envelope but delivers the same speed as a Xeon v3

Look at Cavium ThunderX: 48 ARMv8 cores, very high performance, 95W.

there arn't a lot of workloads a laptop is going to care about that utilize 4 cores, let alone 48. Each indivual core is going to be 1/10th the preformance or less of an intel one.

It would feel dreadfully slow by comparison.

You won't be running a Cavium part in your laptop, ever. They specialize in parts for enterprise servers and telecoms.

However, my phone has 6 x 64 bit ARMv8 cores.

I think you're talking about the Snapdragon 808 which your phone has. That's a big.LITTLE CPU which means that it's meant to use the 4 little cores almost all of the time and use the big cores only when high performance is really required. What I mean to say is that the 4+2/4+4/4+4+4 configurations aren't meant to parallelize across all cores but rather to minimize power consumption.

The key question is which workloads will actually benefit from many low single-thread performance cores. While there are some where such a node coupled with a GPU compute cluster will probably be the useful but it seems like a vast majority of server deployments just want good single-thread perf with 2 or more hardware threads.

Don't necessarily assume that all ARM cores will have low single-thread performance. Especially with 64 bit server-class ARM SoCs, the performance can be rather good.

> Intel manufactures different state-of-the-art laptop CPUs. Apple SoCs can compete with the ultra-low voltage Intels, but not with high-end i7 HQ/MQ series.

I think his point was that a $15 Apple chip crushes a $150 Intel Core M. Core M chips have extremely little bang/buck value, as do the new "Celerons" and "Pentiums" that are dressed up $110+ Atom chips.

That $15 vs $150 comparison is stupid - where I can buy a A9 for $15?

The A9 is a great jump forward by Apple - and I'm impressed by my iPhone 6s - but I'm yet to see the "crushing" benchmark, except for those that are skewed by SHA results.

Not to mention people buying the x86 chip were buying performance and x86 compatibility. Proper comparisons for cheap or energy-efficient x86 would be best made against VIA's CPU's. Maybe also the Loongson with x86 emulation since they can give you speed in MIPS mode and x86 compatibility where needed. Don't know their price, though.

It's funny that the Celerons you call expensive make up for almost all the processors in cheap laptops such as Chromebooks which sell around the $200 price point[1]. Heck I bought one of those Kangaroo portable desktop things[2] last week for $100 which has a dual core Atom.

Now maybe Core M was an especially gimped CPU but it might be a part of Apple's strategy to push users from entry level Macbooks to the iPad Pro. But all of these other devices with Celerons / Atoms are really capable and people use them for actual work instead of using an iPhone/iPad despite what the benchmarks say.

[1] http://www.amazon.com/gp/product/B00MMLV7VQ [2] http://www.engadget.com/2015/10/26/infocus-kangaroo-pc/

For anyone else wondering what "SoC" means, here you go: https://en.wikipedia.org/wiki/System_on_a_chip

One should also note that the x86 platform has taken on some SoC like structures over the years, with AMD as a driving force.

For example they lumped the memory controller onto the CPU die with the Athlon 64, and then a GPU on their A series.

Yeah, and the integrated GPUs, at least the entry level ones, do suck. My Acer netbook is barely able to play 360p videos, 720p if you're lucky and 1080p - no hope. My cheap smartphone can play 1080p without a single dropped frame.

I suspect that could be a driver/software issue.

On neither of those platforms is the GPU doing the video decode. On your phone there's a dedicated DSP. On your netbook the CPU is pulling up the slack. There's a lot of serial work in decoding videos (like huffman decoding) that doesn't translate to standard GPUs without a little extra silicon.

An entry level AMD apu will give you a dedicated video decoding unit.

https://en.wikipedia.org/wiki/Bobcat_%28microarchitecture%29

I've never had an APU but on paper they sound really nice. Intel probably has similar but you may have to be more careful picking your CPU because of market feature segregation.

AMD tends to dominate the GPU half of the APU battle. Intel has made great strides, and I'm interested in their newest pieces, but from what I've seen so far, AMD still wins this niche.

One of them should just rip off and modernize the SGI Onyx or Octane approaches. They did amazing stuff with 150-200Mhz CPU's by great, HW architecture. :)

There's something seriously wrong with your install to not be able to play 360p videos. Really

I would check video drivers, DMA enabled on drives, excess ram or cpu usage by another process(es)

My 7 year old notebook (the one I retired already) had absolutely no problem with videos of this size.

The problem with the 360p videos is that Chrome eats up RAM like I drink beer, and add the neccessities of the Windows world like anti-malware software, adblockers, crappy 3G stick software... you get it :/

I'm guessing YouTube is using VP9 on your netbook which is probably decoded on the CPU because most GPUs don't do VP9 decoding yet. On your smartphone it's probably h264 that's being decoded on the GPU. You can try the h264ify extension[1] to force h264 on your netbook.

[1] https://github.com/erkserkserks/h264ify

On the other hand, my higher end, but certainly not highest Intel Ivy Bridge Xeon E3-1225 V2 (3.2/3.6GHz turbo) plays 1080p without a glitch using Debian Wheezy or Ubuntu Trusty. I'm sure I wouldn't want to play an intense 3D game with it, but for 2D purposes it's great and relatively cheap with all the integration in the chip, the Supermicro X9SAE-O it's on has a real paucity of additional chips, and one of those is for a not strictly needed 2nd gigabit Ethernet port (for a workstation board like it, most useful for running older systems software instead of depending on the southbridge's integrated controller).

Also, as the author missed, Haswell now integrates power regulation on the CPU.

Which evidently was a mistake, because it's been dropped with Skylake.

I'm really, really longing for one of these "modular" phones where I can exchange and upgrade components at will...

Fairphone 2 is shipping now, iirc.

The site says "Designed for use and service in Europe only.", but regardless of what it was designed for does it work with any North American carriers?

http://shop.fairphone.com/fairphone2.html#technical-specific...

That lists the various bands for the mobile radios.

I saw that, but it's strange that they'd make it w/ a capable quad-band radio and -still- declare it's only designed for Europe. I'm not a cellular expert -- are there other considerations outside the frequencies that it supports ? If not, it seems like a good global phone, no ?

I can imagine concerns with stuff like allowed WiFi bands...

GSM quad band, UMTS and LTE seem Europe only. So getting anything more than EDGE in USA seem out of the question.

In this day and age i suspect finding GSM radios that are not quad band are rare to say the least.

Never mind that this phone is unlikely to ever be sold with a carrier subsidy, as it will be damn hard to lock in something that can have its SoC assembly replaced by removing a few screws (and ship with the required screwdriver).

This is a bit silly. A9X and other mobile SoC are in absolutely no way comparable in performance, on any metric, with desktop CPUs. You do not get 150W performance from a 10W chip, particularly not when you are lagging in process node.

I don't think there is much lagging in the node size with the A9 family in comparison to Intel's mobile chips. A9 is a 14-16 nm FinFET design. Not much different than Intel's 14 nm FinFET for core-M. The A9X is clocked at 2.26 GHz compared with a Skylake core-m7 clocked at 1.2-3.1 GHz.

Despite what you said, you in no way addressed or contradicted his point. Clockrate and nm don't tell you performance. Intel's internal architecture has tons of tricks to get more bang for buck out of their cycles. Plus lots of accelerators with extra instructions. Plus modifiable with microcode update. So, a proper comparison would be a number of workloads on each that people use for desktops.

Note: I still haven't seen a mobile that outperforms my old Core Duo 2 laptop. Always slower in both delay and throughput.

Intel biggest inroads in SoC market happened when it was giving away Atoms below cost ($5, Mediatek/Rockchip price point) while losing $3-4 Billion. Year after year since 2013.

Their plan for 2016 is selling LTE modems below cost to Apple.

I was wondering why every darn Android tablet seemed to be using Atom, if they didn't use their own SoC or one of the big names.

How is it even legal ?

If Intel wants to lose a few billion dollars while gaining almost no market share, why shouldn't it be legal? It was a massive failure all around, with basically only low-end tablets and some Asus phones using Intel's fire-sale chips.

Had it succeeded I'm sure we could have seen some lawsuits but it didn't so we don't.

Its totally cool. Learning from Qualcomm experience (milking Chinese companies for IP resulted in $1B penalty hehe) Intel "partnered" with Rockchip and Spreadtrum (read: was forced by Chinese regulators to give away its Atom IP to Rockchip) and "invested" in a bunch of companies, not to mention will spend almost $6B building a fab there.

Reply on news.ycombinator.com