More importantly, it would almost certainly be a broken implementation, without the ability to be updated. 5G is extremely complex, even when processing in software.
If you want to do real time 5G calls (as you do when you prototype, test and develop) using software defined radio you buy one of those USRP boxes. They have CPU's DSPs that run software and then programmable FPGA's talking to ADCs and DACs.
Here is an interesting paper on the architecture of an LTE modem. Seems like a reasonable approach, although I don't know how prevalent this approach is. RFIC, some dedicated hardware for the PHY DSP stuff and an ARM chip for the high level logic, all connected via AXI bus.
It's not a feature set of a CPU, this is an SOC, or System On Chip. It's not like an Intel 4004 that just does the processing and relies on external chips for peripherals, they bake in all the typical stuff that the system it's designed for (a cell phone) is expected to use. It contains 8 ARM cores (one of which is "the CPU"), a GPU, an AI DSP, a memory controller, 3 image signal processors for connection to cameras, video encoder/decoders, Bluetooth, Wifi, NFC, GPS/GNSS, dual SIM support, probably some general-purpose inputs and outputs, plus, yes, a 5G modem.
To sum up the explanations given for why it's part of the SOC and not an external device on a bus:
> Last year’s ... external 5G modem such as the X55 helped the 5G transition as it was available to vendors earlier than the Snapdragon 865 SoC itself, allowing them to design their RF systems before having access to the newest SoC.
> The platform’s reabsorption of the modem into the SoC die should signify better power efficiency, lower platform cost as well as lower PCB complexity for smartphone vendors.
How does this compare to the latest A14 (i think that's the latest) when it comes to "baked in" radio set? Is it still external on apple device and it's just a matter of time until they bake in what they acquired from Intel?
Or are they already baking it in and just paying royalties to Qualcomm?
Apple use a separate modem chip. iPhone12 has a Qualcomm SDR865 baseband and SDX55M radio. There have been rumours about Apple making their own modem, and presumably integrate it in an A-series SoC, but nothing yet.
A 5G modem is pretty complicated tech. I used to work on Nvidia's Icera, but that's gone now. A lot of chip makers have given up, and Qualcomm is the undisputed leader. MediaTek dominates in low-mid end Androids. I think Huawei & Samsung both have something.
The A14 uses an external Qualcomm radio, much like Android phones based on the previous Qualcomm Snapdragon 865. Take a look at the yellow and green boxed SDR865 and SDX55M in this image of the iPhone 12 with A14 motherboard from the iFixit teardown:
The SDR865 and SDX55M are the external 5G parts that would disappear with the 888. Compare the PCB area used for those (they're both almost as large as the whole SOC; I'm sure much of that is driven not by silicon area or thermal requirements but by a lot of power and pins for high-speed host communication) to this teardown of the Galaxy S10, which has the Snapdragon 855 with an integrated 4G modem:
Apple bought Intel’s modem division in 2019. That purchase included related patents. The expectation is that they will move to integrate their own modems into their SOC in the next couple of years. That will help increase the efficiency of their chips.
It may also make it feasible to include 5G in laptops. Currently Apple licenses Qualcomm modems. Qualcomm charges a fee that is proportional to the full retail price of the device. That would be significant for a laptop. If Apple can use their own modems and get around that licensing cost, it makes a laptop with a cellular connection more likely.
For most people their phone is their most important tech device where a lot of critical data transit (banking, social etc) so using an unsecured phone isn't a good idea.
You can keep using your old phone, but no updates means no security updates. This means your browser and other apps are susceptible to compromise. As time goes on the exploits are automated, this becomes less the provenance of determined organizations, and more accessible to script kiddies or casual hackers.
If the drivers are open source you can at least run something like LineageOS or another Android derived open source ROM and still run an updated OS on your old hardware.
It's a gamble, I surely wouldn't use a KitKat phone for internet browsing, it is probably no different from using an ancient version of IE from 20 years ago, it is just a matter of time before something bad happens. The biggest problem is, most OEMs nowadays lock bootloaders and make it a lot tougher to use custom ROMs for longer support, non malicious pre-installed applications, and removing Google services - almost non-existent FOSS support from Qualcomm devices doesn't help either. I only buy for myself and suggest to others devices that are more likely to get community driven support, that for example bars some manufacturers with ridiculous lock-ins.
In addition to security, sometimes new versions of Android have a feature that I would like to use. And knowing that the hardware I am using is perfectly capable of supporting it but just because the greedy Android ecosystem wants to sell me new devices bothers me. And not only new features but there can also be bugs. For example I have an Android tablet which stopped getting updates after just 1 year. It has bugs which did not get solved during this period so they will never get solved. Also I like using dark mode that is introduced with Android 10, which will never come to my tablet. So I need to buy a new tablet if I want dark mode or less buggy experience.
This is exactly my issue with Qualcomm. They don't compete in device support or performance metrics. They literally offer nothing competitive. They only exist because of the way the market is segmented.
one could regulate this. Share source and toolchain or don't sell in this market. The EU has something coming with their repairability laws, but the software side isn't too underlined yet.
The early to mid 2010 android devices are piling up left and right, but the software is old. As phones they're still good enough but network operators build back tower hardware for 3G, so there's another EoL date coming at least for their data channels.
For days, I'm reading about the 888, its Cortex X1 and a 35% uplift and that it's coming (some day) but I couldn't find a single word about how it will perform against the M1. Even Anandtech dares to bring four, full pages of fluff.
This is what most are interested in and Qualcomm should asap disclose the performance and/or give an outlook on what is coming, always referencing M1's performance as the current benchmark.
Exactly my gut feeling and the press is not questioning anything just rephrasing press releases from Qualcomm, writing about so important details like why the name 888 and so on.
Qualcomm and Intel should come out asap and tell what they plan to do in the next 12-18 months, e.g. embedding RAM or whatever, to match M1' performance. But they rather work on original naming schemes. Good that I sold my Qualcomm stocks years ago.
Can you explain why 888 will be n-times slower? Assuming, with n greater than 1.
I just looked up the Snapdragon performance vs Apple A14, and it is within ~10% [1]. I have also compared it to an intel processor, and it also ahead [2].
From a quick search, it looks like Qualcomm is a competitive contender, if we're going to ARM. If true, that would be a pretty exciting new world of ARM PCs :)
Ok, I'm sorry, I haven't been reading CPU reviews for the past... 10 years. Can you explain why I would add 35% on top? What is "miles away" converted to percent difference in operations per second?
I think the 35% on top is the expected improvements of the 888 compared to the 865.
On Geekbench, the top 865 is at 887 for single thread compared to 1585 for the A12. If we add 35% to the 865, the expected 888 score is ~1200, still quite behind the A12.
I see, got it. So, if we look at single core performance, I can see it is much less performant, but does not seem like more than a magnitude. It would also make it more efficient than Intel. Still, overall, would be very interesting to see a high performance Snapdragon ARM laptop!
Most of that is difference in the process technology. Snapdragon 865 Plus uses N7P, A15 uses 5nm. It's not Qualcomm vs Apple, It's TSMC vs TSMC. Apple pays premium for first digs.
Rest is microarchitecture design choices and integration. Apple designs only Apple in mind and can afford to choose expensive solutions for performance gains. Qualcomm processors are sold for a larger number of customers and they have to compromise.
Since the 855 or even earlier, Qualcomm had always a notebook derivate SoC (which were not much faster than the mobile phone version). However, the M1 is now the reference for any ARM licensee and anyone at Qualcomm's scale should have some answer to Apple's SoC, especially then when the create some press buzz around a new SoC with a fancy name.
The magic of M1 is its x86 emulation speed with it's memory flag which I'm guessing is not something that has been on the roadmap for Qualcomm until recently. Even then you need integrate all the hardware/software partners instead of Apple's vertical integration.
Even if they have M1 level performance without the good emulation, what's the point if it's stuck with windows RT?
Windows RT isn't too bad. You wouldn't actually notice a difference. A decade ago (or less), I had the first RT-based Surface running native Word processing a 300-pages PhD paper, it was running buttersmooth and cool, all the time.
I think this is more about MS' commitment to ARM than a software problem.
Weird, my experience with that same device when it launched was noticible input lag while trying to use explorer and microsoft word at the same time. The stuttering was so bad, scrolling and typing could take 100's of ms - now I wonder if I wrote off the whole RT experiment because I got a lemon or malware.
While Apple did really stellar job with M1, Apple's R&D counts less than 50% of M1 performance gain over competition.
Rest is TSMC's 5nm and perfect timing. Apple generates largest revenue per processor by a wide margin. This means they can afford to pay premium to be first at everything and have 6-12 month lead. TSMC's 5nm volume production and Apple's 2020 launch was timed to go together since 2017 (if go back to TSMC conference call you can hear them talking about it without mentioning details).
A lot of the lead is process, no doubt, but a lot of it is simply a better design. Every step of the way, even when they're at process parity, Apple chips have been better.
Your math doesn't make sense. 25% because Apple chooses better solutions, and then you finish with only 10% better design? Is it only a better design if it's not more expensive? So you get to decide that? Does BMW not get credit for better design than a Ford Focus because it's more expensive?
I guess it doesn't really matter. Apple makes a better product. That's kind of the story here at the end of the day.
What are these numbers based on? Just your "best estimate"? It's dismissive of Apple's uarch for what reason?
Arm has architecture licenses, unlike x86: anyone could've designed an Arm CPU from the ground up. NVIDIA tried, AMD tried, Intel tried, Qualcomm tried, Samsung tried, Huawei tried, etc. Everyone had a chance (and they still do). Arm is the most level playing field available in high-perf CPU design.
And "50% process technology" is an exaggeration even embarrassing for HN. The A13 was built on 7nm and still beat perf/watt of any x86 CPU and total 1T performance rivals Tiger Lake. A14 / M1 are a natural evolution of that same uarch.
What really breaks down your argument: Samsung has had nearly every advantage as Apple, yet its Exynos line is some of the worst-perf/watt Arm uarch today: its own OS (Tizen), its own foundry, its own phones / tablets / laptops, and a massive conglomerate for funding. What happen to Samsung? What money doesn't Samsung have? Hell, Samsung is even MORE integrated than Apple, as Apple still needs to outsource its fabrication to TSMC.
There's a reason Samsung is giving up on its uarch and moving to Arm stock cores (X-1, A78, A78C, etc.), just like Qualcomm + NVIDIA.
Don't tell me "you need trillion dollar valuation to make a top-class CPU". AMD was nearly bankrupt 6 years ago and now has the fastest general compute x86 arch in the world.
This argument reeks of "well, if Apple has the fastest CPU uarch today, then I'm going to ensure everyone else has an excuse."
Apple didn't even have an Arm architectural license, much less a custom high-perf CPU, 13 years ago. 13 years from "never designed a high-perf CPU" to "dominating x86 perf / watt while at the heels of total perf" is actually notable and actually impressive.
It's really disappointing that as Apple comes out with some of the most exciting processor designs we've ever seen, some look to dismiss the effort and make excuses for the other players that aren't producing.
It can't possibly be that Apple is doing impressive, innovative work. No, it has to be some nebulous "other" thing that's really happening.
JavaScript perf is a GPU test? I guess I'm learning new things every day.
(The link I shared is all CPU leaning performance tests.)
Nanoreview? Really dude? In the "Pro" section for the 865+ it lists two more cores (factual but largely irrelevant) and in the "Pro" section of the A13 it lists "Compatible with the latest DirectX 12.1" (laughably hilariously untrue).
I love how nano review has become the Android fan bible, despite its very obvious flaws.
Also, the Antutu developers said scores between different OSes are not comparable: https://www.antutu.com/en/doc/119646.htm but some people just love to boast how Android is better based on Antutu...
>13 years from "never designed a high-perf CPU" to "dominating x86 perf / watt while at the heels of total perf" is actually notable and actually impressive.
They did buy PA-Semi, a company with a lot of low power chipmaking experience. They had a plan and invested in it and continued making it better and better, but its not like they developed this exclusively from the ground up. Its impressive but more evolutionary.
A throwback to when computers doubled in performance every couple years.
P.A. Semi was partially responsible for the A4 and maybe A5? What about A6, A7, A8, A9, A10, A11, A12, A13, A14? You casually state "making it better and better", as if that's a regular occurrence for CPU architects over the past two decades. It's not, unfortunately.
Iterating a uarch is not a given. Ask Intel & AMD.
20% to 40% IPC gain / year is not a given. Ask Intel & AMD.
Iterating 20% to 40% IPC gain / year / nearly identical TDP is absolutely not a given. Ask Intel & AMD.
Likewise, P.A. Semi (in 2008) is a very recent addition in Apple's Arm journey.
// the history //
In the 1980s, Apple's ATG (Advanced Technology Group) pushed to move to on RISC and worked with Acorn in England to co-develop ARM III CPUs for Apple's "next-gen" devices
In 1990, Arm Ltd. is founded with the two largest shareholders as Apple (43%) and Acorn (43%).
Apple was reportedly responsible for Arm's current name: it's "Advanced", not "Acorn", because Apple didn't want to be associated with a former competitor.
Apple mainstreams Arm far earlier than the iPhone: in 2001, the first iPod launched with a 90 MHz ARM-derived CPU.
// P.A. Semi //
People always forget the history of the P.A. Semi acquisition, even as it's linked there.
The short version: Intel was literally gifted a major high-perf Arm uarch and screwed it up; that same CPU architect refuses to join Intel, founds P.A. Semi some years later, and Apple acquires P.A. Semi and that CPU architect joins Apple.
That "lot of low power chipmaking experience" was literally gifted to Intel and they couldn't do jack. Apple needed to acquire it after decades of trying to get Arm to work, far more effort than Intel.
In Q4 2005, Intel made $23b in profit. Apple made $4b in profit. Intel was a CPU IP & foundry. How did they miss Arm or other RISC designs? That's the kicker: Intel didn't miss Arm. Intel has repeatedly failed & floundered with "side technologies" that they thought inferior to x86: heterogenous dies, chiplets, Arm, XScale, etc. = Intel explicitly decried and denounced these "cheap" tactics by their competitors.
Intel in the 1990s, 2000s, and 2010s genuinely exhibited "not made here means it's crap and automatically inferior to anything we can make" philosophy.
1993: Daniel Dobberpuhl develops StrongARM, a high-performance ARM uarch.
1997: Intel is gifted StrongARM through a legal settlement and the "keys to the kingdom". Dobberpuhl doesn't agree to join Intel.
2003: Unrelenting on Arm's future, Dobberpuhl starts P.A. Semi after working at Broadcom.
2007: Intel sells off StrongARM, saying they don't know how to make money with ARM as AMD had applied unprecedented pressure with K8 & Athlon64.
2008: Apple acquires P.A. Semi.
2009: Dobberpuhl retires from Apple.
It's not like AMD, Intel, etc. were ignorant x86 had severe perf/watt limitations. They knew and they know. They tried repeatedly to overcome them and then simply lost interest as they couldn't make money.
In 2005 to 2008: Intel made 2x to 6x more profit per quarter than Apple. A short example:
Q4 2005 at Intel: $24b profits
Q4 2005 at Apple: $4b profits
Not revenue. Not stock price. Profit, that can be put anywhere in the business. Intel had insanely higher profits than Apple. The iPod (which ran an ARM-derived 90MHz CPU) was already out in 2008. Apple had just acquired P.A. Semi. Intel still had vastly, vastly, vastly more free cash flow and profit.
Don't tell me Intel didn't have money for Arm or hiring CPU architects or even literally acquiring entire Arm firms.
Intel had hired these very people who later joined Apple; it absorbed Arm architects, uarches, designers, portfolios, etc.
They failed. As The Register wrote in 2006,
>Intel's failure wasn't for a lack of talent or investment - or even luck - over the years. Intel threw $1.6bn on a DSP company in 1999, and followed up with a host of smaller investments. And luck blessed Intel on several occasions. When DEC's StrongARM processor fell into Intel's hands in the fall of 1997, it was as a result of a legal settlement, and an unsought and unexpected prize. Pundits at the time thought that Intel had been handed the keys to the kingdom. But billions of dollars later, Intel could only claim two significant design wins from lower tier phone OEMs RIM and Palm. Texas Instruments, by contrast, will cash $14bn in revenue from phone chip sales this year.
>
>So how did Intel fail to capitalize? In a nutshell, it failed to live up to its name. Intel may stand for 'INTegrated ELectronics', but it failed to integrate the electronics that mattered when it mattered.
>
>A series of poor management decisions ensured that StrongARM was well positioned for a market that was on the decline, and rarely competitive in a market that boomed. Early on, Intel decided against integrating dedicated digital signal processing into the StrongARM chip, later renamed XScale. While this decision was justifiable for fixed embedded markets and for PDAs, it put the chip at a huge disadvantage for lower cost devices that needed voice capabilities. In response, Intel copied the PC strategy of adding new floating point instructions, introducing MMX for Xscale. Since phone manufacturers preferred cheaper custom chips for devices that needed a multimedia flip, this was a wasted investment.
Hardware discussion on HN has always been subpar, and Apple's M1 discussion sort of confirm that. I guess most developers dont really give a damn about hardware.
Given that this TSMC 7nm->5nm process improvement has lead to the smallest gen-on-gen performance improvement for Apple's iPhone chips - the performance gains from the A13 to the A14 are the smallest in years - I really don't buy that at face value. Surely a better process helps a substantial percentage, but I think a 7nm M1 chip would have impressed in most of the same ways.
TSMC is a very conservative company. they will not throw money into a new node without its customers backing consider the amount of cash needs for each new gen (node).
TSMC's 5nm must be good enough for Apple to back it.
I think you're right, but also a little bit backward on things.
TSMC is conservative, but, with the continued guaranteed high-dollar buys from Apple, they have tremendous freedom to invest in R&D in an almost fearless way that any company would envy.
This is also why people trying to give all credit to TSMC for the performance of A-series chips are also selling Apple short on that side of the ball there too. TSMC would not be in the position they are right now without Apple's partnership powering their capital investments.
So in short, Apple's backing isn't proof that TSMC 5nm is good, it's actually a large part of why TSMC 5nm is good, and even exists in the first place.
Processor designs, as with many things is a question of tradeoffs. Apple is vertically integrated and has the luxury of designing cores based on a different set of requirements that drive the rest of the industry. As Apple makes a profit off per device sold, and not per chip, they can strategically approach chip design differently. For example, the area of die used.
But obviously not having the highest level of performance doesn't prevent from Qualcomm from producing more chips than Apple. So what's at play here is more performance != better. For you and me yes, more performance is good, but there are a lot of factors in play and not everybody looks to performance being the main metric.
I'm sorry, I just don't think your point is that interesting though.
Qualcomm is still in business so sure they're moving product and making sales. I'm not denying them that.
Apple makes an objectively better chip based off performance. I really don't care about much else. Are you a Qualcomm stock holder? What's your definition of better and why is it better than using performance?
Apple's "little" cores are roughly on par with the performance of the big cores in a Galaxy S10 while only having the power draw of the current gen ARM "little" cores.
>The performance showcased here roughly matches a 2.2GHz Cortex-A76 which is essentially 4x faster than the performance of any other mobile SoC today which relies on Cortex-A55 cores, all while using roughly the same amount of system power and having 3x the power efficiency.
The fact of the matter is the Apple ASIC team is better than Qualcomms and even if the 888 was at the same 5mn node the M1 would be faster. I find it very frustrating how people always seem to say "well, Apple only is better because of ... and wait until ..."
Apple has been kicking Qualcomms butt for a number of years now and up until now building ASIC was not primary part of their company. Now that they have moved the Mac platform over to their own chips I believe you can expect to gap to widen.
I hope, truly that this becomes a awake up call for Intel and Qualcomm.
>Apple's R&D counts less than 50% of M1 performance gain over competition
>TSMC announced the first details about its 5 nm manufacturing technology which is going to enable TSMC to aggressively increase its transistor density versus prior generations. However, when it comes to performance and power improvements, the gains do not look very significant.
Some Geekbench results have appeared and seem to show the 888 still at a pretty big disadvantage compared to the A14. e.g. the big cores in the A14 offering 40% greater performance, so much so that the 2 big+4 little of the A14 has higher MT performance than the 4 big+4 little 888.
These results could very well be fake, however early in a product release OEMs do often intentionally release these teaser numbers. This would make it the fastest Android chipset by a good margin, and bring it closer to equalling the A12 of 2018.
Several Big Tech vendors are designing ARM SoCs as outlined in the recent HN post The Tech Monopolies Go Vertical [1]. The Snapdragon 888 seems to compete with the Apple A14. I wonder why there isn’t more comparisons between the A14 iPad Air and the M1 MacBook Air. Microsoft’s SQ2 chip for the Surface integrates a better GPU into a Qualcomm based SoC. Perhaps GPU performance is a key differentiator. I find it odd that the 2+4 big.LITTLE A14 isn’t trounced by the 4+4 big.LITTLE Qualcomm SoCs.
A dark horse, especially with respect to GPU, is Nvidia after their ARM acquisition. These are interesting times.
What I find the most interesting is Mediatek, who largely gets ignored among the major players. Their affordable SoC's powered the Lenovo Duet into being one of the most compelling affordable Chromebook Tablets of the year. They've also mentioned plans to put more powerful chips into chromebooks in 2021.
ARM powered chromebooks run Android apps far better than their x86 cousins plus still include linux support (crostini) which has a growing set of ARM applications every year.
I'm pretty excited for where the future of efficient computing is going. Between my M1 Mac Mini and overclocked Pi4 8GB I've managed to sell off and move on from x86 machines for personal use. Interesting times indeed.
M1 competes with Qualcomm's 8cx / SQ2. It's certainly reasonable to show the comparison between M1 and 888 but the products that M1 show up in (Macbooks) don't generally go head-to-head with the products 888 shows up in.
A78 is basically a mild refinement to A77. X1 is a souped-up A78.
* Move from 4 to 5-wide decoder (A14 is 8-wide)
* Double number of 128-bit SIMD from 2 to 4 (A14 is 4)
* Moving multiplication into a second ALU allows simultaneous multiplication and division (on both A78 and X1)
* Minimum L1 cache at 64kb instead of 32kb with a 64kb option (A14 is 128 d-cache and 196 i-cache)
* L1 bandwidth on A78 and X1 doubled over A77 (probable 2-4x as wide on A14)
* Double L2 cache from 512kb to 1mb (A14 has 8mb L2, but no L3)
* Double maximum amount of L3 cache from 4mb to 8mb (for a potential 9mb cache vs 8mb for A14)
* Still ARMv8.2 (A14 is ARMv8.4 which brings some extra instructions for pointer security, virtualization, JS integer conversion, complex number SIMD, SHA512/SHA-3 hardware, int dot products, and some other stuff -- probably the biggest performance difference here will be the 1-2% for JS conversion)
* Increase BTB from 64 to 96 entries (Not sure about A14, but probably more)
* Increase TLB from 1k to 2k pages (A14 is 3072 pages)
* mops bandwidth goes from 6 to 8 mops (not sure about A14, but 8+ for sure)
* uop cache goes from 1.5k entries to 3k entries (I seem to remember hearing 2k entries for the A12, but I don't know)
* mops dispatch also goes from 6 to 8 (once again , not sure about A14, but 8+ for sure)
* uops dispatch increases from 10 in A77 to 12 in A78 to 16 in X1 (I'd guess Apple's at least this wide)
* reorder buffer increasing from 160 to 224 entries (A14 ROB is somewhere around 630 entries)
ARM's claims that with all three processors at 3GHz, the A78 is 7% faster than A77 and the X1 is 20-30% faster depending on the workload. I'd guess that SIMD float-heavy performance between X1 and A14 is very close.
More interesting is the x86 comparison (let's look at zen 2).
X1 has twice the L1 cache (64 vs 32k), but considering AMD lowered this amount, I doubt it's a real advantage more than different ISA and microarchitectures. Likewise X1 with 1mb of L2 vs 512kb for zen 2 probably doesn't matter too much (AMD redid the cache in zen 3 and stuck with 512kb while Intel still uses 256k). Zen 2 has 16 L0 BTB vs 96 L0 BTB for X1, but this might be microarchitecture more than anything as well as X1 drops to 2k L2 entries vs 7k for Zen 2. I don't know if 8mb L3 is shared in X1, but I'd assume so in which case 16mb for zen 2 is a definite advantage.
The decode width is now 20% wider than zen, but x86 instructions are around 15% more dense, so I'd guess it's about the same. X1 has 3k uop cache entries vs 4k on x86, but I suspect x86 needs more to avoid hitting those complex decoders at all costs. AMD dispatches 6 int uops and 4 float uops per cycle while X1 does 16 uops per cycle.
I don't know about ARM, but AMD's perceptron predictor is probably better (it seems like Apple or ARM would make a big deal about this if they'd made the investment -- Samsung did). Given AMD has 19 stages instead of 13 in the X1, a better predictor is needed to offset penalties. X1 and Zen 2 both have identical 224-entry reorder buffer.
X1 has 2 AGU, 1 AGU/load and 2 load/store while zen 2 has 3 AGU, 2 load/store, and 1 store. Zen 2 and X1 AGU are both 256-bit wide. I can't find store queue length for X1, but bigger queues could be an advantage to AMD (I suspect so as ARM probably wouldn't want the extra power draw).
A76 seems to have similar instruction cycle counts to Zen 2 and I assume they haven't gotten worse since. X1 has four 128-bit SIMD while Zen 2 has two 256-bit SIMD. If you can use wider SIMD, AMD has an advantage, but if there's lots of smaller SIMD, then X1 should have an advantage).
Zen 2 has 4 integer ALU with one capable of multiplication, another capable of division, and another with a CRC (cyclic redundancy checking). X1 is identical (except I don't know if it has a CRC).
X1 very likely comes close to Zen2 and even A14 in integer performance while being a bit ...
87 comments
[ 5.0 ms ] story [ 133 ms ] threadTrying to do a wholly SDR free 5G modem would be a monumental feat, and first attempts at implementations would probably be gigantic.
If you want to do real time 5G calls (as you do when you prototype, test and develop) using software defined radio you buy one of those USRP boxes. They have CPU's DSPs that run software and then programmable FPGA's talking to ADCs and DACs.
https://community.arm.com/cfs-file/__key/telligent-evolution...
I believe srsLTE, used to turn SDRs into eNodeB/UEs, adds support for 5G soon (tm).
To sum up the explanations given for why it's part of the SOC and not an external device on a bus:
> Last year’s ... external 5G modem such as the X55 helped the 5G transition as it was available to vendors earlier than the Snapdragon 865 SoC itself, allowing them to design their RF systems before having access to the newest SoC.
> The platform’s reabsorption of the modem into the SoC die should signify better power efficiency, lower platform cost as well as lower PCB complexity for smartphone vendors.
Or are they already baking it in and just paying royalties to Qualcomm?
A 5G modem is pretty complicated tech. I used to work on Nvidia's Icera, but that's gone now. A lot of chip makers have given up, and Qualcomm is the undisputed leader. MediaTek dominates in low-mid end Androids. I think Huawei & Samsung both have something.
https://www.ifixit.com/Teardown/iPhone+12+and+12+Pro+Teardow...
The same are in yellow and red on the Samsung Galaxy S20 with Snapdragon 865 teardown:
https://www.ifixit.com/Teardown/Samsung+Galaxy+S20+Ultra+Tea...
The SDR865 and SDX55M are the external 5G parts that would disappear with the 888. Compare the PCB area used for those (they're both almost as large as the whole SOC; I'm sure much of that is driven not by silicon area or thermal requirements but by a lot of power and pins for high-speed host communication) to this teardown of the Galaxy S10, which has the Snapdragon 855 with an integrated 4G modem:
https://www.ifixit.com/Teardown/Samsung+Galaxy+S10+and+S10e+...
Apple has said they want to integrate the 5G modem by 2022.
It may also make it feasible to include 5G in laptops. Currently Apple licenses Qualcomm modems. Qualcomm charges a fee that is proportional to the full retail price of the device. That would be significant for a laptop. If Apple can use their own modems and get around that licensing cost, it makes a laptop with a cellular connection more likely.
By the way I think Qualcomm do share sources (through code Aurora?). But they don't upstream them.
Writing this from a zl1 running LOS17.1 using Snapdragon 821.
P.S. Big thanks to mosimach for being device and specialised kernel fork maintainer.
If the drivers are open source you can at least run something like LineageOS or another Android derived open source ROM and still run an updated OS on your old hardware.
The early to mid 2010 android devices are piling up left and right, but the software is old. As phones they're still good enough but network operators build back tower hardware for 3G, so there's another EoL date coming at least for their data channels.
This is what most are interested in and Qualcomm should asap disclose the performance and/or give an outlook on what is coming, always referencing M1's performance as the current benchmark.
I believe X1 will still be able to beat Firestorm on size/performance ratio.
Exactly my gut feeling and the press is not questioning anything just rephrasing press releases from Qualcomm, writing about so important details like why the name 888 and so on.
Qualcomm and Intel should come out asap and tell what they plan to do in the next 12-18 months, e.g. embedding RAM or whatever, to match M1' performance. But they rather work on original naming schemes. Good that I sold my Qualcomm stocks years ago.
I just looked up the Snapdragon performance vs Apple A14, and it is within ~10% [1]. I have also compared it to an intel processor, and it also ahead [2].
From a quick search, it looks like Qualcomm is a competitive contender, if we're going to ARM. If true, that would be a pretty exciting new world of ARM PCs :)
[1] https://nanoreview.net/en/soc-compare/qualcomm-snapdragon-86...
[2] https://gadgetversus.com/processor/qualcomm-sm8250-snapdrago....
This is just so off. Look at Geekbench's figures for a 865 Plus, add 35% on top and you are still not close to an A14 and miles away from the M1.
On Geekbench, the top 865 is at 887 for single thread compared to 1585 for the A12. If we add 35% to the 865, the expected 888 score is ~1200, still quite behind the A12.
So Qualcomm didn't even compare the 888 to their latest 865+ CPU: it'll be less than 25%, but how much?
Source: https://www.anandtech.com/show/16271/qualcomm-snapdragon-888...
The 25% number is compared to the 865, not the 865+
The one big core is only 13% faster than the big core on the 865+.
>Compared to the Snapdragon 865+ which clocks in at 3.09GHz, this performance advantage should reduce to only 13%, which is less impressive.
https://www.anandtech.com/show/16271/qualcomm-snapdragon-888...
Rest is microarchitecture design choices and integration. Apple designs only Apple in mind and can afford to choose expensive solutions for performance gains. Qualcomm processors are sold for a larger number of customers and they have to compromise.
Even if they have M1 level performance without the good emulation, what's the point if it's stuck with windows RT?
I think this is more about MS' commitment to ARM than a software problem.
Rest is TSMC's 5nm and perfect timing. Apple generates largest revenue per processor by a wide margin. This means they can afford to pay premium to be first at everything and have 6-12 month lead. TSMC's 5nm volume production and Apple's 2020 launch was timed to go together since 2017 (if go back to TSMC conference call you can hear them talking about it without mentioning details).
25% for the fact that Apple can choose more expensive solutions in every turn. Better performance is often more expensive.
15% because Apple can design only for Apple products. Qualcomm, Intel, AMD .. have to design more generic products.
~10% at most for something else.
I guess it doesn't really matter. Apple makes a better product. That's kind of the story here at the end of the day.
Arm has architecture licenses, unlike x86: anyone could've designed an Arm CPU from the ground up. NVIDIA tried, AMD tried, Intel tried, Qualcomm tried, Samsung tried, Huawei tried, etc. Everyone had a chance (and they still do). Arm is the most level playing field available in high-perf CPU design.
And "50% process technology" is an exaggeration even embarrassing for HN. The A13 was built on 7nm and still beat perf/watt of any x86 CPU and total 1T performance rivals Tiger Lake. A14 / M1 are a natural evolution of that same uarch.
https://images.anandtech.com/doci/16226/perf-trajectory.png
What really breaks down your argument: Samsung has had nearly every advantage as Apple, yet its Exynos line is some of the worst-perf/watt Arm uarch today: its own OS (Tizen), its own foundry, its own phones / tablets / laptops, and a massive conglomerate for funding. What happen to Samsung? What money doesn't Samsung have? Hell, Samsung is even MORE integrated than Apple, as Apple still needs to outsource its fabrication to TSMC.
There's a reason Samsung is giving up on its uarch and moving to Arm stock cores (X-1, A78, A78C, etc.), just like Qualcomm + NVIDIA.
https://news.samsung.com/global/samsung-electronics-announce...
Don't tell me "you need trillion dollar valuation to make a top-class CPU". AMD was nearly bankrupt 6 years ago and now has the fastest general compute x86 arch in the world.
This argument reeks of "well, if Apple has the fastest CPU uarch today, then I'm going to ensure everyone else has an excuse."
Apple didn't even have an Arm architectural license, much less a custom high-perf CPU, 13 years ago. 13 years from "never designed a high-perf CPU" to "dominating x86 perf / watt while at the heels of total perf" is actually notable and actually impressive.
It's really disappointing that as Apple comes out with some of the most exciting processor designs we've ever seen, some look to dismiss the effort and make excuses for the other players that aren't producing.
It can't possibly be that Apple is doing impressive, innovative work. No, it has to be some nebulous "other" thing that's really happening.
> The A13 was built on 7nm and still beat perf/watt of any x86 CPU
We can directly compare Snapdragon 865 Plus and Apple A13 Bionic because they use the same process (N7P).
865 Plus is equal or better in most benchmarks.
What in the world are you talking about? Anandtech shows the A13 handily besting a phone equipped with an 865+.
Here is Nanoreview's Snapdragon 865 Plus vs A13 Bionic CPU performance tests: https://nanoreview.net/en/soc-compare/qualcomm-snapdragon-86...
(btw. My argument is not that Apple is worse)
(The link I shared is all CPU leaning performance tests.)
Nanoreview? Really dude? In the "Pro" section for the 865+ it lists two more cores (factual but largely irrelevant) and in the "Pro" section of the A13 it lists "Compatible with the latest DirectX 12.1" (laughably hilariously untrue).
I love how nano review has become the Android fan bible, despite its very obvious flaws.
They did buy PA-Semi, a company with a lot of low power chipmaking experience. They had a plan and invested in it and continued making it better and better, but its not like they developed this exclusively from the ground up. Its impressive but more evolutionary.
A throwback to when computers doubled in performance every couple years.
https://en.wikipedia.org/wiki/P.A._Semi
Iterating a uarch is not a given. Ask Intel & AMD.
20% to 40% IPC gain / year is not a given. Ask Intel & AMD.
Iterating 20% to 40% IPC gain / year / nearly identical TDP is absolutely not a given. Ask Intel & AMD.
Likewise, P.A. Semi (in 2008) is a very recent addition in Apple's Arm journey.
// the history //
In the 1980s, Apple's ATG (Advanced Technology Group) pushed to move to on RISC and worked with Acorn in England to co-develop ARM III CPUs for Apple's "next-gen" devices
In 1990, Arm Ltd. is founded with the two largest shareholders as Apple (43%) and Acorn (43%).
Apple was reportedly responsible for Arm's current name: it's "Advanced", not "Acorn", because Apple didn't want to be associated with a former competitor.
Apple mainstreams Arm far earlier than the iPhone: in 2001, the first iPod launched with a 90 MHz ARM-derived CPU.
// P.A. Semi //
People always forget the history of the P.A. Semi acquisition, even as it's linked there.
The short version: Intel was literally gifted a major high-perf Arm uarch and screwed it up; that same CPU architect refuses to join Intel, founds P.A. Semi some years later, and Apple acquires P.A. Semi and that CPU architect joins Apple.
That "lot of low power chipmaking experience" was literally gifted to Intel and they couldn't do jack. Apple needed to acquire it after decades of trying to get Arm to work, far more effort than Intel.
Likewise, in the 2000s, Intel had significantly higher profits than Apple: https://www.macrotrends.net/stocks/stock-comparison?s=gross-...
In Q4 2005, Intel made $23b in profit. Apple made $4b in profit. Intel was a CPU IP & foundry. How did they miss Arm or other RISC designs? That's the kicker: Intel didn't miss Arm. Intel has repeatedly failed & floundered with "side technologies" that they thought inferior to x86: heterogenous dies, chiplets, Arm, XScale, etc. = Intel explicitly decried and denounced these "cheap" tactics by their competitors.
Intel in the 1990s, 2000s, and 2010s genuinely exhibited "not made here means it's crap and automatically inferior to anything we can make" philosophy.
1993: Daniel Dobberpuhl develops StrongARM, a high-performance ARM uarch.
1997: Intel is gifted StrongARM through a legal settlement and the "keys to the kingdom". Dobberpuhl doesn't agree to join Intel.
2003: Unrelenting on Arm's future, Dobberpuhl starts P.A. Semi after working at Broadcom.
2007: Intel sells off StrongARM, saying they don't know how to make money with ARM as AMD had applied unprecedented pressure with K8 & Athlon64.
2008: Apple acquires P.A. Semi.
2009: Dobberpuhl retires from Apple.
It's not like AMD, Intel, etc. were ignorant x86 had severe perf/watt limitations. They knew and they know. They tried repeatedly to overcome them and then simply lost interest as they couldn't make money.
Sources (a great read / watch, if you're bored):
https://www.theregister.com/2006/06/28/intel_mobile_failure/ https://www.youtube.com/watch?v=AADZo73yrq4&t=449s samfisher83 ↗ > 13 years from "never designed a high-perf CPU" to "dominating x86 perf / watt zipityzi ↗ That's because Intel (back then, the only real high-perf CPU) rejected ARM as an inferior, slow technology. therealbilly ↗ AMD's recent agreement with China and the resulting monies allowed AMD to really move ahead. We have to keep that in mind. ksec ↗ >is an exaggeration even embarrassing for HN
The people who designed these chips are from intel, amd, samsung etc. Apple can spend way more money than amd or intel.
And blatantly ahistorical on "money": where is your citation?
https://www.macrotrends.net/stocks/stock-comparison?s=gross-...
In 2005 to 2008: Intel made 2x to 6x more profit per quarter than Apple. A short example:
Q4 2005 at Intel: $24b profits Q4 2005 at Apple: $4b profits
Not revenue. Not stock price. Profit, that can be put anywhere in the business. Intel had insanely higher profits than Apple. The iPod (which ran an ARM-derived 90MHz CPU) was already out in 2008. Apple had just acquired P.A. Semi. Intel still had vastly, vastly, vastly more free cash flow and profit.
Don't tell me Intel didn't have money for Arm or hiring CPU architects or even literally acquiring entire Arm firms.
Intel had hired these very people who later joined Apple; it absorbed Arm architects, uarches, designers, portfolios, etc.
They failed. As The Register wrote in 2006,
>Intel's failure wasn't for a lack of talent or investment - or even luck - over the years. Intel threw $1.6bn on a DSP company in 1999, and followed up with a host of smaller investments. And luck blessed Intel on several occasions. When DEC's StrongARM processor fell into Intel's hands in the fall of 1997, it was as a result of a legal settlement, and an unsought and unexpected prize. Pundits at the time thought that Intel had been handed the keys to the kingdom. But billions of dollars later, Intel could only claim two significant design wins from lower tier phone OEMs RIM and Palm. Texas Instruments, by contrast, will cash $14bn in revenue from phone chip sales this year. > >So how did Intel fail to capitalize? In a nutshell, it failed to live up to its name. Intel may stand for 'INTegrated ELectronics', but it failed to integrate the electronics that mattered when it mattered. > >A series of poor management decisions ensured that StrongARM was well positioned for a market that was on the decline, and rarely competitive in a market that boomed. Early on, Intel decided against integrating dedicated digital signal processing into the StrongARM chip, later renamed XScale. While this decision was justifiable for fixed embedded markets and for PDAs, it put the chip at a huge disadvantage for lower cost devices that needed voice capabilities. In response, Intel copied the PC strategy of adding new floating point instructions, introducing MMX for Xscale. Since phone manufacturers preferred cheaper custom chips for devices that needed a multimedia flip, this was a wasted investment.
Good reading: https://www.theregister.com/2006/06/28/intel_mobile_failure/
https://en.wikipedia.org/wiki/StrongARM
https://www.computerweekly.com/news/2240049543/Intels-Strong...
Hardware discussion on HN has always been subpar, and Apple's M1 discussion sort of confirm that. I guess most developers dont really give a damn about hardware.
Given that this TSMC 7nm->5nm process improvement has lead to the smallest gen-on-gen performance improvement for Apple's iPhone chips - the performance gains from the A13 to the A14 are the smallest in years - I really don't buy that at face value. Surely a better process helps a substantial percentage, but I think a 7nm M1 chip would have impressed in most of the same ways.
The 5nm chips do better per watt, which is part of the wow factor of the M1. It performs at a high level and delivers industry leading battery life.
TSMC's 5nm must be good enough for Apple to back it.
TSMC is conservative, but, with the continued guaranteed high-dollar buys from Apple, they have tremendous freedom to invest in R&D in an almost fearless way that any company would envy.
This is also why people trying to give all credit to TSMC for the performance of A-series chips are also selling Apple short on that side of the ball there too. TSMC would not be in the position they are right now without Apple's partnership powering their capital investments.
So in short, Apple's backing isn't proof that TSMC 5nm is good, it's actually a large part of why TSMC 5nm is good, and even exists in the first place.
Define better.
Processor designs, as with many things is a question of tradeoffs. Apple is vertically integrated and has the luxury of designing cores based on a different set of requirements that drive the rest of the industry. As Apple makes a profit off per device sold, and not per chip, they can strategically approach chip design differently. For example, the area of die used.
As an end user with an Apple chip in my hands, I get better performance than competing products.
I know what you're saying--all those market factors are why the chip in an iPhone achieves better performance than competing phones.
So maybe you can stretch better to mean "well Qualcomm still manages to extract profit". Bully for them--as a user, an A series chip is better.
Qualcomm is still in business so sure they're moving product and making sales. I'm not denying them that.
Apple makes an objectively better chip based off performance. I really don't care about much else. Are you a Qualcomm stock holder? What's your definition of better and why is it better than using performance?
Apple's "little" cores are roughly on par with the performance of the big cores in a Galaxy S10 while only having the power draw of the current gen ARM "little" cores.
>The performance showcased here roughly matches a 2.2GHz Cortex-A76 which is essentially 4x faster than the performance of any other mobile SoC today which relies on Cortex-A55 cores, all while using roughly the same amount of system power and having 3x the power efficiency.
https://www.anandtech.com/show/16192/the-iphone-12-review/2
Apple has been kicking Qualcomms butt for a number of years now and up until now building ASIC was not primary part of their company. Now that they have moved the Mac platform over to their own chips I believe you can expect to gap to widen.
I hope, truly that this becomes a awake up call for Intel and Qualcomm.
>TSMC announced the first details about its 5 nm manufacturing technology which is going to enable TSMC to aggressively increase its transistor density versus prior generations. However, when it comes to performance and power improvements, the gains do not look very significant.
https://www.anandtech.com/show/12727/tsmc-details-5-nm-proce...
These results could very well be fake, however early in a product release OEMs do often intentionally release these teaser numbers. This would make it the fastest Android chipset by a good margin, and bring it closer to equalling the A12 of 2018.
A dark horse, especially with respect to GPU, is Nvidia after their ARM acquisition. These are interesting times.
[1] https://news.ycombinator.com/item?id=25251229
ARM powered chromebooks run Android apps far better than their x86 cousins plus still include linux support (crostini) which has a growing set of ARM applications every year.
I'm pretty excited for where the future of efficient computing is going. Between my M1 Mac Mini and overclocked Pi4 8GB I've managed to sell off and move on from x86 machines for personal use. Interesting times indeed.
* Move from 4 to 5-wide decoder (A14 is 8-wide)
* Double number of 128-bit SIMD from 2 to 4 (A14 is 4)
* Moving multiplication into a second ALU allows simultaneous multiplication and division (on both A78 and X1)
* Minimum L1 cache at 64kb instead of 32kb with a 64kb option (A14 is 128 d-cache and 196 i-cache)
* L1 bandwidth on A78 and X1 doubled over A77 (probable 2-4x as wide on A14)
* Double L2 cache from 512kb to 1mb (A14 has 8mb L2, but no L3)
* Double maximum amount of L3 cache from 4mb to 8mb (for a potential 9mb cache vs 8mb for A14)
* Still ARMv8.2 (A14 is ARMv8.4 which brings some extra instructions for pointer security, virtualization, JS integer conversion, complex number SIMD, SHA512/SHA-3 hardware, int dot products, and some other stuff -- probably the biggest performance difference here will be the 1-2% for JS conversion)
* Increase BTB from 64 to 96 entries (Not sure about A14, but probably more)
* Increase TLB from 1k to 2k pages (A14 is 3072 pages)
* mops bandwidth goes from 6 to 8 mops (not sure about A14, but 8+ for sure)
* uop cache goes from 1.5k entries to 3k entries (I seem to remember hearing 2k entries for the A12, but I don't know)
* mops dispatch also goes from 6 to 8 (once again , not sure about A14, but 8+ for sure)
* uops dispatch increases from 10 in A77 to 12 in A78 to 16 in X1 (I'd guess Apple's at least this wide)
* reorder buffer increasing from 160 to 224 entries (A14 ROB is somewhere around 630 entries)
ARM's claims that with all three processors at 3GHz, the A78 is 7% faster than A77 and the X1 is 20-30% faster depending on the workload. I'd guess that SIMD float-heavy performance between X1 and A14 is very close.
More interesting is the x86 comparison (let's look at zen 2).
X1 has twice the L1 cache (64 vs 32k), but considering AMD lowered this amount, I doubt it's a real advantage more than different ISA and microarchitectures. Likewise X1 with 1mb of L2 vs 512kb for zen 2 probably doesn't matter too much (AMD redid the cache in zen 3 and stuck with 512kb while Intel still uses 256k). Zen 2 has 16 L0 BTB vs 96 L0 BTB for X1, but this might be microarchitecture more than anything as well as X1 drops to 2k L2 entries vs 7k for Zen 2. I don't know if 8mb L3 is shared in X1, but I'd assume so in which case 16mb for zen 2 is a definite advantage.
The decode width is now 20% wider than zen, but x86 instructions are around 15% more dense, so I'd guess it's about the same. X1 has 3k uop cache entries vs 4k on x86, but I suspect x86 needs more to avoid hitting those complex decoders at all costs. AMD dispatches 6 int uops and 4 float uops per cycle while X1 does 16 uops per cycle.
I don't know about ARM, but AMD's perceptron predictor is probably better (it seems like Apple or ARM would make a big deal about this if they'd made the investment -- Samsung did). Given AMD has 19 stages instead of 13 in the X1, a better predictor is needed to offset penalties. X1 and Zen 2 both have identical 224-entry reorder buffer.
X1 has 2 AGU, 1 AGU/load and 2 load/store while zen 2 has 3 AGU, 2 load/store, and 1 store. Zen 2 and X1 AGU are both 256-bit wide. I can't find store queue length for X1, but bigger queues could be an advantage to AMD (I suspect so as ARM probably wouldn't want the extra power draw).
A76 seems to have similar instruction cycle counts to Zen 2 and I assume they haven't gotten worse since. X1 has four 128-bit SIMD while Zen 2 has two 256-bit SIMD. If you can use wider SIMD, AMD has an advantage, but if there's lots of smaller SIMD, then X1 should have an advantage).
Zen 2 has 4 integer ALU with one capable of multiplication, another capable of division, and another with a CRC (cyclic redundancy checking). X1 is identical (except I don't know if it has a CRC).
X1 very likely comes close to Zen2 and even A14 in integer performance while being a bit ...