Qualcomm Snapdragon 845 Benchmarks Show An Incredible GPU, Faster CPU (hothardware.com)

MojoKid writes: Though the company has been evangelizing its new Snapdragon 845 Mobile Platform for a while now, Qualcomm is lifting the veil today on the new chip's benchmark performance profile. At the heart of the Snapdragon 845 is the new Kyro 385 CPU, which features four high-performance cores operating at 2.8GHz and four efficiency cores that are dialed back to 1.7GHz, all of which should culminate in a claimed 25 percent uplift over the previous generation Snapdragon 835, along with improved power efficiency. In addition, the Snapdragon 845's new Adreno 630 integrated GPU core should deliver a boost in performance over its predecessor as well, with up to a 30 percent increase in graphics throughput, allowing it to become the first mobile platform to enable room-scale VR/AR experiences. Armed with prototype reference devices, members of the press put the Snapdragon 845 through its paces and the chip proved to be anywhere from 15 to 35 percent faster, depending on workloads and benchmarks, with graphics showing especially strong. Next-generation Android smartphones and other devices based on the Snapdragon 845 are expected to be unveiled at Mobile World Congress in Barcelona at the end of this month.

Comments

Sigh.;

[furiousgeorge]

"Incredible GPU" (I'll just leave for a moment this is a chip just to be maybe 'unveiled', while the the latest iOS chip which you can buy for 6 months is already way ahead as usual...)

Unless they've fired their entire OpenGL/Vulkan driver engineering department and started over, I can't get excited. It'll just be *another* big bag of pain and busted features.

As God as my witness, I wish somebody would make the investment to give Qualcomm some actual competition, cause they are a nightmare.

Signed : Mobile Games Graphics Engineer.

Re:Sigh.;

[Hal_Porter]

Lies.

Finally we will crush the lamers who just use ARM IP unmodified with our brilliant and revolutionary Kryo core, the crown jewel of Qualcomm which we've spent a fortune optimising.

Wait, actually we're going take ARM A73 cores, hack them around and just call it Kryo and just pull in a cheap ARM GPU hard macro unmodified. Actually we've pretty much given up on doing microarchitecture design and are gradually turning into a company that just takes hard macros from ARM and puts them on chips. Oh dear.

Yours,

A Qualcomm Engineer.

Re:Sigh.;

[jareth-0205]

"Incredible GPU" (I'll just leave for a moment this is a chip just to be maybe 'unveiled', while the the latest iOS chip which you can buy for 6 months is already way ahead as usual...)

Unless they've fired their entire OpenGL/Vulkan driver engineering department and started over, I can't get excited. It'll just be *another* big bag of pain and busted features.

As God as my witness, I wish somebody would make the investment to give Qualcomm some actual competition, cause they are a nightmare.

Signed : Mobile Games Graphics Engineer.

I mean it could come from Apple but they're too busy walling off their garden. Full marks though, after years of marketing bluster that didn't really stack up, Apple seem to have created a monster CPU. It's just a shame it's trapped only in their devices.

Hmm.

[Presence Eternal]

Is it stupid of me to wonder why Intel/AMD don't do the whole "two fast cores, and lots more slower cores" bit?

Re:Hmm.

[Hal_Porter]

Qualcomm doesn't own big.LITTLE. The original implementation was by ARM. In fact Qualcomm stuck to having one Snapdragon core for both high performance and low power tasks until fairly late. E.g. the Galaxy S5 had two models. The Qualcomm Snapdragon one didn't do big.LITTLE/Heterogeneous Multi Processing.

https://www.cnet.com/news/the-...

For some Galaxy S5 models, Samsung will use the Exynos 5422, which also was announced at Mobile World Congress. The processor boasts eight ARM cores versus the four in the Snapdragon 801. Samsung has employed ARM technology that allows for four big cores that run at speeds up to 2.1GHz and four small cores for speeds up to 1.5GHz. When the phone requires heavy computing, all eight cores can run at the same time. The phone can also employ just one of the small cores for minor activities

And companies other than Qualcomm and ARM have patents on various implementations of big.LITTLE

https://patents.google.com/pat...

Also I'm sure Intel could do HMP. E.g. they could have a chip which has some Coffee Lake i5/i7 type 'big' cores and some Gemini Lake/Goldmont Plus Atom type 'little' ones. Sure there are patents but Intel have a patent portfolio of their own. They could just launch, and if they get sued agree on a cross licence.

NVidia sued Intel and got some cash but no x86 licence or the right to make chipsets for newer Intel CPUs.

https://www.anandtech.com/show...

And of course a hypothetical Intel chip with Coffee Lake/Gemini Lake might not work like big.LITTLE. On a big.LITTLE chip you can migrate threads from a big core to a little core quickly because they share the same registers. A Coffee Lake/Gemini Lake chip might not need to do that - you could have cores sharing an L3 cache but not much else and just relying on the OS to do the migration. And if you look at the Lenovo patent the patented thing is not 'having big and little cores on the same chip' but rather the way you decide whether a thread runs on the big or little core. Intel could invent their own scheme. Or just leave it to the OS.

Actually it's not a bad idea. The USP of Windows on ARM is 'better battery life'. Unfortunately that comes with Atom like performance. A Coffee Lake/Gemini Lake HMP design would offer the low power of an Atom when the system is idle but also the high power of a Coffee Lake when it is not. And you can have a whole bunch of Gemini Lake cores in the space taken up by a Coffee Lake one. So you might sacrifice 1-2 Coffee Lake and have 4-8 Gemini Lake in the same space.

If Intel can spend R&D on a package which has Coffee Lake and an AMD GPU on it which is basically just for Apple, I reckon they could spend R&D on a HMP chip to attack Qualcomm with.

Actually another option would be to dumb down the integrated graphics. I bet you could get away with just a frame buffer for most users who are just running a GUI and not doing an 3D. So gate all the 3D stuff with a Mosfet and make sure you can power it down.

Intel clearly cares about power consumption and battery life and Windows on ARM shows they've got some way to go to yet. HMP and more work on optimizing the power consumption of the integrated graphics solution in the lowest power/performance state is the way to go.

Re:Hmm.

[TheRaven64]

Three reasons. The first is that Intel has control over their process, which currently gives them a big advantage in the analogue electronics side of chip design. ARM cores have to be synthesisable on any process and even Qualcomm likes to have a couple of options for fabs so that they can negotiate a good rate. This means that Intel can do clock and power gating at a much finer granularity than anyone else.

The second is that the workloads where big.LITTLE makes sense are fairly limited. If you're buying machines for a datacentre, they're consuming so much power just by existing that if you're not using them at close to capacity then you're wasting money. Even on laptops, the CPU in its lowest power state is often using less power than keeping the RAM refreshed. It's only on very mobile devices that it makes sense to have a simple in-order core or two that can do background processing while the main core sleeps and the user isn't actively doing anything. Apple is apparently considering adding a small ARM core for this purpose to laptops (for things like checking email while the machine is in suspend mode, for which it currently wakes up the main CPU for a few seconds). It's likely to add a little bit of time to the suspend life of a laptop, but most laptop users care about the time that the machine can be on and in use a lot more than they care about standby time. In contrast, mobile phone users care a lot about standby time.

Finally, there's a much higher lower-limit on the complexity of an x86 chip. A low-end ARM core has no microcode. It can have a trivial branch predictor, because the pipeline is so short that the mispredict penalty is small. The decoder is trivial. The cache controller is simple, because there's no implicit synchronisation between instruction and data cache. In contrast, an x86 chip needs to have a bunch of microcode, an instruction parser (x86 instructions are so complex that it's a stretch to call it a decoder) which has little area impact on a high-end superscalar core but adds a significant fraction to a trivial core, and must ensure that it has enough store forwarding that writes are instantly observable in the decode stage.

Random tech shows incremental improvement

[xxxJonBoyxxx]

>> random chip company releases numbers to buttress its marketing claims

Nice, er, incremental improvement. Now, if you don't mind, I have a nap to take...ZZZzzz

Re:Random tech shows incremental improvement

[Lunix Nutcase]

The bigger issue is real devices won’t see as much gain because the prototype devices used to benchmark are likely to have a much larger thermal headroom than the average/tablet using this CPU.

Still better than some alternatives.

[Anonymous Coward]

Adreno has at least some open source stack, as does Tegra K/X via nouveau, vivante via etnaviv. VC4, while anemic has signifcant support as well as general enough purpose to run OCL code on (albeit not securely thanks to the lack of an mmu.. but then GPU code has the potential for just as much risk.

Honestly the worst is still Imagination Technologies or whoever has inherited their PowerVR GPU architecture, followed by ARM with their cut rate Mali drivers.

Seriously, these assholes just need to openly document this shit then either provide some code submissions, some development hardware, or hire a few open source driver developers to do the jobs their own staff obviously can't.

Re:How does that compare to ...

[Anonymous Coward]

Not well, on a per core basis, here's Geekbench scores for single threaded:

Apple A11 @ 2.4GHz: 4246
Core i7 8700k @ 3.7GHz: 6245
Ryzen 1800X @ 3.6GHz: 4127
Qualcomm 845 @ 2.8GHz: 2476

So per clock:
Apple A11: 1777 Geekbench points per GHz
Core i7 8 series: 1687 Geekbench points per GHz
Ryzen: 1146 Geekbench points per GHz
Snapdragon 845: 884 Geekbench points per GHz

Long story short - Apple's current ARM CPUs are in the same ballpark as high end desktop CPUs for single threaded work (but way off for multithreaded), while Qualcomm's are roughly half as fast. Apple's ARM CPUs have the best IPC of any major CPU on the market just now.

Snapdragon 845 Is King (Of Android Phones For Now)

[Humbubba]

Quote from GizmoChina:

So how does Qualcomm's new chip perform against those in the market currently? Long story short, it is not the king... It was ran through benchmarking apps Geekbench and AnTuTu and then pitted against other phones and chipsets. The test device was compared to the Huawei Mate 10 Pro with its Kirin 970 SoC, the OnePlus 5T with Snapdragon 835, the Exynos 8895 toting Galaxy Note8, and the Apple A11 Bionic iPhone X...

Qualcomm's new chip beats all but one - the Apple A11 Bionic. Apple's chipset not only trumps it but does so with at least 2000 points in both the single-core and multi-core tests. Qualcomm's joy as the king of Android chipsets will actually be short-lived as the Exynos 9810 is said to be ahead in performance too.https://www.gizmochina.com/2018/02/12/snapdragon-845-battles-snapdragon-835-exynos-8895-kirin-970-apple-a11-bionic/

Re:How does that compare to ...

[Bert64]

Apple only use their CPUs in small, passively cooled battery powered devices... Would be interesting to see how they could scale if coupled with the typical cooling and power supply that an i7 or ryzen chip uses, or if they could add more cores.

Next big thing for Gaming

[jamesfelix]

I think snapdragon 845 will take over the big games. Affordable Salon management software

Re:Who cares?

[EETech1]

I just got the Moto E4 plus
5000 mAH battery!
Right now...
  Last full charge 1 day and 18 hours ago.

79 percent charge approx 1 day and 23 hours left.

I didn't charge it last night, and I'm not charging it tonight. It has quick charge, so it only needs a couple hours to fully charge anyways.

Easy to unlock (not Verizon model) from Motorola, root it, do whatever you want with it!

Incredible GPU?

[ShamblerBishop]

Damn, and I should bought an 1080Ti - I better chuck that in the bin now and wait for my new phone.

Re:How does that compare to ...

[TheRaven64]

Are you really telling me that apple who has gone from no chip to having a cpu in the last couple years has all the top CPU talent and companies like Intel and AMD who have been at this game forever are really loosing out to them?

Note that Apple bought PWRFicient 10 years ago. This gave them a bunch of people with 10-20 years of CPU design experience, so it's not as if they had a standing start. Their first (recent) in-house-designed CPUs were in pre-Touch iPods, so iPhone CPUs weren't even their first ones.

That said, I'm still skeptical of the Geekbench numbers. The Apple cores are pretty impressive, but I don't believe that you can get higher IPC in a passively cooled CPU in a phone than Intel gets from a 95W core.

Re:Who cares?

[TheRaven64]

You know the best way to get good power usage? Finish what ever compute needs doing quickly and put the device to sleep. A faster CPU in the same power envelope means that it can have the network and RAM out of their low-power states for less time, which translates to better battery life.

Re:How does that compare to ...

[drinkypoo]

Apple only use their CPUs in small, passively cooled battery powered devices... Would be interesting to see how they could scale if coupled with the typical cooling and power supply that an i7 or ryzen chip uses, or if they could add more cores.

x86 doesn't scale down, and ARM doesn't scale up. You could have more cores, but you couldn't practically use them. The GPU is actually pathetic (30% faster than shit is just slightly faster shit) and would impress nobody.

Re:Snapdragon 845 Is King (Of Android Phones For N

[AmiMoJo]

I'm starting to think that Geekbench scores for the iPhone are bullshit. The A11 Bionic has 2 high performance cores, but somehow out performs chips with 4 high performance cores. Yet iPhones don't appear to be any faster than Android phones, and in fact they are often quite a bit slower in real world use due to having only 2GB of RAM.

Even if Apple has somehow managed to get >2x the performance per core, Geekbench seems to have little relation to real performance.

Re:How does that compare to ...

[imgod2u]

It's higher IPC in Geekbench. IPC varies by workload. Note that there isn't that big of a difference in IPC between the 5W i7-7Y75 (Kaby Lake) and the 91W i7-7700k (Kaby Lake).

You could walk away from that data point thinking that there's no difference between the 91W desktop processor and the 5W laptop processor. Or you can walk away thinking that Geekbench is a limited benchmark that only measures a tiny subset of processor functions.

I doubt Apple optimizes for Geekbench. But whatever internal iOS workload they do optimize for apparently has similar stress points as Geekbench.

Re:Who cares?

[imgod2u]

In reality though, all a faster CPU gets you is software developers that write more complicated code. So you end up with the same amount of absolute time doing compute work but with much higher energy use.

But yes, if you ran iOS 7 (or Jelly Bean) on a modern phone CPU, it'd be both blazing fast and very power efficient.

Re:Snapdragon 845 Is King (Of Android Phones For N

[Humbubba]

AmiMoJo said

I'm starting to think that Geekbench scores for the iPhone are bullshit. The A11 Bionic has 2 high performance cores, but somehow out performs chips with 4 high performance cores. Yet iPhones don't appear to be any faster than Android phones, and in fact they are often quite a bit slower in real world use due to having only 2GB of RAM.

Many have criticized the Geekbench processor benchmarks, unbelievably, even Linus Torvalds. But he relented with version 4.0, saying it looks much better. Version 4.2's GPU test fixes put it in line with OpenCL and CUDA results. I don't see any problem.

I've not tried either the SD845 nor the A11 Bionic processors. If you have, you're a better geek than me, which isn't saying very much. I'm sure you're right about the 2GB bottleneck. As I look over their different specs, there are two other things that stand out in the SD845's favor: the GPU, and the core\cache organization.

1) Snapdragon 845's has modest CPU improvements over the SD835, but the GPU upgrade is 32%-40% better, depending on the graphics test. And it beats the A11 in all but two of those tests.

2) The A11 does have a better CPU performance than the SD845, hands down, but there may be more to it than that. The A11 can use all six cores simultaneously, and has AI hardware called a "Neural Engine" that can perform 600 billion operations per second. Some or all of this may help explain why it's a speed demon at multi-core tasks. But not so much at single-core tasks. Just guessing, but maybe that's because it has discrete core clusters and caches. In contrast, the SD845 uses ARM's DynamiQ CPU cluster organization, letting different cores be hosted within the same cluster and cache hierarchy.

That's all I got, except the links below.

http://bgr.com/2017/09/14/iphone-x-vs-iphone-8-a11-bionic-benchmarks-macbook-pro/

https://www.neowin.net/news/qualcomms-snapdragon-845-benchmarks-show-massive-gains

https://www.anandtech.com/show/12420/snapdragon-845-performance-preview

https://www.geekbench.com/blog/2017/11/geekbench-42/

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

https://www.extremetech.com/mobile/263774-qualcomms-snapdragon-845-strong-gpu-performance-less-cpu-improvement-advertised

Re:Snapdragon 845 Is King (Of Android Phones For N

[Dixie_Flatline]

The iPhone X and 8 Plus have 3GB of RAM, only the iPhone 8 has 2GB.

iPhone apps are rarely RAM limited, and indeed, when you watch those speed tests on YouTube where they open a whole bunch of apps to see how fast each phone can process things, iPhones win will considerable regularity. The iPhone 7 was winning those tests right up until the iPhone 8 was released.

iPhone single core performance has always been better than the multi-core, because most day-to-day tasks on phones don't parallelize well, outside of games.

If you think that iPhones are underperforming in the real world, let alone because of RAM issues, I'd like to see a test or some evidence for that other than, "I feel like this phone is slower." iOS has notoriously elaborate transition animations which can make it feel like it's lagging, even though it's usually prepared to accept your input before you can see what you're tapping at.

But to a large degree, nobody will ever notice the speed difference in most situations because most tasks just aren't intense enough to slow down a modern iPhone OR Samsung or Pixel. The only real time that comes up is in stuff like image processing, and devices like the Pixel are now running custom silicon JUST for that, which makes a difference for both quality and speed.

Re:Snapdragon 845 Is King (Of Android Phones For N

[AmiMoJo]

The main issue with having little RAM is that you end up waiting as apps reload after being forced out. Even within apps such as the web browser some parts might get swapped out, e.g. background tabs.

The next biggest influence on phone performance is flash memory speed. Samsung is king, but other phones are quite good too. iPhones are competitive.

Re:Snapdragon 845 Is King (Of Android Phones For N

[Dixie_Flatline]

Until recently (that is, the iPhone 8, 8 plus and X), iOS tended to be *better* about keeping apps in memory. If you watch any speedtest involving the iPhone 7, it decidedly trounced any other handset, even with only 2GB of RAM. (Web tabs have always been reloaded, which is good or bad depending on your point of view—I usually want the tab reloaded anyway.)

Anyway, it's still *very* close between the iPhone 8 plus and the Note 8, for instance. Despite having twice as much RAM, the Note 8 only barely beats the iPhone.

THAT SAID, I think Apple could probably afford to put more RAM in if they wanted to. It's a bit of a tradeoff, and over time, I think we'll see the amount of memory increase. iOS is just really memory efficient in general, so it'll probably always lag Android devices, but if Apple wants the biggest and best games developed on its platform, they'll have to yield here and just jam more in. But it still doesn't represent a major loss of performance for most day to day use.

Re:Snapdragon 845 Is King (Of Android Phones For N

[AmiMoJo]

I dunno, the X seems noticeably slower than the Pixel 2 in many regards. Especially where RAM really helps like keeping the camera app in memory all the time for speedy access.