I've run experiments/developed on both Applied Micro's X-Gene 1 and Cavium's ThunderX (the original). Both of these have been touted as "ARM servers".

The X-Gene was complete crap in terms of both performance and power consumption. The ThunderX was only really useful for highly-parallel applications that don't require a lot of horsepower, and even then it's a power hog.

I'm cautiously optimistic with Qualcomm's Centriq and Cavium's ThunderX2, but unless they can make a compelling argument in terms of power consumption, I'm skeptical of them making any inroads into the market. Sure, Intel Xeons and IBM Power8s may suck power, but they're _damn_ fast, so you can execute your workloads quickly and then just shut the machine down or enter idle power states.

I am not sure if you have seen this review from Qualcomm's Centriq by Cloudflare, but at least for others it might be interesting: https://blog.cloudflare.com/arm-takes-wing/

In the conclusion they mention: "The largest win by far for Falkor is the low power consumption." (Falkor is the name of Qualcomm's Core).

Not to mention the amount of hardware standardization that still needs to happen. Having a different image for each manufacturer's hardware is going to get real old, real fast.

5+ years on, ARM on anywhere but mobile seems like a dud.

You don't need a different image for different manufacturers' platforms. One Linux image works across vendors. The standardisation effort started some years ago resulting in SBSA/SBBR, which allows a common image to be used.

I really want to see someone attempt an Android Treble-like project for ARM Linux distros.

I just wrote a Device Tree Overlay for a Real Time Clock module on FreeBSD RPI3. There is something to be said for PNP hardware.

Who would imagine that the computer industry would go so far back in time as to leave off the RTC.

Oh but wait they added Bluetooth and Wifi. Yet no system time. Still waiting for a PCIe bus as well... Guess I need some more of that koolaide they are drinking.

> Still waiting for a PCIe bus as well

RPI3 is not a 'server grade' ARM -

See e.g. the AMD 'Opteron A' boards:

http://www.lenovator.com/product/103.html#params

    CPU Quad-core ARM Cortex-A57 64 bit
	
    DRAM Two DDR3 SO-DIMM sockets
	
    SATA Two SATA ports
	
    USB Two USB 3.0 ports
	
    Console USB-micro port for console support
	
    Ethernet 1 GBe Ethernet
	
    PCIe x16 PCIe G3 slot
	

not sure about the RTC, but the PI is way far from any of the arm 'standardization' efforts, being a low end 'standard' unto itself..

Not all workloads require powerful CPU's: there are plenty of servers that are bandwith-limited (either network or I/O), like static webserving or mail-relay, and plenty workloads where performance is of secondary concern (like out-of-band batch processing, backup systems).

Yesterday: ARM in your phones

Today: ARM in your toys (RasPI, Nintendo Switch)

Tomorrow: ARM in your servers

Next week: ARM in your desktop/laptop?

The iMac Pro will have an ARM co-processor: https://www.macrumors.com/2017/11/19/imac-pro-a10-chip-hey-s... - Arguably for "Hey Siri" and other mobile-inspired functionality, but wouldn't it be interesting if it were also an ARM-on-the-desktop beachhead?

I, for one, welcome our new ARM overlords.

> The iMac Pro will have an ARM co-processor: https://www.macrumors.com/2017/11/19/imac-pro-a10-chip-hey-s.... - Arguably for "Hey Siri" and other mobile-inspired functionality, but wouldn't it be interesting if it were also an ARM-on-the-desktop beachhead?

The current MBP already has ARM: https://techcrunch.com/2016/10/28/apples-new-intel-driven-ma...

But this is similar to ARM (or MIPS or whatever) cores that already exist in all sort of peripherals (what's in your Wifi chipset? Bluetooth? SMC?), so it's nothing new.

I'm dismissive of any talk of switching conventional Intel-based desktop/laptops to ARM. It's not equivalent to Apple's switch from PPC to Intel, because in that case there was a clear performance advantage. ARM may have the performance-per-watt crown on the low-power end, but Intel maintains the pure-performance lead everywhere else. Conventional desktops still care about performance (and if you don't, likely you can switch to a Chromebook), and there aren't any advantages to switching a heavily Intel-ISA-dependent desktop ecosystem (macOS/Windows) to ARM.

I care about performance in my phone too, so where can I buy an Intel phone?

The truth is that my phone doesn't have to be the fastest; it just has to be fast enough in balance with other factors important to me (size, battery life, etc).

And the same goes with my laptop. I currently have an MBP with an i7 processor, and I'm a web developer - it's overkill. I also play games with it now and then, but it's clear that the iGPU is the bottleneck there and not the processor. Very rarely do I do heavy compiling or video encoding or other CPU-maxing tasks on this thing.

So would I trade the i7 for an A-series plus a couple more hours of battery life and/or a lower (or at least not further increased, damn it Apple) price? Yes, I would take that trade, and probably never suffer for it.

I've had Intel phones running Android. For example the Asus ZenFone as mentioned below, also the Lenovo K80. [0] I also used to have a Motorola RAZR ages ago, that I think was one of the first Intel powered phones? Not just phones, watches with x86 series Intel CPUs exist! The Tag Modular 45 [2] has an Intel Atom Z3400 [3] in it...

0. https://www.gsmarena.com/lenovo_k80-7216.php

1. https://en.wikipedia.org/wiki/Motorola_RAZR_i

2. https://www.tagheuer.com/en-us/watches/tag-heuer-connected-m...

3. https://en.wikipedia.org/wiki/Atom_(system_on_chip)

Why you bought a laptop with an i7 then in the first place? There are already more energy efficient alternatives - i5, i3, core m.

I have a theory that probably when Apple will release a laptop with their own SoC it will be more expensive...

Because it checked the other boxes I was looking for with regards to screen size, memory, and so on.

Perhaps an A-series laptop would be more expensive for consumers (damn it, Apple), but I don't think it would cost more for Apple to manufacture, since they can do so much more of it in house.

Asus Zenfone smartphones use Intel processors.

While I agree with you that we almost certainly won't see an ARM macintosh, I wouldn't be surprised to see an ARM laptop exclusively running iOS at some point. We probably will see an ARM core with single threaded performance in Intel's realm once Intel is fully backed against the wall that is the end of Moore's law. It'll probably always be slower than an intel core if it has to emulate x64, but the iOS ecosystem gives Apple control of it's on chips in a world after the commodification high performance chips.

> an ARM laptop exclusively running iOS

So an iPad with a Bluetooth keyboard? Apple already sells those!

It's a little more than a stretch to refer to a current iPad with Bluetooth keyboard as a laptop. But sure, it shows how Apple is inching in that direction.

Well, it's a stretch for me to consider iOS as a usable laptop OS, so I guess we're both stretching. :P

But to play the devil's advocate, how is it not sufficiently laptop-ish for you? Because the keyboard is detachable? Because it has a touchscreen?

I think it will morph into a usable laptop OS as Apple pushes it into that direction.

> But to play the devil's advocate, how is it not sufficiently laptop-ish for you? Because the keyboard is detachable? Because it has a touchscreen?

Lack of a mouse equivalent. Having to move your hands off the keyboard to do anything other than type is a pain in the ass.

"Intel maintains the pure-performance lead everywhere else"

The point of the article is that ThunderX is claimed to be faster than Xeon Gold for HPC applications.

I see one HPC application where ThunderX 30% is faster (TeaLeaf). For the others shown, it's neck-and-neck with the Xeon. Reaching parity with Xeon is admittedly impressive even on this subset of applications, but it doesn't look consistently faster based on the information presented. I'm personally interested in how this system stacks up for quantum chemistry (GAMESS, NWChem, PySCF, Psi4...) and I hope that more benchmarks will become available soon.

But no where near single threaded performance.

ARM is a bit of a misnomer, there are many segments, Cortex M, Cortex A. Only the "A" series could be considered intel/amd class I guess.

But yes, arm cores are everywhere.

> Next week: ARM in your desktop/laptop?

And it's not just coprocessors, next year we will see real ARM SoCs - https://www.theverge.com/2017/10/18/16495010/microsoft-windo...

Next year^H^H^H^Hdecade: RISC-V in your ______

I'd love to see RISC-V become a huge presence, but ARM IP seems a lot more refined & hassle-free.

A bit similar to Linux/FOSS not having any decent CAD alternative to proprietary ones.

OpenSCAD is very good, having designed quite a few things using it, and FreeCAD is quite serviceable for working with legacy shapes made in other CAD software. The ecosystem has definitely improved thanks to 3D printing.

To be fair, ARM has been in Mobile Nintendos since 2001.

And laptops have had ARM coprocessors for quite a while. Power management controllers, disk controllers, NICs, etc.

Hey yeah, I remember my first ARM programming was on the NDS' two 66/33MHz processors. I think it basically dedicated one to each screen, fun stuff.

Are the laptop co-processors you're talking about the 'R' lines? I know 'M' is for low-power mobile stuff and 'A' is for "Real Computer" stuff, but I was never too clear on the 'Realtime' lines.

I've seen all three in that use case (and ARM11 cores and earlier where they didn't really make the distinction as overtly). Mainly R and M though.

But the distinction is something like:

M - Thumb2 only, simple two or three stage pipelines typically. Optional everything like selectable (at chip design time) multiply unit that's either 1 or 32 cycles of latency. MPU (if present) only adds protection information, but doesn't remap memory like a full MMU. An ascetic M has something like 12k gates which is insanely low for a 32 bit processor. It mainly competes with AVRs, 8051s, and maybe really low end MIPS if anyone is still using R3000s. In a laptop, I've seen them in power management controllers, bluetooth, and sometimes in WiFi.

R - More powerful than M. Runs Thumb and A32. Clear, deterministic. In order pipeline, tightly coupled memory, and an MPU so that you can still run hard real time cycle counting kind of code twiddling GPIOs carefully, but then also running DSP on the results afterward. I've seen a lot of these in hard disk controllers. They're sort of spiritual successors to the ARM9/7 you know from the DS. Competes with most MIPS.

A - Sort of all over the place. Some are 32bit, some are both 32 and 64. Has a real, page table MMU. Typically run real, big boy kernels like Linux. Competes with Atoms, higher end MIPS (see a trend?) and POWER at the very high end if you squint enough. Typically they're complex enough that you can't cycle count and get deterministic results (longer pipelines, TLB that'll refill 'randomly', cache refills, more complex branch predictor), but give higher overall performance. I think I've seen an A8 as a sound processor? Typically though they're not really a coprocessor, the gate count is too high to make sense.

Huh, cool. I've actually been playing with the STM32 Cortex-M processors, and I gotta say - some of those high-end M7 chips could easily run a game system and are way out of the range of 8-bit Arduino-y chips. 400MHz with a 24-bit parallel LCD peripheral? That could probably give the PSP a run for its money.

And they're moving to a ~40nm process on those performance lines soon.

They're a lot more powerful once you start utilizing their hardware peripherals like DMA. With a STM32F4 it's possible to saturate a 100mbit PHY that pushes out that data to 100+ GPIOs while using all of the CPU just for bit interleaving and control signals. If your use case allows it, these microcontrollers can punch far above their weight with dedicated hardware when compared to general purpose CPUs

ARM's already been in most of those places, in nearly reverse order of your timeline. It started as a desktop CPU, was then in a PDA (Newton), game console (3DO) and phone (Nokia) - all of them by 1997.

Unless the iMac Pro (unlike the current MacBookPro) has a generally addressable ARM core, it's not news that they're using ARM for the purpose of overall ARM dominance.

It's like when people were all excited they were testing LiquidMetal in their SIM ejection tool (not sure if that was apocryphal).

ARM has been in Chromebooks for a while. 90% of the desktop has been Microsoft Windows for about 2 decades so it's really up to Microsoft to try again.

Ideally, Android and iOS will evolve so phones and tablets replace the desktop for more people.

ChromeOS is already capable of running Android apps.

What makes a laptop a laptop is a decent keyboard and a reasonably large screen. There is a fundamental mismatch between the screen size comfortable for the eyes, keyboard size comfortable for typing, and device size comfortable for holding in hand. Because of this, I don't think phones and laptops will ever converge; tablets and laptops are already actively converging, though.

So, you simply need to add an optional large screen, keyboard, and mouse to your phone. It has already been done. Now it just needs to be done a little better?

https://www.samsung.com/us/explore/dex/

See also: Microsoft Continuum, Motorola Atrix (probably the first Android that turned into a laptop)

Yay, and with Emui and DeX growing up, it seems to be next revolution about smartphones replacing your PC's. For instance, the Mirabook from Miraxess : https://miraxess.com/samsung-note-8-laptop/

Someone should make a discrete gpu with an arm cpu core or 50 sharing the same memory and maybe ssd. Then you have a really nice hierarchy of parallelism, 2 fast arm cores, 10 medium, 50 slow, then gpu.

Would be great for stuff like multigrid methods where you are balancing parallelism and latency:

https://devblogs.nvidia.com/parallelforall/high-performance-...

I'm very happy about this development as well. Especially looking at Intel's fiasco with ME... I just don't trust proprietary chip makers anymore.

And yes, I am willing to trade in performance, for more security. My MacBook should be every bit as power conserving as my iPhone, perhaps with a GPU for multi screen and playing videos.

Who will make the SoC if not the proprietary chip maker? ARM does not make this thing better.

I guess my point is that the market would punish anyone for being foolish enough to do that in the form of competitors without egregious backdoors. Whereas with Intel... it is hard if not impossible to design/make competitors (AMD being possibly the sole exception)

I had the idea about a beachhead too but I think it would be different. What if you had an A10 to do all of the iPhone features that you would want on your laptop or desktop. Then use the Intel for the performance . You would increase battery life and both platforms would have feature parity (something I would like a bunch ).

Not just for Macs, but for full Windows 10 laptops, too (soon):

https://www.theverge.com/2017/10/18/16495010/microsoft-windo...

Even cheap SD cards have been running on ARM for a while https://www.bunniestudios.com/blog/?p=3554

Finally, a chance to have a desktop CPU without a black-box coprocessor like Intel ME? (A quick web search did not find anything definitive about its presence / absence in Cavium.)

Every ARM SoC is already a black-box containing black-boxes. That's one of the reasons why almost all SoCs are on some ancient version of Linux kernel. They will never be mainlined.

Next week + 1 day: your laptop is no longer supported

Which ARM SoC has mainlined drivers in Linux kernel? I fear that the ARM trend will only help to make the the e-waste mountain higher.

ARM64, at least, is moving towards UEFI/ACPI for booting/peripheral support, so ideally a lot of things will "just work" once people are shipping mobos/silicon with it enabled (and they already are).

That said it's unclear what the volume on those parts will be once higher-grade CPUs come out, which will impact pricing/availability. And newer parts are going to be necessary for competitive performance/power. If the only place you can get a 14nm Thunder X2 is some high-end SuperMicro workstation, it'll be somewhat limiting in the end, anyway. (You can buy a Thunder X1 prebuilt desktop on the order of $1500 USD right now, at least.)

ARMv7 of course is a different bag of worms; Mali tends to ruin everything. But if you can avoid it, a lot of systems can get by with upstream U-Boot and upstream Linux... You can even get some cheap Asus chromebook laptops that can run libreboot and upstream everything (including WiFi if you buy a dongle.)

That’s exactly why the Arm servers adopt the same specs or functional equivs to x86. A single binary distribution of an OS will boot on any SBSA compatible machine - because all the weird bits are hidden through ACPI, UEFI and PSCI. No BSPs.

I’d go as far as saying Arm servers are compatible to x86 designs at a platform level (eg. how PCIe and config space are wired up, how caches and DMA work, NUMA, SMP, MSI(X), and so on....

Is there any hope that it could be used to some extent on smartphones? Or maybe there is a similar effort in that space?

I understand that some works may be hard to miniaturize.

IMO yes. Although SBSA and SBBR target the servers, already smaller gear (think vCPE, set top box and IoT gateway chips) are adopting the specs as well, and I anticipate adoption everywhere except devices where a) cost is everything b) power and perf need to be squeezed out even at the cost of maintainability. At some point following the everything-needs-its-own-BSP mentality for embedded SoCs will start hurting the bottom line, because there will be plenty of competitors willing to make more “normal”-looking gear.

This looks even more promising than Qualcomm's chip?

https://blog.cloudflare.com/arm-takes-wing/

Either way, competition is definitely heating-up in the server space. If I were them I would price these chips super-aggressively (near-cost) for first few generations to get rapid mass adoption and software support. Worry about profits later, when they are established as serious players in the server chip market.

I'd sort of been expecting that ARM would begin making HPC inroads when SVE[1] was ready to go. But I guess those two Neon units provide enough horsepower?

[1]https://community.arm.com/processors/b/blog/posts/technology...

They don't for fp32 or int8 workloads until SVE IMO. High end Skylake as an example has 2 AVX-512-F ports (and not considering AVX-512-4FMAPS), which give you do 64 FLOP/cycle (512 / 32 * 2 * 2). Even 2 NEON units, assuming FMA's execute in 1 cycle/unit, gives only 16 FLOP/cycle. For ML inference workloads you can even get away with lower precision and even do away with floating point (which helps motivate AVX-512-4VNNIW).

From a HPC perspective, I think the current crop is interesting mostly [1] for tire kicking, software porting etc. Real production deployments will start to happen once SVE is available.

[1] Though they could be interesting for applications that are memory BW bound and don't vectorize that well.

It seems like Intel/ARM processors are just needed to launch CUDA kernels these days.

ARM on desktop or server will be tightly controlled closed SOCs with zero driver support. These will be nothing like the relatively open systems in the PC world, so there is nothing to look forward to.

This is ARM's model and they have shown no interest in improving things beyond kicking open source developers between the ARM and SOC vendors each blaming the other for years on end.

Intel, even with something as nasty as ME, look like an angel in comparison.

That’s not true at all.

That has been true only for the embedded and mobile space.

The server space is exactly being made open and compatible between SoCs (in the same way AMD and Intel machines are). Platform crud (i2c, gpio, i2c, clock trees and pinnmuxes) is hidden in firmware behind ACPI, PSCI and UEFI just like on x86. SBSA and SBBR specs mandate stringent compatibility requirements. Machine error handling is firmware-first. IO is basically PCIe now.

For an OS vendor or an IT guy these machines look almost exactly like an x86 box. And then they look the same between the Arm vendors too.

> This is ARM's model

ARM's model is to develop/license the ARM processor ISA - What the vendors choose to do with this is their own business.

If anyone wants to play around with a 96-core Cavium ThunderX system for 10 cents an hour, https://www.packet.net/ has some deals on their spot pricing sometimes.

With their out-of-band console, you can watch it boot from the very beginning.

Not affiliated, just sometimes use them when I need to play around with a physical box, and can't do so at work...

Interesting that the graph shows ICC getting 20% more performnance than GCC - This seems a much bigger difference than normal, as people usually don't find the ICC speedup enough to warrant using. Anyone know if it's a CPU2017 phenomenon, or has ICC recently dramatically improved, or was ICC always 20% faster?

or ICC was optimized for benchmarks

Larger ROB/OoO window than Xeon? Wow.

I would enjoy seeing a Cavium based system that I was allowed to actually run my own code on to see what was what. Over the years, in various roles, I've listened to them trying to sell me servers but they have never actually had hardware that they could just let me run code on.