Intel is running rings around AMD and Arm at the edge Supermicro launched a wave of edge appliances using Intel's newly refreshed Xeon-D processors last week. The launch itself was nothing to write home about, but a thought occurred: with all the hype surrounding the outer reaches of computing that we call the edge, you'd think there would be more competition from chipmakers in …
The article covers it later on but really this seems to be all about definition. A few years ago, "edge" was used to refer to devices running on large networks, such as CDNs, that were offloading processing such as DDoS attacks to nodes topologically closer to the client and these were of necessity small devices that, for example, could be fitted into DSLASM cabinets. As they never needed to run x86 code, x86 was not a popular architecture given the power draw. So talking about Xeons and edge really is something else. As is doing telemetry.
AND monetizing that data stream is the next big growth area.
CEO to CIO:
Why did we spend all that money on telemetry? What is in it for my paycheck?
CIO:
Don't worry boss, we are about to sign a deal with Google to sell all that lovely customer data. That deal will bring in ten times what it cost us to put it in the first place.
CEO: Great. Carry on.
CEO to wife later that day,
Darling... you can close the deal on the new villa in Aruba...
At the very edge you will have an army of CCTV, home routers, Alexa enabled devices, SmartTV’s, set-top boxes and Mobile phones all running ARM.
Edge is a meaningless cover-all term for this useless specious article… I presume originated from some SuperMicro press release ??
The article is really talking about edge services - one step in from the actual edge. It would certainly have helped if this had been reflected in the title, I too was going "WTF?!" until I read enough for that to become clear.
I get the same kind of mental failure from city dwellers who cannot understand that, out here in the sticks, there is a distinction between an outlying house and the nearby village where we go to do our shopping and visit our GP. To a Londoner, pushing something out to the edge means relocating to Watford.
AMD had basically ZERO experience with anything related to Edge, and not a lot of experience in IoT/Embedded. Most of the time, mind you, their embedded solution (prior to Zen based CPU's) was just a desktop CPU. AMD made most of it's money from the semi-custom: Sony, Microsoft etc..
Yes, the EPYC for embedded is quite old, but AMD needed something that can have some good & tested architecture - like Zen 3, something which will they need to support for at least 7-10 years, can have tons of core and tons of memory, be cheap, be reliable - enter AMD Siena (already available to some major OEM's, btw), which should compete directly with Xeon D or anything based Ice lake and probably Sapphire Rapids (if Intel will ever release those chips...)
The Ryzen Embedded is for customers who needs the basic stuff - 4 cores are more than enough, cheap, has a long official support (10 years) from AMD. The OEMs seems to like it. Yes, it's not based on Zen 3, it's based on Zen+, but in this segment, you roll your own board with an ODM, and you add your own custom/3rd party IP if you need something fast.
All in all - yes, Intel still rules the market, but I'm pretty sure that both ARM based solutions as well as AMD based solution - will take market share from Intel. We've been there before on other segments.
Intel and AMD have the x86 tool chains that have matured over 30+ years. And they’ve been able to focus on just a few variants, with some AMD and Intel specifics diverging (looking at you vector offloads). But nearly everyone has been maturing these tools, libraries, support libraries, desktop environments, etc., for literally decades.
One thing ARM has something to catch up on is still the tool chain. Even Apple has had to come to terms with that with the M1 and M2 chips. Want to run a VM? Well, sorry, your normal tools don’t work…yet. For many years, the ARM tool chains have been fragmented with all the different hobby boards, SBC’s, and server chips. Only until the recent few years has it really solidified and started converging on something solid for many ARM chips. But I think this is still a very huge hill to climb; get the tools in place, then convince your upper management signing the PO’s that the tools in place are adequate replacements for what nearly everyone in the industry is accustomed to. And: is it worth it to switch, with potential re-learnings of things like debugging binaries (different asm), bootloaders, etc.? I think yes, but that could be a significant learning curve for established departments.
I do believe ARM has everything to gain here, and is positioned quite well to take it. But there’s still quite a bit of work to be done at the edge server space, despite the success of Apple’s laptops.
It's not really the CPU cores anymore, it's the ISA/BIOS/UEFI (call it what you will).
Every ARM application processor has a different set of interfaces, with no standardised way of discovering what's where.
You have to build your own bootloader and kernel with all that info hardcoded. Sure, Das Uboot, buildroot and others make that task far easier, but it's still a big hurdle.
Raspberry Pi, Jetson and similar ARM platforms do well because all that work is already done. If ARM had a standard startup and probe then it'd really start to give amd64 a run for it's money.
Debugging tools (including remote debugging) are basically the same. Post-mortem is a bit of a pain as gdb and cdb both have annoying issues if the debugger and debuggee aren't the same CPU type.
ARM servers, lots of ARM servers that are easy to buy without having to phone up for a quote. Intel isn't ahead because they have superior products, Intel is ahead because it is extremely easy to buy their kit. If it was just as easy to spec up an ARM server and purchase one, Intel would lose a lot of it's "edge" as it were....pun not intended.
I can have an x86 server in front of me in less than 10 days...I haven't a clue how quickly I could get an ARM server because I don't know where to get one without having to deal with a sales vampire.
If I could buy ARM kit as easily as x86, I'd switch to ARM in a heartbeat. Even if it costs 50% more than x86 for 50% less performance, I'd still buy it because I could have twice as many servers because I would output about 80% less heat and could therefore have much more dense rack setups, use way less power and not have to crank the air con up so high. I would save considerable amounts of money if I went ARM.
I would save money on space, energy and cooling have increased redundancy and an insignificant performance penalty.
But alas, ARM servers are as rare as rocking horse shit.
ARM servers are rare because they have to be designed from the ground up for a specific purpose to have a considerable savings on power, so effectively at the server level ARM becomes an ASIC. Without this design intention, far too much is lost from the wall clock. The idea that ARM is simply a lowered powered x86/x64 CPU is very wrong. ARM development is so specific that you can't simply buy a "AI" ARM CPU, you have to be specific on which type of AI... sub-categories within sub-categories.
Nonsense, there are plenty of general purpose ARM based processors out there...the vast majority of servers out there aren't doing anything particuarly niche...they just hosting websites and databases etc...ARM CPU's can do this easily without any specialist configuration / tooling etc...
Most common server technologies that are designed to run on Linux have an ARM build available and they have been available for a very long time.
The Ampere Altra at full tilt uses 50W less than a competing Epyc CPU. Given that most servers don't often operate at full tilt, I'd imagine the Ampere is even more efficient than that! Doesn't sound like much, but over the scale of a rack, it adds up...50W is about £6-7 a week at least, I think, hard to say given the energy price changes of late...might be a lot more!
Over 10 servers, that's £70 a week saved, or £280 a month, £3,360 a year. Now scale that across an entire datacentre and you've got whacking great big potential savings. Even higher if you run ARM CPUs in a low power state (which is probably more than enough for most use cases). In that situation, I'd expect to save 100W per CPU. So your saving doubles...potentially even more. None of this takes into account the savings you'd make on cooling as well...your aircon won't have to work anywhere near as hard which will save even more energy.
https://www.extremetech.com/computing/322988-ampere-will-launch-128-core-arm-cpus-by-the-end-of-2021
For applications like NGINX proxies, web servers and the like, it's very hard to argue against ARM...trouble is, if you win the argument...you can't buy one because at the moment, most ARM servers seem to be geared up towards tackling the top end of the market and cost at least £8000 (which you're essentially getting for £5k with the power savings) for a spec that makes sense...there's nothing down near the bottom...if you want a 64-core or bigger beast, you're sorted...if you just want a simple, really efficient, 1U, 8-16 core box with 16GB RAM, there's nothing.
What we need is a 16 core ARM equivalent of the Dell PowerEdge R240 in the £300-£500 range (excluding storage)...they would fly off the shelves in my opinion. I would snap a load up in a heartbeat and cluster them.
How many years did it take for the Xeon D chips to get the present level of adoption? At least a decade. I would say - fairly slow.
At present, the area of greatest change seems to be centered around the Raspberry Pi. There was already significant interest, even when the boards were not really suitable. With the CM4 and extended temperature range, the main roadblocks are gone - with a now-familiar platform. Expect ARM, centered around the Raspberry Pi, to take over this market fairly quickly.
In this context, for AMD try and push x86 edge/embedded does not look like a great investment.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
