AMD lifts the veil on Opteron, ARM chip plans for 2014 AMD has unfolded its server-chip roadmap for next year, and the road ahead appears to be a sensible motorway with no hair-rasing hairpin turns or unexpected switchbacks – although there is one bright shiny new vehicle on the road. You can forgive AMD for being somewhat cautious – heck, you should congratulate them for their …
It takes the process of making a CPU down from three and a half years and $350m and $400m down to 18 months and $30m
He's not wrong on that. Kudos to AMD to understanding just how commodified the market is becoming. On the face of it the Warsaw, Berlin, Seattle models seem to have something for everyone.
I was burned more than a few times by using AMD chips over the past 25 years. I have to admit, I haven't bothered with them in a while. This is mainly because they're great at consumer and great at data center, but somewhere in-between, they're lacking when it comes to supporting small scale developers. Their development tools have always been lacking and their pathetic support for GPU code developers really chased me away from them. That said, I am happy to hear that a real CPU vendor is taking on ARM.
There are a few major problems I see here :
- 10Gbe support on die, this is just amazing, but to be honest, I don't any mention for FCoE/DCB here. To make the 10Gbe controller useful, it needs to be more than an Ethernet controller or it's just a waste of die space. A modern 10Gbe controller needs to support priority flow control and enhanced transmission selection. This is a minimum requirement for supporting FCoE. In addition, added support for RDMA over Ethernet is a major requirement in today's environment. Add to that VNtag support and you have a controller worth using. Yes, I know only a handful of vendors have those features today. Broadcom, Cisco and it looks like soon Intel will all have them. But two 10Gbe controllers with those features are a minimum requirement in a data center network adapter for 2014.
- Virtualization support. ARM has it, but it's borderline crap at the moment. This is really not ARMs fault. They're just noobs to this category of computing and it'll take some time to get it worked out. They should be actively working with VMware, Microsoft, RedHat and Citrix to make this happen. They should get silicon out there as soon as possible so those companies can get bare metal hypervisors ready for the ARM processor in the real world.
- Compiler support. ARM/AMD need to stop screwing around with the ARM compiler which has always been a pain in the ass except when developing boot loaders or code close to the hardware. They need to put together a team of real engineers to take LLVM seriously. Thankfully, the guys over at Apple take ARM seriously, but they're interested in the 32-bit core from ARM. I haven't heard a single rumor of a chip from Apple which will employ 64-bit instructions. In fact, in a smart phone, it's probably almost a disadvantage to waste space on wider word width. What's the point of adding a huge amount of CPU power using 64-bit in a phone when 99% of what you want to accelerate would profit more from custom cores and better GPUs. In server land, it's all about CPU and therefore neither ARM or AMD can bank on someone like Apple taking on the optimization of back end code generation for 64-bit ARM as there's just no profit in it for them. AMD needs to invest to make this happen.
- Wide busses. Most people simply believe that PCIe busses come for free.The fact however is that the PCIe controller of the CPU takes a tremendous amount of bandwidth, requires direct memory access within the CPU cache as well as system memory causing major issues. This increases the amount of multiplexing that has to occur within the CPU cache especially when trying to support cache coherency. This generally is accomplished by placing a much larger burden on the cache logic itself, the effect of this is to either slow down the cache and the speed at which the CPU cores can access it OR by increasing transistor count and power consumption. Want to see all those great benefits of ARM technology go bye bye? Add more PCIe lanes.
- Microsoft support. Yeh... I know.. we all hate Microsoft, but to be fair, unless Microsoft throws some server and data center love at ARM, there's little hope for this being a really useful technology outside of corner cases. Sure, you can do more web serving. You can maybe run some monster apps like hadoop, but in the end, companies run on Windows Server, Exchange and others. I doubt there's a whole lot of processor specific code in Exchange, but there's bound to be some. In addition, for developer workstations, there should be a version of Windows which runs on ARM with desktop support and Visual Studio for example. Most developers really don't enjoy debugging their apps remotely. There's something more natural about debugging on the machine you're coding on.
Guys, it could happen... I hope it does, but unless ARM and AMD take the real world problems seriously, I don't see this being more than a niche market item. And worse, if AMD comes to market with a half assed solution, it'll become a problem like Surface vs. Surface Pro. People still go to the store and buy computers with either Intel or AMD chips in them. If they get a machine with an AMD and it's an ARM and they end up running Windows RT or worse, an Android variant, people will stop looking for AMD systems because it's too confusing.
You make some good points there, but on one I disagree - Microsoft support is not needed to push ARM to servers. Yes it will help, but is not really important. If companies are given a choice to migrate existing Exchange/SQL/MSAD/terminals etc. servers to new ARM architecture, or stay on x86, I very much doubt they will pick a newcomer. And it does not matter, either.
These servers might be significant investment for some companies, but in total these are not really significant part of servers in todays' server plants (no offence to those who maintain these). It is Linux servers doing the simulations, financial computations, transactions, handling real data loads which take the largest footprint. All that AMD/ARM need to ensure entry to this market, is to work with large Linux distributions on the new platform, to work with compilers of choice (gcc, llvm) and make them really good at generating and optimizing code for the new architecture.
Some interesting points - I agree with Bronek that it's Microsoft who has to come to the ARM party and not the other way round - Warsaw sounds like a good option for Exchange servers for the immediate future. Windows Server 2012 can't be expected to be done for a new architecture anyway, so nothing before end of next year. By then MS will probably have a new CEO and may seriously be thinking about rejigging its infrastructure to suit the market. Who knows.
AMD's coming to the party brings 64-bit, hypervisor and Open-CL experience which will make a big difference if ARM gains traction. The key point is that things like OpenCompute are all about commodification and price and interoperability over absolute performance. AMD becomes and added value reseller of ARM cores, chips largely comparable to mid-range Intel but at a tenth or less of the price.
Apple isn't really interested in high performance - it wants just enough performance and endurance on consumer devices to keep the creatives loyal and dumb servers in its data centres for storing and mining their data.
"Microsoft support. Yeh... I know.. we all hate Microsoft, but to be fair, unless Microsoft throws some server and data center love at ARM, there's little hope for this being a really useful technology outside of corner cases."
As others have said, I don't think this is a large issue. I think you over-estimate MS in the server market.
Yes, there are a lot of MS servers out there. But there are many more Unix based systems. If companies can migrate their Unix/Linux boxes on to ARM hardware for less money and less power consumption at the same performance, why would they care about MS support?
As this happens, MS may see a migration of customers over to Unix/Linux systems. That is the point where they would start putting effort into an ARM Windows Server. It is always the case: MS is behind the curve. It's understandable. They only want to invest in stuff which will make them money, so they wait a bit to see how successful something is before they commit to it.
In truth, MS's commitment to ARM servers is unimportant at this point. It will become more important as the platform develops and the market grows (if it does), but MS wont let a profitable market segment go unloved for long.
Oh, you mean Microsoft has 75% of market when it comes to sales of an Windows Server operating system. Or SQL Server, but why are these two mixed?
I'm shocked, positively shocked. It is not like you can just download and install Linux operating system on your server, without paying license fee to anyone, right?
> Windows Server has about 75% market share
The measurement of this share is by value (or cost to the buyers) of sales for machines built and installed with a 'server' OS, and of server OS sold without hardware. Windows servers on the same hardware will be thousands of dollars more than, say, CentOS.
Many large server farms will build their own racks and may even have their own Linux distro, this won't show up in the 'market share' as it hasn't been bought through the 'market'.
The best figures of numbers of servers shows that each of Linux, Unix and Windows has about 1/3.
Erm, no. There are many more Microsoft based servers - Windows Server has about 75% market share - http://www.trefis.com/company?hm=MSFT.trefis#/MSFT/n-0582/0657?c=top&from=rhs
The text on that page (which you get if you have Flash switched off) only refers to the market share of Windows PC's - and puts that at around 73%. Servers don't get a mention.
But it depends on what you count. Most Windows servers around here seem to have a few cores at most. My Linux ones have 24, 32 or 40.
So if there are 3 dual-core MSWin servers, and one of my 40-core one them MS has 75% of the servers. But I know where most of the work gets done.
> There are many more Microsoft based servers - Windows Server has about 75% market share
Netcraft measures websites and Windows has about 17% share. However, this is based on domains. Very large sites behind a single domain tend to be Linux based and have dozens or hundreds of servers. For example Google or Facebook. Even Skype is reputed to run on a thousand Linux servers since MS took over. OTOH MS went on a campaign to host 'parked' domains, registered domain names which pointed to a single hosted server. They paid the hosting company to use a Windows machine for the hundreds of parked domains they hosted as this would boost the Netcraft figures.
Why go to great lengths to explain some things when we don't know if they are an issure or not? Like 10Gbe, I assume they know what they need to support their customers. And are there not many servers today that rely on MSFT for nothing? AMD actually has a decent software team that develops the tool kits, and likely provide that for ARM 64-bit computing before the end systems ever get released obviously.
That's rather an expensive facility for the top end and not exactly clean yet. According to http://researcher.ibm.com/researcher/files/us-pengwu/BGQPerfPaper-final-PACT12.pdf AMD have been looking at it just not announced anything. You can see there the various problems and opportunities at the moment and a comparison with the software equivalents. I think it is also interesting in that it may improve straight line high performance computing in the future as it is the same sort of hardware as is needed for speculative execution which allows for a longer pipeline and yet might give another 10-20% performance boost.
Although at some point you have to suspect that AMD will kick out an ARM based ceepie-geepie chip with a load of GPU compute cores on it, as that seems to be the way they're going with everything else.
That might put the cat amongst the pidgeons. You could even speculate on the possibility of having the things jigged so that the CPU cores use their GPU counterparts for FP work without the running code being any the wiser.
I really wish I could be privy to Intel's boardroom's view of AMD.
My theory is that Intel regards AMD as an asset because Intel fears becoming a monopoly. That would mean (a) government interference and regulation, and (b) nothing to stop it increasing margins by decreasing R&D ... short-term gain ... until it missed something radical from a company that was completely off its radar. At that point it would be unable to respond fast enough, because its R&D would by then be a shadow of its former glory.
So AMD gets to play with exotic CPUs, while Intel gets to say it's not a monopoly because folks can buy from AMD instead. Some do, as proof thereof. Most don't. Intel has an intrinsic advantage from its process technology and scale. It does boring mainstream stuff extremely well. It can follow wherever AMD shows there is a serious new direction to move in (like x86-64, for example). As long as it keeps its process technology advantage, its designs only have to be 2/3 as effective to benchmark as equals, and it can improve on 2/3 x 3/2 on its next design iteration. In short, AMD keeps Intel on its technological toes without making a serious dent in its profits. Intel for its part keeps AMD in financial distress, but not terminally so. It could kill AMD with a price war, but doesn't want to.
Intel could probably also put AMD out of business through the courts for patent violations, but doesn't want to. Intel probably believes that it can get away with copying anything AMD has patented, because in a nuclear war between patent lawyers, the company with the deepest pockets will win even when it loses.
Well, that's my theory. Thoughts?
Not sure what you mean about AMD and patent violations. Intel uses AMD patents and AMD uses Intel patents, they have a cross-licensing agreement, for free use.
It seems odd to me that Intel comes up with a product only when the competition provides the customers through their advances and benchmarks. Intel copies the strategy of others.
"it's a deep bin sort and improving yields that are at work. (Shhhh.)"
This is not true, it will be a new stepping, revision of the die. You dont get 20% performance per watt out of sorting. Who told you this or was this comment something you fabricated?
"AMD used Taiwan Semiconductor Manufacturing Corp to fab its Kyoto Opteron X-Series processors using its 28-nanometer processes, and the kicker Berlin Opteron X chips will also come out of the same wafer baker and use the same 28nm processes.
Neither the Warsaw nor the Berlin Opteron chips will have PCI-Express 3.0 peripheral controllers on the die, but the Berlin chip has one PCI-Express 2.0 controller on die and another one on the system controller hub."
Where did you get this information??? AMD only got permission from GlobalFoundries to build the Jaguar X86 cores at TSMC. AMD is manufacturing 28nm Steamroller at GlobalFoundries where they have an exclusivity contract on the big cores. And GlobalFoundries has a more optimized LPH 28nm process for low power and high performance.
Additionally, Berlin DOES have PCIe3, it says so right on the darn picture you posted, PCIe Gen 2/3. I realize the block diagram doesn't call out PCIe3, likely because some moron at AMD marketing made it. There are versions of that diagram the show PCIe 2/3.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
