TSMC and ARM Tape-Out First ARM Cortex-A57 Processor on 16 nm FinFET Technology

[beginner99]

By top end I meant the server/HPC market that makes up majority of Intel's revenues & profits, IIRC one of the recent AT articles showed that calxeda(ARM) delivers greatly on the power/efficiency front & seeing as how sensitive the enterprise is in that regard I don't why the likes of AMD, Samsung, Qualcomm even Nvidia can't make a dent to Intel's dominance provided they're in it for the long haul !

I commented in that review which was extremely biased towards the ARM server and completely ignored flexibility of the product and administrative effort required.

Yes, Servers (and power) are expensive but what is much more expensive are employees. Running virtual machines on a intel server is much, much more flexible than running real physical server on such an ARM box. And in certain cases you might fare best with "cloud services" like from Amazon.

With the intel server you can run 1 heavy weight server or just deploy a ton of VM's (web-servers). And since it's virtual it does not matter if one web-server needs more CPU power than another. With the ARM one each one needs to have equal load and if one has a too high load, the other 100 idling cores are still useless. So no, I don't see it.

 

[R0H1T]

I commented in that review which was extremely biased towards the ARM server and completely ignored flexibility of the product and administrative effort required.

Yes, Servers (and power) are expensive but what is much more expensive are employees. Running virtual machines on a intel server is much, much more flexible than running real physical server on such an ARM box. And in certain cases you might fare best with "cloud services" like from Amazon.

With the intel server you can run 1 heavy weight server or just deploy a ton of VM's (web-servers). And since it's virtual it does not matter if one web-server needs more CPU power than another. With the ARM one each one needs to have equal load and if one has a too high load, the other 100 idling cores are still useless. So no, I don't see it.

What you're ignoring is that running a virtual machine itself is highly inefficient be it on Intel or an ARM server ! So even though the review didn't reflect much of the real world usage of most servers it however did show that under such a case ARM beats Intel, given a bit of time for the platform to mature I see ARM maintaining their energy efficiency whilst improving on their performance numbers.

 

[Tuna-Fish]

So a no-name chinese design house managed to upstage AMD in getting 28nm functional silicon out of GloFo? Really?

The Cortex A9 macro they are using is much more lenient for process restrictions than anything AMD wants to ship. Basically, their chip is easier to make. Also slower.

ARM servers is nothing new, they have been there for a long time. And gotten nowhere.

They have had quite a bit of restrictions that have been deemed too onerous. (Notably, the memory pipeline on most current arm chips is just too horrible for server use.) ARM and the various server vendors have worked to remove those, one at a time. Their chips are getting better. I expect the 64-bit arm servers to sell quite well to certain segments. However...

While ARM just taped out the A57 core, Intel has been sampling its "Avoton" micro-server SoC since at least October, if not earlier.

They won't capture all that much of the dollar value of servers, and Avoton won't do much to hurt them, because it's all about cost.

The segment they are aiming for is the lowest of the low cost servers. Things running php or Java or Python to serve web pages. The workload parallelizes really well, because the only point of communication is the database, which will live on a fast box. This means that single-thread performance won't matter that much. The workload is interpreted/jit compiled, so instruction set won't matter. And the workload is horrible from a performance standpoint, full of branches and stalling loads, meaning a top-end core has much less performance to extract compared to a weaker one.

There are a lot of boxes running this kind of code, and the only thing that matters to them is throughput per dollar, counting both power use and cost of equipment. All the presumptive ARM server vendors are talking about the power advantages of ARM. This is smoke and mirrors. ARM has some inherent architectural advantage, but Intel's process advantage buries this. ARM presently has some platform integration advantage, but Avoton will likely match this. In the end, ARM will be barely able to match, or more likely fall a little behind in the throughput/watt race. Leaving the one huge advantage the ARM servers will have, but which is one the vendors don't want to talk about. Once there are several competent ARM microserver chips out there, they will compete each other to near zero margins, and Intel won't want to match the system prices because that would mean losing more profit on some other segments where x86 can fetch a premium.

I predict that Intel will simply price Avoton out of the market, conceding it to the ARM vendors. ARM won't so much capture this market as it will destroy it. It's a significant part of the server market, but even today it's not nearly as profitable as the rest of it, and after the dust settles it will be effectively commoditized.

 

[Tuna-Fish]

I think Nvidia will try to sell ARM/Tesla combo's in the near future.

Certainly. However, the HPC market they will try to sell those into is *completely* different from the microserver market the arm chips we are talking about will go into. Basically everything about the markets and their workloads are the exact opposites.

 

[Torn Mind]

They have had quite a bit of restrictions that have been deemed too onerous. (Notably, the memory pipeline on most current arm chips is just too horrible for server use.) ARM and the various server vendors have worked to remove those, one at a time. Their chips are getting better. I expect the 64-bit arm servers to sell quite well to certain segments. However...



They won't capture all that much of the dollar value of servers, and Avoton won't do much to hurt them, because it's all about cost.

The segment they are aiming for is the lowest of the low cost servers. Things running php or Java or Python to serve web pages. The workload parallelizes really well, because the only point of communication is the database, which will live on a fast box. This means that single-thread performance won't matter that much. The workload is interpreted/jit compiled, so instruction set won't matter. And the workload is horrible from a performance standpoint, full of branches and stalling loads, meaning a top-end core has much less performance to extract compared to a weaker one.

There are a lot of boxes running this kind of code, and the only thing that matters to them is throughput per dollar, counting both power use and cost of equipment. All the presumptive ARM server vendors are talking about the power advantages of ARM. This is smoke and mirrors. ARM has some inherent architectural advantage, but Intel's process advantage buries this. ARM presently has some platform integration advantage, but Avoton will likely match this. In the end, ARM will be barely able to match, or more likely fall a little behind in the throughput/watt race. Leaving the one huge advantage the ARM servers will have, but which is one the vendors don't want to talk about. Once there are several competent ARM microserver chips out there, they will compete each other to near zero margins, and Intel won't want to match the system prices because that would mean losing more profit on some other segments where x86 can fetch a premium.

I predict that Intel will simply price Avoton out of the market, conceding it to the ARM vendors. ARM won't so much capture this market as it will destroy it. It's a significant part of the server market, but even today it's not nearly as profitable as the rest of it, and after the dust settles it will be effectively commoditized.

Fascinating stuff.

 

[beginner99]

What you're ignoring is that running a virtual machine itself is highly inefficient be it on Intel or an ARM server ! So even though the review didn't reflect much of the real world usage of most servers it however did show that under such a case ARM beats Intel, given a bit of time for the platform to mature I see ARM maintaining their energy efficiency whilst improving on their performance numbers.

But it's much more efficient in terms of administration and employees cost more than servers or power.

 

[R0H1T]

But it's much more efficient in terms of administration and employees cost more than servers or power.

Yes assuming its run on an Intel machine, who's to say that the same VM wouldn't deliver better throughput(as in power/perf) on ARM servers ? This all is hypothetical of course but there is no reason to believe that ARM can't deliver this level of performance based on their recent past(last ~5yrs or so) while maintaining similar levels of efficiency !

 

[Gideon]

ARM servers is nothing new, they have been there for a long time. And gotten nowhere.

And what does that prove ? How long did x86 require to really take off in the server market ?I'm not saying, ARM servers WILL succeed but that argument is just stupid IHMO.

 

[Torn Mind]

Yes assuming its run on an Intel machine, who's to say that the same VM wouldn't deliver better throughput(as in power/perf) on ARM servers ? This all is hypothetical of course but there is no reason to believe that ARM can't deliver this level of performance based on their recent past(last ~5yrs or so) while maintaining similar levels of efficiency !

Throughput is measured in bps, not performance/watt, which is a ratio.

 

[wsw1982]

Again, if this is any indication as to where ARM is headed, I see no reason why Intel shouldn't be worried just like their unsuccessful scramble in the mobile space thus far has shown us, the only difference being that they're the incumbents in this case !

The ARM server cost more than twice the price of Xeon server (with similar performance in ARM's best working scenario). And consume compatible amount of power. How well will it go?

http://www.anandtech.com/show/6757/calxedas-arm-server-tested/13

 

[wsw1982]

That's what they said about their mobile offerings as well & yet here we are at a stage where the cortex A15 & something like a Snapdragon 800 is the best in ultra/low power segment even though Intel is half a node ahead of'em, granted they don't have atom at 22nm but still !

Just to add ~ 64bit extensions will give them something what they didn't have earlier i.e. the ability to add massive amounts of memory & who knows 64bit ARM might deliver greater benefits as compared to the jump from x86 to x86-64 !

I don't understand you logic: Because Intel has 22nm trigate processer, so the 32nm planer processer ATOM become half a node ahead of 28nm ARM?

And by way, the 32nm clover trail+ is compatible with 28nm Octa & Snapdragon 600 even with worse manufacture tech.

 

[ShintaiDK]

But the thing is performance doesn't scale up with increasing power so the upside is limited, that's why we see the tapering off in their performance graph from SNB to IVB & now Haswell, however with energy efficiency in mind you can incrementally up the performance till it reaches a certain level. Now with severs you can add lots of cores to increase the performance & SoC's deliver even greater efficiency so there certainly is alot more headroom than what most would like to believe !

You seem to forget the entire cost of the rest of the server and ecossytems. And after we got virtualization. ARM servers are essentially dead. ARM and Atom servers only fit a very tiny niche segment.

 

[Ajay]

You seem to forget the entire cost of the rest of the server and ecossytems. And after we got virtualization. ARM servers are essentially dead. ARM and Atom servers only fit a very tiny niche segment.

I'm really beginning to believe that Intel _has_ to introduce it's many cored high end server CPUs ~ 1 year after the desktop variants (if possible given validation requirements) just to take the momentum away from the ARM server lobby (Intel is losing the PR battle, ATM).

Perception being reality and all that (and IT people really are susceptible to this).

 

[wsw1982]

They won't capture all that much of the dollar value of servers, and Avoton won't do much to hurt them, because it's all about cost.

The ARM server is not cheap.

 

[R0H1T]

Throughput is measured in bps, not performance/watt, which is a ratio.

You're getting into semantics here, it should be clear enough what I was talking about, though power efficiency would be more apt right ?

The ARM server cost more than twice the price of Xeon server (with similar performance in ARM's best working scenario). And consume compatible amount of power. How well will it go?

http://www.anandtech.com/show/6757/calxedas-arm-server-tested/13

As someone said earlier this is a niche market so its not wrong to assume that with multiple ARM vendors coming in & certain other costs going down it'd be naive to think that ARM servers would remain prohibitively expensive as they are now !

I don't understand you logic: Because Intel has 22nm trigate processer, so the 32nm planer processer ATOM become half a node ahead of 28nm ARM?

And by way, the 32nm clover trail+ is compatible with 28nm Octa & Snapdragon 600 even with worse manufacture tech.

I'd like to see when/where is the clover trail comparable to Exynos Octa or Snapdragon 600 ?

 

[zsdersw]

I see ARM has become the new AMD in the stupid fanboy wars... as the OP and his/her subsequent posts demonstrates.

 

[Tuna-Fish]

The ARM server is not cheap.

That's because it's "the". It's a pathfinder product alone in the market. No-one will pay what they ask for it for anything other than evaluation purposes. Also, the A9s in it are hardly worth any money, the baseline of what is useful is probably A15.

The big deal about arm servers is that after this stuff matures, anyone and their uncle can make one. And then they will compete each other to zero margins.

 

[Idontcare]

I see ARM has become the new AMD in the stupid fanboy wars... as the OP and his/her subsequent posts demonstrates.

Its mostly about being anti-Intel than it is about being pro-anything (be it pro-AMD or pro-ARM, etc).

Intel is the enemy, which makes any enemy of Intel an instant friend of theirs. ARM is just the flavor of the month.

 

[Arachnotronic]

Its mostly about being anti-Intel than it is about being pro-anything (be it pro-AMD or pro-ARM, etc).

Intel is the enemy, which makes any enemy of Intel an instant friend of theirs. ARM is just the flavor of the month.

Of course. Same thing with Nvidia, Microsoft, Apple, etc.If you're the market leader, then you've got people rooting for the opposition. It's fun for some, I suppose.

 

[R0H1T]

I see ARM has become the new AMD in the stupid fanboy wars... as the OP and his/her subsequent posts demonstrates.

What others don't seem to realize is that while Intel is barely able to give a ~10% boost YoY across node shrinks & major architectural improvements, AMD on the other hand is giving 10~20% gains all on the same node not to mention ARM is upwards of 40~50% in this case !

Granted AMD/ARM have started from a lower base but they're not moving up the performance ladder any slower than what Intel is down their energy efficiency path, infact it wouldn't be wrong to say that ARM/AMD will beat Intel more & more going forward across a myriad of fields they compete in !

Its mostly about being anti-Intel than it is about being pro-anything (be it pro-AMD or pro-ARM, etc).

Intel is the enemy, which makes any enemy of Intel an instant friend of theirs. ARM is just the flavor of the month.

Nicely put, although a bit sarcastic, but ask yourself this ~

Is Intel delivering the same gains that we saw during the early core 2 era ?

Are AMD/ARM outperforming Intel in the gains they've brought forth over the last couple of years ?

If this isn't the beginning of a new trend then I don't know what is !

 

[ShintaiDK]

Its mostly about being anti-Intel than it is about being pro-anything (be it pro-AMD or pro-ARM, etc).

Intel is the enemy, which makes any enemy of Intel an instant friend of theirs. ARM is just the flavor of the month.

Spot on.

 

[zsdersw]

What others don't seem to realize is that while Intel is barely able to give a ~10% boost YoY across node shrinks & major architectural improvements, AMD on the other hand is boosting it by 10~20% all on same node not to mention ARM is upwards of 40~50% in this area !

Granted AMD/ARM have started from a lower base but they're not moving up the performance ladder any slower than what Intel is down their energy efficiency path, infact it wouldn't be wrong to say that ARM/AMD will beat Intel more & more going forward across a myriad of fields they compete in !Nicely put, although a bit sarcastic, but ask yourself this ~

Is Intel delivering the same gains that we saw during the early core 2 era ?

Are AMD/ARM outperforming Intel in the gains they've brought forth over the last couple of years ?

If this isn't the beginning of a new trend then I don't know what is !

It's always easier to look like you're making a lot of progress when you're far behind than when you're in the lead.

 

[Ajay]

Granted AMD/ARM have started from a lower base but they're not moving up the performance ladder any slower than what Intel is down their energy efficiency path, infact it wouldn't be wrong to say that ARM/AMD will beat Intel more & more going forward across a myriad of fields they compete in !Nicely put, although a bit sarcastic, but ask yourself this ~

Is Intel delivering the same gains that we saw during the early core 2 era ?

On the server side, yes - faster clocks, lower power and more cores.

ARM servers are largely a niche product, IMHO, that are being over hyped because they happen to work well for Google, Facebook an other prominent high traffic interactive web content providers because the workload breaks up into nice chunks that are easily distributed (especially with new platforms like Hadoop, etc.).

 

[Exophase]

Are AMD/ARM outperforming Intel in the gains they've brought forth over the last couple of years ?

If this isn't the beginning of a new trend then I don't know what is!

That doesn't sound like a good argument to me. Intel's improvements are slowing down because there are diminishing returns to improving performance along certain metrics. Physics dictates that. Nothing sustains exponential progress forever.

If rate of progress is naturally dampened then you can progress faster when you're further behind. That doesn't mean you'll eventually catch up with and surpass the more advanced competitor instead of hitting the same slow down.

But ARM's placement vs Intel is nothing like AMD's, that invites a totally different comparison..

There are people who are blindly anti-Intel here, to be sure.. there are also those who are confident Intel will dominate any market they enter because they're Intel. No point singling out side over the other.

 

[Phynaz]

AMD on the other hand is giving 10~20% gains all on the same node

Are you including when they moved backwards in performance? Or is that conveniently left out of your calculation?