Intel planning for thousands of job cuts, internal sources say

[imported_ats]

But you know a lot about Intel; can they develop a SoC with, say, 72 Atom cores of some description? - or would it be prone to 'dark-silicon', due to heat dissipation?

You might want to google Knights Landing. Though their use for anything outside of heavy HPC workloads is extremely limited.

ARM have announced that they are developing IP for SoCs of up to 256 cores, presumably for a case like web-serving, perhaps, or network processing. This is enabled by SoC-level design, not by the ISA so much, although that does play a role, so I understand, in the efficiency of some SoCs.

Um, good for them? Probability of it being useful for 256 cores? almost zero. And for web-serving, you don't want to pay for an ultra low volume botique processor, so probability of sales? almost zero.

Are Intel developing IP for up to 256 cores on a SoC?

Unlikely to be currently as there are no use cases given the physics based realities and restrictions.

The problems ARM most overcome in the server market dwarf any problems that Intel needs to overcome in the phone market to put things in perspective.

 

[imported_ats]

Caucasian US citizen males, of course. By far the most expensive part of their workforce.

The really good stuff at Intel hasn't even been designed in the US for a long time now. Most of the engineering in Intel's modern CPU and chipset products happens at their Haifa complex.

This is flat out false. Haifa and Oregon have been switching off on mainstream core development and basically all server CPU development is US based.

 

[imported_ats]

What features in particular?

Lots of synchronization, performance, performance monitoring, etc functionality.

 

[imported_ats]

If you asked most small to medium sized web companies if they would re-compile and switch to ARM in order to cut their AWS bill in half, 100% of companies would "yes, how about before 5pm tonight?"

Um nope! Most would want to know how their existing application ran on the hardware, would have to figure out how much NRE would be required to modify their applications so that they would work, etc. There isn't a viable ARM alternative on the market. And that's not my opinion, that's Facebook's as well. They've tested it out and found it severely lacking compared to the Intel options.

The ARM dream: ARM-based AWS servers that cost half as much (iso-performance) as Intel servers to AWS end-customers.

And there is a reason it is a dream: because it isn't reality and is unlikely to ever be a reality.

My understanding is that there are still many parts of the stack missing to offer something like that. So if it happens, it's still a few years out.

Not least of which is viable single thread performance.

 

[videogames101]

Um nope! Most would want to know how their existing application ran on the hardware,

That's why I specifically said "iso-performance" mate. :\

would have to figure out how much NRE would be required to modify their applications so that they would work, etc.

The people I've talked to indicate that if the price/perf was there the application modifications would be done overnight (metaphorically).

There isn't a viable ARM alternative on the market. And that's not my opinion, that's Facebook's as well. They've tested it out and found it severely lacking compared to the Intel options.

I completely agree.

And there is a reason it is a dream: because it isn't reality and is unlikely to ever be a reality.

Nice opinion I guess? Forever is a long time.

Not least of which is viable single thread performance.

It turns out many modern website architectures lend themselves to scale-ability, but of course increased ST performance would help make the dream into reality. I'm not saying it will happen for certain, but there is weight behind it right now.

 

[myocardia]

It turns out many modern website architectures lend themselves to scale-ability, but of course increased ST performance would help make the dream into reality. I'm not saying it will happen for certain, but there is weight behind it right now.

Wouldn't you say that there is less weight behind it today, than there was ~18 months ago, when Xeon-D appeared?

 

[videogames101]

Wouldn't you say that there is less weight behind it today, than there was ~18 months ago, when Xeon-D appeared?

I guess I'm not sure what you mean by "weight". We can all see that the tech is not ready. But specifically, I would say that between the various ARM partners there has been an increasing number of engineering hours being spent to develop the necessary tech to take on Intel in the data center. That's what I mean by "weight". Go check out Qualcomm, ARM, Avago/Broadcom, Samsung, etc. investor presentations. They are putting money/man-hours ("weight") into this. Only time will tell if they end up with any success, but those are companies with serious capital.

 

[beginner99]

You might want to google Knights Landing. Though their use for anything outside of heavy HPC workloads is extremely limited.

It's not really limited at all just not very cost efficient. Knights Landing also ships as a CPU not just co-processor and fully supports x86. It can run Windows. So if you already have a very parallel application running on x86 it would need 0 work to port,you can directly execute. Of course it will however not make use of AVX-512.

 

[ShintaiDK]

"PayPal releasing information demonstrating the benefits of using ARM based servers with a 50% reduction in capex, an 85% reduction in running costs and a 10-fold increase in server density compared to traditional data center equipment"

ARM's 2015Q2 results refers:
http://media.corporate-ir.net/media_files/priv/ccbn/event_help/smalldownload/pdf.gif

That's a cute...PR statement. ARM continues its miniscule market share in servers.

Now tell me, why isn't everything shifting to ARM but staying on x86.

Your ARM server dreams are as far away as ever.

 

[imported_ats]

That's why I specifically said "iso-performance" mate. :\

which doesn't and never will exist.

The people I've talked to indicate that if the price/perf was there the application modifications would be done overnight (metaphorically).

Well I'll say good luck with that. Seriously, MASSIVE engineering firms have looked at it and found the actual software costs too costly. The problem is that your ST performance is going to take a significant hit, which requires significant refactoring in your application stack, and beyond that there are further system bottlenecks.

It turns out many modern website architectures lend themselves to scale-ability, but of course increased ST performance would help make the dream into reality. I'm not saying it will happen for certain, but there is weight behind it right now.

The problem is that there is a floor below which ST performance is basically all that matters from a TCO perspective. Its why flock of chicken has never taken off. There is significant NRE in just keeping current app stacks running while adding in the required new features (and as we all know, there are always new features). When you add in the additional NRE in that app stack development for basically another order of magnitude in continuous optimization, even large massive installations like facebook/google can't afford it. It is basically an order of matter cheaper to do the software and system engineering to run on hardware with X ST performance than X/2 ST performance.

The latest V5 Xeons are a perfect case in point. They aren't area constrained and could of increased the core count but that would require a significant reduction is ST performance, in most cases below the baseline thresholds.

 

[imported_ats]

It's not really limited at all just not very cost efficient. Knights Landing also ships as a CPU not just co-processor and fully supports x86. It can run Windows. So if you already have a very parallel application running on x86 it would need 0 work to port,you can directly execute. Of course it will however not make use of AVX-512.

That's assuming that there are no Amdahl's law based constraints on your application. One of the major concessions that KNL makes is greatly reduced ST performance for anything that's not AVX-512. KNL is probably at 1/4-1/8th (or less even) ST performance compared to Xeon-D/E5/E7.

 

[imported_ats]

I guess I'm not sure what you mean by "weight". We can all see that the tech is not ready. But specifically, I would say that between the various ARM partners there has been an increasing number of engineering hours being spent to develop the necessary tech to take on Intel in the data center. That's what I mean by "weight". Go check out Qualcomm, ARM, Avago/Broadcom, Samsung, etc. investor presentations. They are putting money/man-hours ("weight") into this. Only time will tell if they end up with any success, but those are companies with serious capital.

Yes they are companies with serious capital and they are quite literally throwing it away to be a largely a bargaining chip with which the cloud vendors will negotiate the same or lower costs from the Intel chips they already buy.

Generally, all the cloud vendors are perfectly happy with Intel's solutions, they just want them to be cheaper, they don't really want to switch to a different vendor.

 

[Thala]

Lots of synchronization, performance, performance monitoring, etc functionality.

Hmmm, with other words, you have not the faintest idea regarding this topic.

 

[Dresdenboy]

Now tell me, why isn't everything shifting to ARM but staying on x86.

App diversity?

 

[ShintaiDK]

App diversity?

Nope.

 

[videogames101]

That's a cute...PR statement. ARM continues its miniscule market share in servers.

Now tell me, why isn't everything shifting to ARM but staying on x86.

Your ARM server dreams are as far away as ever.

It's because the platform is not ready from a technical perspective. It's not even remotely close to plug-in-play.

which doesn't and never will exist.

Never is a long long time.

Well I'll say good luck with that. Seriously, MASSIVE engineering firms have looked at it and found the actual software costs too costly. The problem is that your ST performance is going to take a significant hit, which requires significant refactoring in your application stack, and beyond that there are further system bottlenecks.

The problem is that there is a floor below which ST performance is basically all that matters from a TCO perspective. Its why flock of chicken has never taken off. There is significant NRE in just keeping current app stacks running while adding in the required new features (and as we all know, there are always new features). When you add in the additional NRE in that app stack development for basically another order of magnitude in continuous optimization, even large massive installations like facebook/google can't afford it. It is basically an order of matter cheaper to do the software and system engineering to run on hardware with X ST performance than X/2 ST performance.

The latest V5 Xeons are a perfect case in point. They aren't area constrained and could of increased the core count but that would require a significant reduction is ST performance, in most cases below the baseline thresholds.

Depends on the application, as you well know. There is a reason companies use Xeons more often than systemZ mainframes. It's not because the Xeons have better ST performance, it's because cost is (in fact) a factor. And interestingly, do you really believe not a single ARM partner is going after Intel class ST performance? I think you're really caught up in the status quo, but I suppose we'll look back on this thread in 4 years and see if you were right.

 

[jpiniero]

Nope.

You have to remember that chips have a pretty long lead time. The more I think about it, the more the severity of the job cuts seem kind of excessive. It could just be that they are going to rehire cheaper workers at a later point (however they do it) but maybe the real purpose is to make earnings not look so bad 12-18 months from now when it becomes apparent that ARM has gotten a foothold at the cloud companies.

 

[Systems analyst]

It seems that Intel are responding to the threat to their business by developing products such as Knights Landing and Xeon-D; they know that they are vulnerable, and may have difficulty in maintaining their 60% profit-margin. They have purchased Altera, someone else's technology, which indicates the recognisation of the need for change.
Currently, there are a dozen or more companies developing ARM SoCs for networking or servers; the use of proprietary accelerators allows a rich field of SoCs to be developed.
ARM are recruiting engineers to design server-specific IP; they say that it can take up to seven years to go from the start of designing a processor core, to its availability in full production, but other ARM based solutions seem to be well advanced, with server SoCs being evaluated at present.
We should hear more next year, if they are successful in beginning to penetrate the server market; they already have 15% of networking.

 

[Systems analyst]

It seems that Intel are responding to the threat to their business by developing products such as Knights Landing and Xeon-D; they know that they are vulnerable, and may have difficulty in maintaining their 60% profit-margin. They have purchased Altera, someone else's technology, which indicates the recognisation of the need for change.
Currently, there are a dozen or more companies developing ARM SoCs for networking or servers; the use of proprietary accelerators allows a rich field of SoCs to be developed.
ARM are recruiting engineers to design server-specific IP; they say that it can take up to seven years to go from the start of designing a processor core, to its availability in full production, but other ARM based solutions seem to be well advanced, with server SoCs being evaluated at present.
We should hear more next year, if they are successful in beginning to penetrate the server market; they already have 15% of networking.

 

[imported_ats]

Depends on the application, as you well know. There is a reason companies use Xeons more often than systemZ mainframes. It's not because the Xeons have better ST performance, it's because cost is (in fact) a factor. And interestingly, do you really believe not a single ARM partner is going after Intel class ST performance? I think you're really caught up in the status quo, but I suppose we'll look back on this thread in 4 years and see if you were right.

They use Xeons because they are cheap, support all the RAS features required, are well supported from a tools and infrastructure perspective, have the best or nearly the best performance on a perf/w, ST, and MT perspective, etc. They aren't going system/Z because system Z costs a bloody fortune, has much worse perf/w, and worse performance.

And no, I don't believe that a single ARM partner is going after Intel class ST performance, largely because they can't afford it. There are only a handful of companies with the resources to achieve it, most aren't interested in ARM, the one that is won't share its toys. None of the ARM vendors have a market to offset the development costs for that level of ST performance. Intel offsets the costs via the PC market for instance. Apple does it due to their insanely high profits compared to the rest of the market in mobile. Everyone else is left battling over raw price in the mobile market with razor thin margins. The cost to just develop a server class uncore alone is in the 100s of millions range and rising. And I'm saying this as someone who has designed and built uncores across multiple architectures across multiple companies for the server space. NRE for EV7 was hundreds of millions of dollars and is rather simple compared to what's required these days for instance.

 

[videogames101]

They use Xeons because they are cheap, support all the RAS features required, are well supported from a tools and infrastructure perspective, have the best or nearly the best performance on a perf/w, ST, and MT perspective, etc. They aren't going system/Z because system Z costs a bloody fortune, has much worse perf/w, and worse performance.

And no, I don't believe that a single ARM partner is going after Intel class ST performance, largely because they can't afford it. There are only a handful of companies with the resources to achieve it, most aren't interested in ARM, the one that is won't share its toys. None of the ARM vendors have a market to offset the development costs for that level of ST performance. Intel offsets the costs via the PC market for instance. Apple does it due to their insanely high profits compared to the rest of the market in mobile. Everyone else is left battling over raw price in the mobile market with razor thin margins. The cost to just develop a server class uncore alone is in the 100s of millions range and rising. And I'm saying this as someone who has designed and built uncores across multiple architectures across multiple companies for the server space. NRE for EV7 was hundreds of millions of dollars and is rather simple compared to what's required these days for instance.

I wonder if in 5 years someone will say the same thing:

They aren't going Xeons because Xeons cost a bloody fortune, have much worse perf/w, and worse performance.

 

[imported_ats]

I wonder if in 5 years someone will say the same thing:

They aren't going Xeons because Xeons cost a bloody fortune, have much worse perf/w, and worse performance.

It highly unlikely. System/Z is from an entirely different era than microprocessors like x86/ARM. Its goal these days is primarily to allow legacy software stacks to continue to run and provide IBM with service revenue. To do that they expend a large amount of money and resources to increase the performance to allow old applications to continue to scale in performance. A side effect of that is the massive MCMs used along with the liquid cooling which allow them to run at much higher power lvls. AKA the fundamental design/targets of System/Z is at odds with what the majority of server customers want these days. The same cannot be said of ARM/x86 which besides the performance issues largely have the same fundamental design targets.

 

[Nothingness]

System/Z is from an entirely different era than microprocessors like x86/ARM. Its goal these days is primarily to allow legacy software stacks to continue to run and provide IBM with service revenue.

And it looks like z Systems isn't doing that well: 16'Q1 revenues fell 42% YoY (vs 14% decline for Power). IBM blames it on the "product cycle dynamics".

I know of a large company that uses IBM mainframes. They are trying to move away from them, but it takes time to port all of their PL/I software to something more modern :biggrin: