Amazon uses custom product numbered Xeons. So I wouldn't be surprised if they get the same advantage + their own optimizations in short term. Intel wouldn't risk to annoy AWS, they are a far bigger customer than Google.
The optimization Amazon wanted (and got) was related to CPU cores, not early next-gen chips. I always assumed it was a way to sell more "dedicated cores" per socket. It might be exclusive, but Google, Facebook and enterprises would not be interested unless the overall chip was more powerful.
Well good. Once upon a time we had PowerPC and x86; and one particularly smart company was able to switch architecture without too much trouble to their customers.
Sure would be nice to have some competition to get things moving forward again.
Intel has been under intense pricing pressure as the server market has consolidated into a few large cloud providers. Some of the largest providers have used their weight to really turn the screws on Intel. Providing exclusive access to top-end child has effectively turned Intel from being viewed as a commodity vendor to a partner.
In the past some of these players have threatened to turn to ARM processors to extract more pricing concessions from Intel. It will be interesting to see if that was a bluff or if Intel's technology advantage does really give them a TCO advantage even with their margins factored in.
It's also interesting that Google and Facebook are leaders in machine learning and I wonder if there isn't some additional collaboration in the works for custom silicon in the works. Google has their own custom silicon for machine learning and maybe could be encouraged to help Intel enter the market if they knew they would have exclusive access to the newest generation of chips.
Given the market caps of Google, Facebook, Apple et-al it makes me wonder why nobody has bought out Intel.
Given the market caps of Google, Facebook, Apple et-al it makes me wonder why nobody has bought out Intel.
They'd over time lose the ecosystem advantages that an independent Intel offers. They'd probably also lose a lot of money since an in-house Intel would become much less competitive. Apple also doesn't need or want what Intel has to offer (highest performance but very much not the lowest cost to manufacture chips), unless my impression of Mac vs. iOS systems sales is way off.
It's hard to imagine Intel being in such dire straights they can let Google and Facebook blackmail them in such a market damaging way, although as noted by t0mas88 other big customers like AWS might be getting their own custom numbered parts. Or, who knows, might not need exactly these types of chips (I have no idea, not familiar with the market beyond what I buy, low/near lowest cost single chip Xeons for my personal workstation class computers).
But pissing off everyone else who buys these class chips, unless contrary to the article they do have a capacity issue, seems breathtakingly unwise, an inflection point indeed.
That may actually be a reason for preferential release to Google.
In general, I think we are in interesting times. Still a guess but I think the crowd close to their forties (I'll be 40 in 4 years) will see a significant shift within a decade.
My kids are an age where they don't really care what is in their game console, TV box or their parents' phones.
I guess I (as a working adult) would really stop caring about what is powering the cloud as long as I keep getting access to the service level I (or my employer) is paying for.
With Google trying non-Intel option, there may be a copy-cat effect for other cloud providers to do the same to remain competitive, possibly diminishing Intel's importance in the server market.
And with "kids" mostly using non-Intel(or x86 in general) platforms (phones, TVs, snapchat-shades, phones etc) where Intel already has very little presence.
Not trying to sound a doomsday prophecy for Intel but... times they are a changing ?
2011-x chips contain custom and undisclosed additions for these biggest customers ("Big seven") already. These are in all 2011-x SKUs made from the same chip, but are said to be disabled through fusing, just like other feature fuses.
I hope someone leaks more info about that, or reverse-engineers something, because I find the whole secretive attitude around processors and other hardware information incredibly offputting as it is; but it's not just Intel who are doing it.
Maybe when used hardware with these customisations starts showing up on eBay or similar...
I think that Google's TPUs (Tensor Processing Unit) use Altera FPGAs (speculation based on [1]). I'm sure there's some degree of collaboration if Google is filling cabinets with high-end FPGAs.
Google is active with ARM and (Open)POWER. ARM remains a threat to the lower end, POWER challenges fat Xeons. Potentially AMD, depending on how their current development efforts are going.
Buying Intel is easier, so Intel gets to guess how far they can go before one of their large customers decides that the risk of going with one of those alternatives is worth it (demonstrating to the world that Intel isn't the only viable option in the process, which might at the end be the bigger damage for Intel)
Given the market caps of Google, Facebook, Apple et-al it makes me wonder why nobody has bought out Intel.
Or for 1/10 the cost, they could buy AMD. Not because their products are equivalent, but because any new owner of AMD also acquires a very special relationship Intel. Intel and AMD have cross-licensed a lot of IP, without which neither would be able to sell an x64 processor. And (as I understand it) if AMD changes ownership, their licensing deal with Intel is nullified and would need to be renegotiated:
However, according to Drew Prairie, director of corporate communications at AMD, once ownership of either AMD or Intel changes, the whole agreement is terminated for both parties. As a result, after a transaction happens, the companies need to negotiate a new cross-licensing agreement.
“Actually, the agreement is pretty clear – if there is a change of control for either company the agreement is terminated,” said Mr. Prairie. “That does not mean a new agreement could not be reached, but in a change of control the agreement is terminated.”
The assumption is that obviously the deal would be renegotiated, since it would be "Mutually Assured Destruction" for both AMD and Intel if either side refused. But what if AMD is bought by a company with deep enough pockets to just absorb the loss? What if AMD was owned by someone who didn't actually need it, and could countenance the MAD?
I'd think that such a company would have tremendous leverage against Intel in such a negotiation. If they demanded it, and were willing to "throw the steering wheel out of the window and play chicken"[1] (ie, structure the acquisition such that both Intel and AMD were prevented from selling x64 processors unless Intel agrees to their terms), exclusive early access (and whatever other special preference they want) would be easily negotiated.
But if this were true, I'd have thought that someone would have already done it. If not Amazon/Google/Facebook (the three largest Internet companies), why not Tencent/Alibaba/Baidu (the next largest)[2]? All of them could easily afford it, and I'd think that all of them could get tremendous benefit from a sweetheart deal with Intel. What am I missing?
Oops, yes. I looked at the linked page, mentally abbreviated the top 3 as "AGF", and then expanded it wrong while typing. It might be way for Apple to maintain their hardware premium in the x64 market, though. Or to wipe it out at the same time they move to something else.
I don't get the impression google nor facebook really want to be the sole CPU producer. Strategically, market regulators could site that as a reason to never again let google or facebook purchase another company. If you're google and facebook you highly value the ability to buy competitors out while they're small. That threat should also exist for Apple, theoretically. Of the three, I could imagine Apple having the business model closest to making a purchase/merger of/with Intel. But it sounds like there are both technological and business reasons for Apple not wanting to become that big.
So basically if I want Skylake Xeons I need to move from AWS to Google Cloud? Interesting that Intel did this. I guess throwing AWS under the bus is worth it?
The whole of the article rests on a single claim that,
"Several trusted sources say that later today, likely at Supercomputing 16, the company will announce they have pulled in Purley aka Skylake-EP Xeons, to this year and will sell them to two key customers."
Customers which the article goes on to name as Google and Facebook. Although, I cannot find any corroborating evidence to support this claim. Supercomuting 16 is now over, and I see no other articles detailing how "Intel is playing favorites with Skylake Xeons in a way that will shatter the industry status quo." SemiAccurate is literally the only site I can find professing this scenario.
I'm not saying it's necessarily untrue, but I'd like a lot more information on why it's just Google and Facebook.
- Was there a silent bidding war on early access?
- Did Intel simply choose two major customers based on previous sales?
- Were Google and Facebook specifically targeted to exploit FPGA for AI development?
> Customers which the article goes on to name as Google and Facebook. Although, I cannot find any corroborating evidence to support this claim. Supercomuting 16 is now over, and I see no other articles detailing how "Intel is playing favorites with Skylake Xeons in a way that will shatter the industry status quo." SemiAccurate is literally the only site I can find professing this scenario.
SemiAccurate has a pretty good track record with these kinds of rumors and analysis. They're not right 100% of the time, but they're right enough that it's clear they aren't just making things up.
Looking at the article date (Nov 17) and the dates for Supercomputing '16 (ended Nov 18), if Intel had actually announced preferential release to specific customers, it would have been confirmed already. Combing through Google News, there doesn't appear to be any article covering Intel's announcements at SC16 that mentions such a preferential release. So, it's probably just a false rumor.
Would the screwed companies that had been promised these chips immediately go to the public?
And with Thanksgiving, I'm not sure there's been enough time for it to further leak out.
On the other hand, it's a claim that Intel will do something that's both extraordinary and extraordinarily stupid for the long term, so we indeed need some more proof.
Although I would be really surprised if this was something they actually announced, it doesn't really help Intel to announce something like that and as the article points out it can anger other customers.
That said I have worked in Intel's System Validation group which was responsible for determining when a new chip was ready for general availability. I have directly witnessed Intel's preferential customer treatment where Intel allocated all of their processor production on a new processor to their largest customer(s). In my case they were selling Nehalem chips to Dell long (6 - 9 months) before others like NetApp could get them to start building filers with them. I have also worked in the platforms group at Google so I have an idea of the scale of their hardware effort and how many chips such a deal with Google might entail.
As a result, I think the actual story is this -
Intel got back first silicon that was acceptably functional on the new parts. This kicked off a discussion between their top customers and the potential early availability of parts. Those customers did a calculation on what advantage they might get by getting hardware early. Then Intel went down their customer list from most strategic to least strategic and offered them access. Either Google or Facebook was on the top of that list (likely Google) and they bought some, but not all of the production capacity, and then Intel offered the remainder to the second customer on the list who agreed to take the rest. Everyone else has to wait until the part is released to general production in all of its target fabs.
Further, it won't cause Intel any grief because all their customers know that the ones with the most volume get the most "love" from Intel. And the situation might be different the next time around.
I don't understand what the big deal is. Almost every integrated circuit manufacturer does this because supporting a few big customers is far cheaper than many smaller ones especially when you're operating on the cutting edge, a chronic condition for Intel. Google and Facebook will only call support when there is a discernable pattern of failures of processors or in previously verified code and that's the most optimal way for Intel to deal with early problems instead of handling thousands of support calls for intermittent failures in tiny sample sizes. The difference here is that microprocessors are a key piece of IT infrastructure and there are now a few providers that can soak up production of new products for months. For better or worse, economies of scale have always been a natural competitive advantage and Intel's dominant position in the server CPU market doesn't change that calculus, nor obligate them to make their products widely available when they first roll off the fab.
Ah, I wasn't aware that Intel promised them early access. Did Intel break a contractual obligation or an implied agreement with previous partners? I'm unclear as to what promise means in this case.
I also wonder if this has anything to do with how Google and Facebook both make their own-ish servers ( https://code.facebook.com/posts/1711485769063510/facebook-s-... // https://www.google.com/amp/s/www.wired.com/2016/02/googles-h... ) which would potentially allow Intel more wiggle room with early release chips vs supplying all the chipsets and reference designs themselves. It could also just be the volume of FaceGoog. It's a pretty normal thing for a business to do. I don't really see the cause for outrage here. I do wish that Intel had more competition, and hopefully I'll be building a new gaming PC based around the AMD Zen SR7 when that comes out.
"New generation server parts significantly reduce operating costs and without a huge price rise, they will also drop TCO more than a little"
Looking at some outside evaluations of their capex, it does not seem to represent a significant drop since CPU cost is only a tiny part of operating a datacenter.
Hence _T_otal _C_ost of _O_wnership. Running a data center is all about getting direct power consumption (servers) and indirect power consumption (cooling) down.
I've heard the assumption that energy costs outweigh cpu costs for TCO, but after some rough guesstimation I'm not so sure.
1MWh of electricity is enough to power a cpu running 114W on average for a year. A 24 core (48 thread) xeon has a 165w tdp, meaning that even such a beast is likely to use less on average (most applications don't have 100% cpu load 100% of the time, and virtualizing up to that load causes serious issues when several clients happen to simulatneously peak). In essence, this should run pretty much any single CPU, and even if you're doing something weird to actually get that 100% load, we're not far off.
Of course, there are power supply losses and cooling overheads, but even if that raises the power consumption by a factor 5 (doubtful), then a several-thousand dollar server cpu is still going to outweigh energy costs for many years.
Furthermore, energy costs are trending downwards and cpu's are still getting more efficient; it's a somewhat risky investment thus to buy a more expensive (but more efficient) chip if that investment needs years to pay off.
So, it looks to me as though plain old cpu costs certainly are far from irrelevant; they may indeed exceed energy costs over the chips lifetime if the energy is acquired at close to wholesale prices and cooling is efficient (say 3x overhead?).
They employ way too many sales and marketing people that would fight against such a straight-forward solution. You would think that the lack of having any actual competitor after AMD more or less folded would make their positions redundant, but Intel moves slowly.
They sell it. About a year ago fb decommissioned an old server design, leading to a huge glut of used Xeon E5-2670s flooding the market. Gamers snapped them up for next to nothing.
I don't really understand the commotion. Intel have been striking global and local deals like this with much smaller outfits than Google and Facebook for at least a decade.
It's an intricate play with variables like marketing opportunities (globally and locally), silicon availability (when, in which numbers), performance per dollar, etc etc. You could even get into this game by buying "just" like 5-10k CPUs in a quarter, if you're in the right region and have the proper local branding quality.
The concept of "launch customers" globally/per market doesn't seem very new either. That window of exclusivity is usually so short that it matters more from a marketing point-of-view (for both seller and buyer) than anything else.
The problem is the claim that Intel reneged on commitments to sell these chips to other companies RSN instead of in 6 months, and it's not a capacity issue.
48 comments
[ 2.8 ms ] story [ 104 ms ] threadDo you have a citation or something showing such numbers? What is this statement based on?
Sure would be nice to have some competition to get things moving forward again.
In the past some of these players have threatened to turn to ARM processors to extract more pricing concessions from Intel. It will be interesting to see if that was a bluff or if Intel's technology advantage does really give them a TCO advantage even with their margins factored in.
It's also interesting that Google and Facebook are leaders in machine learning and I wonder if there isn't some additional collaboration in the works for custom silicon in the works. Google has their own custom silicon for machine learning and maybe could be encouraged to help Intel enter the market if they knew they would have exclusive access to the newest generation of chips.
Given the market caps of Google, Facebook, Apple et-al it makes me wonder why nobody has bought out Intel.
They'd over time lose the ecosystem advantages that an independent Intel offers. They'd probably also lose a lot of money since an in-house Intel would become much less competitive. Apple also doesn't need or want what Intel has to offer (highest performance but very much not the lowest cost to manufacture chips), unless my impression of Mac vs. iOS systems sales is way off.
It's hard to imagine Intel being in such dire straights they can let Google and Facebook blackmail them in such a market damaging way, although as noted by t0mas88 other big customers like AWS might be getting their own custom numbered parts. Or, who knows, might not need exactly these types of chips (I have no idea, not familiar with the market beyond what I buy, low/near lowest cost single chip Xeons for my personal workstation class computers).
But pissing off everyone else who buys these class chips, unless contrary to the article they do have a capacity issue, seems breathtakingly unwise, an inflection point indeed.
They are looking at power PC as an alternative to Xeon.
In general, I think we are in interesting times. Still a guess but I think the crowd close to their forties (I'll be 40 in 4 years) will see a significant shift within a decade.
My kids are an age where they don't really care what is in their game console, TV box or their parents' phones.
I guess I (as a working adult) would really stop caring about what is powering the cloud as long as I keep getting access to the service level I (or my employer) is paying for.
With Google trying non-Intel option, there may be a copy-cat effect for other cloud providers to do the same to remain competitive, possibly diminishing Intel's importance in the server market.
And with "kids" mostly using non-Intel(or x86 in general) platforms (phones, TVs, snapchat-shades, phones etc) where Intel already has very little presence.
Not trying to sound a doomsday prophecy for Intel but... times they are a changing ?
Maybe when used hardware with these customisations starts showing up on eBay or similar...
[1] https://twitter.com/danluu/status/705802676945072129
Buying Intel is easier, so Intel gets to guess how far they can go before one of their large customers decides that the risk of going with one of those alternatives is worth it (demonstrating to the world that Intel isn't the only viable option in the process, which might at the end be the bigger damage for Intel)
Or for 1/10 the cost, they could buy AMD. Not because their products are equivalent, but because any new owner of AMD also acquires a very special relationship Intel. Intel and AMD have cross-licensed a lot of IP, without which neither would be able to sell an x64 processor. And (as I understand it) if AMD changes ownership, their licensing deal with Intel is nullified and would need to be renegotiated:
However, according to Drew Prairie, director of corporate communications at AMD, once ownership of either AMD or Intel changes, the whole agreement is terminated for both parties. As a result, after a transaction happens, the companies need to negotiate a new cross-licensing agreement.
“Actually, the agreement is pretty clear – if there is a change of control for either company the agreement is terminated,” said Mr. Prairie. “That does not mean a new agreement could not be reached, but in a change of control the agreement is terminated.”
http://www.kitguru.net/components/cpu/anton-shilov/amd-clari....
The assumption is that obviously the deal would be renegotiated, since it would be "Mutually Assured Destruction" for both AMD and Intel if either side refused. But what if AMD is bought by a company with deep enough pockets to just absorb the loss? What if AMD was owned by someone who didn't actually need it, and could countenance the MAD?
I'd think that such a company would have tremendous leverage against Intel in such a negotiation. If they demanded it, and were willing to "throw the steering wheel out of the window and play chicken"[1] (ie, structure the acquisition such that both Intel and AMD were prevented from selling x64 processors unless Intel agrees to their terms), exclusive early access (and whatever other special preference they want) would be easily negotiated.
But if this were true, I'd have thought that someone would have already done it. If not Amazon/Google/Facebook (the three largest Internet companies), why not Tencent/Alibaba/Baidu (the next largest)[2]? All of them could easily afford it, and I'd think that all of them could get tremendous benefit from a sweetheart deal with Intel. What am I missing?
[1] https://bpicampus.com/2011/02/28/evening-focus-applied-game-...
[2] https://en.wikipedia.org/wiki/List_of_largest_Internet_compa...
"Several trusted sources say that later today, likely at Supercomputing 16, the company will announce they have pulled in Purley aka Skylake-EP Xeons, to this year and will sell them to two key customers."
Customers which the article goes on to name as Google and Facebook. Although, I cannot find any corroborating evidence to support this claim. Supercomuting 16 is now over, and I see no other articles detailing how "Intel is playing favorites with Skylake Xeons in a way that will shatter the industry status quo." SemiAccurate is literally the only site I can find professing this scenario.
I'm not saying it's necessarily untrue, but I'd like a lot more information on why it's just Google and Facebook.
- Was there a silent bidding war on early access?
- Did Intel simply choose two major customers based on previous sales?
- Were Google and Facebook specifically targeted to exploit FPGA for AI development?
The list goes on ...
https://cloudplatform.googleblog.com/2016/11/power-up-your-G...
SemiAccurate has a pretty good track record with these kinds of rumors and analysis. They're not right 100% of the time, but they're right enough that it's clear they aren't just making things up.
And with Thanksgiving, I'm not sure there's been enough time for it to further leak out.
On the other hand, it's a claim that Intel will do something that's both extraordinary and extraordinarily stupid for the long term, so we indeed need some more proof.
Although I would be really surprised if this was something they actually announced, it doesn't really help Intel to announce something like that and as the article points out it can anger other customers.
That said I have worked in Intel's System Validation group which was responsible for determining when a new chip was ready for general availability. I have directly witnessed Intel's preferential customer treatment where Intel allocated all of their processor production on a new processor to their largest customer(s). In my case they were selling Nehalem chips to Dell long (6 - 9 months) before others like NetApp could get them to start building filers with them. I have also worked in the platforms group at Google so I have an idea of the scale of their hardware effort and how many chips such a deal with Google might entail.
As a result, I think the actual story is this -
Intel got back first silicon that was acceptably functional on the new parts. This kicked off a discussion between their top customers and the potential early availability of parts. Those customers did a calculation on what advantage they might get by getting hardware early. Then Intel went down their customer list from most strategic to least strategic and offered them access. Either Google or Facebook was on the top of that list (likely Google) and they bought some, but not all of the production capacity, and then Intel offered the remainder to the second customer on the list who agreed to take the rest. Everyone else has to wait until the part is released to general production in all of its target fabs.
Further, it won't cause Intel any grief because all their customers know that the ones with the most volume get the most "love" from Intel. And the situation might be different the next time around.
It's the breaking the promise more than anything else that problematic, if the sources, presumably from these screwed companies, are accurate.
Looking at some outside evaluations of their capex, it does not seem to represent a significant drop since CPU cost is only a tiny part of operating a datacenter.
http://continuance.co/revisiting-capex/
1MWh of electricity is enough to power a cpu running 114W on average for a year. A 24 core (48 thread) xeon has a 165w tdp, meaning that even such a beast is likely to use less on average (most applications don't have 100% cpu load 100% of the time, and virtualizing up to that load causes serious issues when several clients happen to simulatneously peak). In essence, this should run pretty much any single CPU, and even if you're doing something weird to actually get that 100% load, we're not far off.
1MWh of electricity is around 40$ at wholesale prices (https://www.eia.gov/electricity/monthly/update/wholesale_mar...). Even if google pays 100, then it looks to me as if straightforward server cpu energy usage is unlikely to be a huge issue.
Of course, there are power supply losses and cooling overheads, but even if that raises the power consumption by a factor 5 (doubtful), then a several-thousand dollar server cpu is still going to outweigh energy costs for many years.
Furthermore, energy costs are trending downwards and cpu's are still getting more efficient; it's a somewhat risky investment thus to buy a more expensive (but more efficient) chip if that investment needs years to pay off.
So, it looks to me as though plain old cpu costs certainly are far from irrelevant; they may indeed exceed energy costs over the chips lifetime if the energy is acquired at close to wholesale prices and cooling is efficient (say 3x overhead?).
https://www.reddit.com/r/pcmasterrace/comments/3smipw/psa_th...
http://www.techspot.com/review/1155-affordable-dual-xeon-pc/
It's an intricate play with variables like marketing opportunities (globally and locally), silicon availability (when, in which numbers), performance per dollar, etc etc. You could even get into this game by buying "just" like 5-10k CPUs in a quarter, if you're in the right region and have the proper local branding quality.
The concept of "launch customers" globally/per market doesn't seem very new either. That window of exclusivity is usually so short that it matters more from a marketing point-of-view (for both seller and buyer) than anything else.
E.g. a 48 core 64 bit ARM processor: http://www.cavium.com/ThunderX_ARM_Processors.html
So, Amazon, Alibaba, you know what to do now.