[Canard PC Hardware] Intel prepares Ryzen's response behind the scenes

[raghu78]

You aren't looking back far enough, sir!

Nehalem for servers launched in Nov. 2008. When did Lynnfield launch, again? That's right -- almost a year later, in September of 2009 🙂

The reason being Nehalem removed the major drawback of Intel Xeons at that time which used FSB and brought a point to point interconnect called QuickPath. Moreover Nehalem launched on a mature 45nm process a year after Penryn desktop launched. Different times and different situations. Lets see if Intel repeat it. Maybe their next grounds up new architecture which might come next decade may launch on server/HEDT.

 

[sm625]

I think Intel's response is a lot more than this, and is a LOT more behind the scenes than this. I think KBL is so mediocre because they dont want to waste a lot of effort improving it because they are focusing on CFL or maybe even further out than that. I get the feeling that when Intel decides to drop this bomb, it will be just like Conroe...

 

[Arachnotronic]

The reason being Nehalem removed the major drawback of Intel Xeons at that time which used FSB and brought a point to point interconnect called QuickPath. Moreover Nehalem launched on a mature 45nm process a year after Penryn desktop launched. Different times and different situations. Lets see if Intel repeat it. Maybe their next grounds up new architecture which might come next decade may launch on server/HEDT.

You're getting warmer!

Look, the first 10nm product from Intel is going to be...*drumroll*...Cannon Lake-Y. The mainstream DT as well as HEDT chips in 2018 will be on some flavour of 14nm. After Cannon Lake-Y has cleaned the ol' 10nm pipe a bit, buying the process guys a chance to work out yield issues, improve performance, etc. I think we will see Ice Lake-S/H/U as well as Ice Lake server come out pretty close to each other.

The key here is that the "Y" series chips will pave the way for all of the other chips, higher-end client and server alike.

At least that's my opinion, take it for what it's worth.

 

[JDG1980]

Do you think Intel told its CPU architects to go for a multi-year vacation? Do you really think that any company with just even baseline competence -- not talking Steve Jobs excellence, but just basic competence -- would, after hearing in May 2015 directly from AMD that it is building Zen (and all of the major performance info was known -- 40% more IPC than XV, SMT, 8 cores, etc.) -- would just sit around and do nothing in response?

That's on top of the many ARM-based SoC makers like Qualcomm, Cavium, Broadcom, HiSilicon, governments, etc. all announcing their own server SoC efforts targeting Intel, with ARM saying that they expect to take 25% marketshare in servers by 2020.

Come on, this defies common sense and basic logic. You know this.

In the 1970s, it "defied common sense and basic logic" that General Motors (the biggest and most profitable company in the world) would just sit around and let the Japanese competitors eat it alive while it released flop after flop. Everyone assumed that, sooner or later, they would come up with the "import fighter" car that would solidify their dominant position once again. But it never happened. Things were all downhill (with a slight blip in the late 90s and early 00s, thanks largely to SUVs and pickup trucks) until bankruptcy hit in 2008.

I see Intel today as being much like GM in the 1970s-1980s. Their arrogance blinds them. They have come to take their dominance for granted and assume, like GM in the bad old days, that their size and wealth means they can use capital spending to drown their competitors. Just as Intel thinks they will always have a process lead, so did GM think that they could build "Lights Out" robot factories that would let them build cars with next to no staff, at low marginal cost, and crush the competition once and for all. They were no match for smaller but leaner, less bureaucratic, and hungrier adversaries. Excessive market segmentation was one of the things that killed GM (google "badge engineering") and it will help to hasten Intel's decline over the coming decades.

 

[[DHT]Osiris]

In the 1970s, it "defied common sense and basic logic" that General Motors (the biggest and most profitable company in the world) would just sit around and let the Japanese competitors eat it alive while it released flop after flop. Everyone assumed that, sooner or later, they would come up with the "import fighter" car that would solidify their dominant position once again. But it never happened. Things were all downhill (with a slight blip in the late 90s and early 00s, thanks largely to SUVs and pickup trucks) until bankruptcy hit in 2008.

I see Intel today as being much like GM in the 1970s-1980s. Their arrogance blinds them. They have come to take their dominance for granted and assume, like GM in the bad old days, that their size and wealth means they can use capital spending to drown their competitors. Just as Intel thinks they will always have a process lead, so did GM think that they could build "Lights Out" robot factories that would let them build cars with next to no staff, at low marginal cost, and crush the competition once and for all. They were no match for smaller but leaner, less bureaucratic, and hungrier adversaries. Excessive market segmentation was one of the things that killed GM (google "badge engineering") and it will help to hasten Intel's decline over the coming decades.

I'd argue more that they're like Cisco for the last 20-30 years. Everyone seems to talk about all the upstarts coming in and beating them at their game, but they have such an immense level of the market share and they're continuing to advance (even if it's minor iterations) that they can't be bothered to address everyone announcing the downfall of their empire.

AMD as as much market share as Intel lets it, I'm confident it could crank up the heat if it felt like it, undercut AMD by 33% with 33% more powerful products, and grind them in the dust without blinking. It doesn't because it'd get Ma' Bell'd out of existence.

 

[Arachnotronic]

In the 1970s, it "defied common sense and basic logic" that General Motors (the biggest and most profitable company in the world) would just sit around and let the Japanese competitors eat it alive while it released flop after flop.

How many "flops" has Intel released in recent years in its core markets (mobile SoCs is a different story altogether)? They have been putting out pretty successful products that have sold as well as could be expected.

Everyone assumed that, sooner or later, they would come up with the "import fighter" car that would solidify their dominant position once again. But it never happened. Things were all downhill (with a slight blip in the late 90s and early 00s, thanks largely to SUVs and pickup trucks) until bankruptcy hit in 2008.

Intel is literally made $10B+ in net income last year. It is in a dominant position in PCs and servers. How is this analogy at all relevant?

I see Intel today as being much like GM in the 1970s-1980s. Their arrogance blinds them. They have come to take their dominance for granted and assume, like GM in the bad old days, that their size and wealth means they can use capital spending to drown their competitors.

Um, no. They are putting out good products that sell well.

Just as Intel thinks they will always have a process lead, so did GM think that they could build "Lights Out" robot factories that would let them build cars with next to no staff, at low marginal cost, and crush the competition once and for all. They were no match for smaller but leaner, less bureaucratic, and hungrier adversaries. Excessive market segmentation was one of the things that killed GM (google "badge engineering") and it will help to hasten Intel's decline over the coming decades.

Market segmentation has helped Intel boost profits over the years. They are dominating the PC processor market, they're now raking in hundreds of millions of dollars in revenue from sales of modems to Apple, their data center biz is nearing $20 billion in annual revenue, etc.

Your analogy doesn't seem to map well to Intel's current situation.

 

[raghu78]

You're getting warmer!

Look, the first 10nm product from Intel is going to be...*drumroll*...Cannon Lake-Y. The mainstream DT as well as HEDT chips in 2018 will be on some flavour of 14nm. After Cannon Lake-Y has cleaned the ol' 10nm pipe a bit, buying the process guys a chance to work out yield issues, improve performance, etc. I think we will see Ice Lake-S/H/U as well as Ice Lake server come out pretty close to each other.

The key here is that the "Y" series chips will pave the way for all of the other chips, higher-end client and server alike.

At least that's my opinion, take it for what it's worth.

When I said 12 months you said you will beat against it now what is your pretty close scenario. A quarter or two ? I will say this by the time Icelake mainstream lands we might see Cannonlake server/HEDT. From the Sandy bridge generation we are yet to see a less than 12 month gap between mainstream and Xeon/HEDT.

 

[JDG1980]

How many "flops" has Intel released in recent years in its core markets (mobile SoCs is a different story altogether)? They have been putting out pretty successful products that have sold as well as could be expected.

But that parenthetical is precisely the issue. Intel has already been completely elbowed out of the massive market for smartphones, tablets, and IoT devices by ARM. This is just like how, in the 1970s and 1980s, GM could still often make good large cars (traditionally their strong point) but all of their attempts at small cars were utter crap. And unfortunately for them, the trend at the time was toward smaller, cheaper, more fuel-efficient vehicles, just as today the trend is away from desktops for an increasing number of tasks.

Intel is literally made $10B+ in net income last year. It is in a dominant position in PCs and servers. How is this analogy at all relevant?

We're still in the early days of Intel's decline and fall. In 1970, GM made $1.7 billion in profits; taking inflation into account, that's equivalent to over $10 billion today, exactly like Intel.

GM had a massively dominant position in the automotive market until fairly recently. Their market share was well over 40% during the post-WWII era until the 1980s, far more than any other competitor. All imports combined were usually 10%-20% of GM's market share alone. That bought them time, but in the long run it did not protect them from the effects of stupidity and arrogance. In fact, it arguably exacerbated these traits, because GM executives had the same unrealistically bullish outlook that Intel's boosters have today.

Market segmentation has helped Intel boost profits over the years. They are dominating the PC processor market, they're now raking in hundreds of millions of dollars in revenue from sales of modems to Apple, their data center biz is nearing $20 billion in annual revenue, etc.

Your analogy doesn't seem to map well to Intel's current situation.

Sure it does. Market segmentation works great when you're the dominant player in the industry. But it falls apart quickly when real, meaningful competition shows up. GM's five different brands were a boon in the 1950s, but a massive albatross in the 1990s. (At least Intel doesn't have to haggle with dealers protected by state law if it wants to simplify its lineup.)

 

[[DHT]Osiris]

We're still in the early days of Intel's decline and fall. In 1970, GM made $1.7 billion in profits; taking inflation into account, that's equivalent to over $10 billion today, exactly like Intel.

You do realize you're attributing success to an inevitable failure, right?

You do understand the logical fallacy behind this, yes?

GM, like other automakers, fell for a variety of reasons. Their success in several market segments, size, and wealth were not primary reasons.

 

[JDG1980]

You do realize you're attributing success to an inevitable failure, right?

You do understand the logical fallacy behind this, yes?

GM, like other automakers, fell for a variety of reasons. Their success in several market segments, size, and wealth were not primary reasons.

I'm not arguing that success inevitably leads to failure. I was countering Arachnotronic's argument that Intel must be doing fine strategically for the future because they are making big profits now. My contention is that too much success for too long can (not necessarily will in all cases, but often can) breed complacency, which leads in turn to failure to anticipate new markets, engage in meaningful self-improvement, and cut off smaller, hungrier competitors at the pass. And in the long run (several decades), this syndrome can bring down even the largest corporate giant.

 

[itsmydamnation]

I'd argue more that they're like Cisco for the last 20-30 years. Everyone seems to talk about all the upstarts coming in and beating them at their game, but they have such an immense level of the market share and they're continuing to advance (even if it's minor iterations) that they can't be bothered to address everyone announcing the downfall of their empire.

AMD as as much market share as Intel lets it, I'm confident it could crank up the heat if it felt like it, undercut AMD by 33% with 33% more powerful products, and grind them in the dust without blinking. It doesn't because it'd get Ma' Bell'd out of existence.

No they couldn't. Intels stock price would tank on the complete lack of revenue and the board would get the boot in an instant. Its not the period of the yearly 2000's where every sector intel was in had massive growth allover it so you could have a war chest and still deliver massive revenue and GP numbers.

Zen competes with Zepplin from 6 core to 32 core, RR will compete in laptops. Vega and GP 100 will compete in highend compute while being paid for by consumer GFX or Sony/MS. Consider things like AMD will have more chip capacity with dual sourcing from Samsung and GF then it has ever had before. That its Zeppelin strategy helps with yields vs monolithic 32 core parts. That AMD with P10 @ 232mm sq with PCB, VRM, GDDR5, packaging/testing and AIB margin is selling from 150USD to 260USD and making something like 30-35% GP.

So if intel wanted to go to WAR with AMD. AMD could sell 24-32core server parts for 600 to 1000 USD and 6core consumer parts @ 150USD and likely increase both revenue and GP from today. If they try to bring down the price of KNL they will have to fight NV as well as AMD and both have other revenue sources for that part to subsidize that fight .

Then look at things like intel has to feed ~110k mouths while amd only 10K. That the price of manufacturing per Zepplin is in the $20-$30 range then fighting on price doesn't look like a good idea.

How about this for a thought, intel trys to sink AMD with massive price cuts, AMD has way less opex costs so can still under cut, that under cut leads to more market share for AMD which actually increases intels manufacturing costs from lower fab utilization, how does the next intel AMG look?

So what exactly can intel do? to flight on price will be bloodly for intel and likely not hurt AMD anymore then it already has. They need to fight on performance and that takes time, so if intel has wasted the last 4 years ( not saying they have, i dont know) then they have a very delicate fight on their hands for the next 4 until the choices they make today flow through to products.

 

[mikk]

When I said 12 months you said you will beat against it now what is your pretty close scenario. A quarter or two ? I will say this by the time Icelake mainstream lands we might see Cannonlake server/HEDT. From the Sandy bridge generation we are yet to see a less than 12 month gap between mainstream and Xeon/HEDT.

Cannonlake Server might be cancelled as well, means there is room for a launch of Icelake Server 2019. Is there a recent server Roadmap with CNL-EP on it?

 

[dfk7677]

How many "flops" has Intel released in recent years in its core markets (mobile SoCs is a different story altogether)? They have been putting out pretty successful products that have sold as well as could be expected.

I would consider KBL a "flop". That is because in 1.5 years after SKL they only managed to increase clocks by 5%, no IPC gain, probably worse perf/W. I am no expert in Intel processor history, but I think they never offered such a low increase in performance (IPC, clock, perf/W) between 2 families of processors.

In my eyes, Intel are dragging their feet, maybe because there is no competition or maybe because they are conceited.

 

[Arachnotronic]

Cannonlake Server might be cancelled as well, means there is room for a launch of Icelake Server 2019. Is there a recent server Roadmap with CNL-EP on it?

I haven't been able to find anything on a Cannon Lake-EP. The only references to Cannon Lake server I have found on LinkedIn for example have been to potential server SoC products (i.e. Xeon D).

 

[Atari2600]

I am no expert in Intel processor history, but I think they never offered such a low increase in performance (IPC, clock, perf/W) between 2 families of processors.

I'd say the "jump" from Coppermine to Willamette would take it...

 

[mikk]

I haven't been able to find anything on a Cannon Lake-EP. The only references to Cannon Lake server I have found on LinkedIn for example have been to potential server SoC products (i.e. Xeon D).

I didn't see anything either recently. In the recent (Saphire Rapid) Server Roadmap there wasny't anything, next after SKL-EP is Icelake. And in the Dell Roadmap below there isn't anything about CNL throughout 2018 either which is in line with the Sapphire Rapid Roadmap. If this is true then we can be sure Icelake Server is planned for 2019, same as the Consumer platform.

 

[Arachnotronic]

I didn't see anything either recently. In the recent (Saphire Rapid) Server Roadmap there wasny't anything, next after SKL-EP is Icelake. And in the Dell Roadmap below there isn't anything about CNL throughout 2018 either which is in line with the Sapphire Rapid Roadmap. If this is true then we can be sure Icelake Server is planned for 2019, same as the Consumer platform.

I have to agree with your analysis. Intel cancelled Sandy Bridge-EX when it became clear that it wouldn't be out in time so they just skipped to Ivy Bridge-EX.

I think Intel is realizing that sticking with Cannon Lake-EP would hurt them so it's better to just kill Cannon Lake-EP/EX to get the server cores back on track.

 

[Eddward]

I would consider KBL a "flop". That is because in 1.5 years after SKL they only managed to increase clocks by 5%, no IPC gain, probably worse perf/W. I am no expert in Intel processor history, but I think they never offered such a low increase in performance (IPC, clock, perf/W) between 2 families of processors.

In my eyes, Intel are dragging their feet, maybe because there is no competition or maybe because they are conceited.

I see it very simple from my point of view. Failed tick-tock. This is the root cause of everything what we see now. Intel have had a very clear roadmap for several years ahead with this in mind. Year 2017 should have seen another tock on 10nm after 2 years of Skylake, but what happened. 10nm is far from ready and instead of that, they have been pushed to release another 3rd 14nm product rush-created from nothing and even 4th is on the way. Clearly they didn't have a proper plan B if 10nm won't go as smooth as planned.

 

[raghu78]

I see it very simple from my point of view. Failed tick-tock. This is the root cause of everything what we see now. Intel have had a very clear roadmap for several years ahead with this in mind. Year 2017 should have seen another tock on 10nm after 2 years of Skylake, but what happened. 10nm is far from ready and instead of that, they have been pushed to release another 3rd 14nm product rush-created from nothing and even 4th is on the way. Clearly they didn't have a proper plan B if 10nm won't go as smooth as planned.

very true. What was Intel's Plan B if 10nm was delayed. None. The earliest we might see Icelake is Q1 2019. No tock for close to 3.5 years is plain stagnation.

 

[Arachnotronic]

very true. What was Intel's Plan B if 10nm was delayed. None. The earliest we might see Icelake is Q1 2019. No tock for close to 3.5 years is plain stagnation.

Plan B was to soup up 14nm by enhancing it by as much as the rest of the industry is enhancing their 10nm processes compared to their 14nm/16nm processes, tweak the circuit design to allow for higher frequencies without increasing power, and add new media capabilities that have a serious impact on user experience.

In mobile, Kaby Lake is arguably a bigger improvement in CPU perf than either Broadwell or Skylake were. But you seem so set in your belief that Intel sucks that you ignore a lot of this goodness 🙁

 

[PPB]

Plan B was to soup up 14nm by enhancing it by as much as the rest of the industry is enhancing their 10nm processes compared to their 14nm/16nm processes, tweak the circuit design to allow for higher frequencies without increasing power, and add new media capabilities that have a serious impact on user experience.

In mobile, Kaby Lake is arguably a bigger improvement in CPU perf than either Broadwell or Skylake were. But you seem so set in your belief that Intel sucks that you ignore a lot of this goodness 🙁

Problem is there is so much you can get out of the process. AMD made the same thing with Kaveri, Godavari, Carrizo and BR but from their uarch perspective as 28nm couln't be extracted more than we have seen these years. Both approaches feel lackuster without the other.

Kaby Lake shouldn't be praised as this was the norm in the Nehalem era. What we should be asking ourselves was: what went wrong with Broadwell on mobile compared to Haswell and considering the node change that a mere more optimized refresh like Kabylake is as nice if not better perf-jump wise? This only proves Intel's 14nm was a turd of a process at launch and we are getting with Kaby lake the things that should be there from the tick in the first place, not 2 gens later (Process-architecture-optimization).

This is why Kaby lake is a snooze fest for everyone but the most loyal fans: it's just 14nm from the renounced best foundry bussines in the world, gotten right at the third try.

 

[Arachnotronic]

*

Problem is there is so much you can get out of the process. AMD made the same thing with Kaveri, Godavari, Carrizo and BR but from their uarch perspective as 28nm couln't be extracted more than we have seen these years. Both approaches feel lackuster without the other.

You can get a lot out of the process if you make significant enough changes.

Remember, die shrinks in themselves don't improve performance -- they reduce chip area which, if everything goes right, improves cost. We saw with TSMC 16FF+ that you could get a big performance boost without shrinking, and what we saw with 14nm -> 14nm+ is like the 20SoC -> 16FF+ transition (though the performance gain wasn't as dramatic).

Kaby Lake shouldn't be praised as this was the norm in the Nehalem era. What we should be asking ourselves was: what went wrong with Broadwell on mobile compared to Haswell and considering the node change that a mere more optimized refresh like Kabylake is as nice if not better perf-jump wise? This only proves Intel's 14nm was a turd of a process at launch and we are getting with Kaby lake the things that should be there from the tick in the first place, not 2 gens later (Process-architecture-optimization).

14nm was problematic at launch, no question of it -- the yields were just challenging (functional and parametric). I also think Broadwell itself was a very troubled child, you can see it with Broadwell-E -- Intel added more cores and IPC went up a little bit, but Fmax came down. This smells like a problem with the physical design more than anything else. Implemented correctly, that's not supposed to happen. Skylake was a much better architecture on 14nm and the 14nm process probably underwent some maturity (though no significant transistor structure changes).

This is why Kaby lake is a snooze fest for everyone but the most loyal fans: it's just 14nm from the renounced best foundry bussines in the world, gotten right at the third try.

Ideally Intel would have brought out a souped up uArch alongside the new 14nm+ process, but that is risky. If you keep the same uArch but migrate to a new process/physical design, it's much easier to know where any potential bugs are. If you change both variables at the same time (uarch and process/physical implementation), it becomes harder to validate your CPU and you either get schedule slips or you get probably what we saw with Broadwell -- kind of a mess.

 

[IntelUser2000]

Do you think Intel told its CPU architects to go for a multi-year vacation? Do you really think that any company with just even baseline competence -- not talking Steve Jobs excellence, but just basic competence -- would, after hearing in May 2015 directly from AMD that it is building Zen (and all of the major performance info was known -- 40% more IPC than XV, SMT, 8 cores, etc.) -- would just sit around and do nothing in response?

No, I believe Intel has been having serious internal problems for a few years now and the direction given to the engineers have been erratic so the people who should have been working towards improving products weren't doing such a good job at it.

I am not too surprised by significant increase in TDP for 100MHz boost.

A 16 years ago, Intel announced the "Terahertz Transistor" project. That is a very early announcement of the now known Tri-Gate transistor used since 22nm in 2012. There's a good reason they named it such. They believed that CPUs would reach 20-30GHz in record time and this "Terahertz Transistor" would enable them to do so.

Fast forward 5 years later, we know its a bust. We won't see 10GHz, let alone 20-30GHz.
Fast forward another 10 years later, I think its becoming clear that we won't see much above 5GHz. The 10GHz barrier has never, I repeat, never been breached, even with exotic cooling solutions and unrealistic setups(Celerons with 1 core for example).

In mobile, Kaby Lake is arguably a bigger improvement in CPU perf than either Broadwell or Skylake were.

I think if we just think of the x86 world, they are doing fine with mobile KBL. But, we can't do that. Apple with likely far better execution(likely resulting from happier people) is doing what KBL is doing without the high idle power, high cost of CPU, and high thermals. Even on the GPU side we get basically Iris level performance.

 

[frozentundra123456]

PCGamesHardware.de posted CPCHardware news this morning and they just added some more information based on their own sources.

http://www.pcgameshardware.de/Kaby-...ntels-Antwort-Core-i7-7740K-i5-7640K-1220177/

The i7-7740K and i5-7640K do exists, but they are 112W Kaby Lake-X (Socket 2066) that should be released for Gamescom. It's the same Kaby-Lake-S die "just" with a disabled iGPU. TIM will still be used.

There will be a 100 Mhz base clock improvement on the 7740K, but the turbo should stay the same @ 4.5Ghz.

You do realize you're attributing success to an inevitable failure, right?

You do understand the logical fallacy behind this, yes?

GM, like other automakers, fell for a variety of reasons. Their success in several market segments, size, and wealth were not primary reasons.

Yea, all the problems attributed to GM were there, but union benefits and labor costs were also a huge albatross around their neck.

 

[CentroX]

What i like about lisa su is that she is going after the classic enthusiast market; the cpu and gpu desktop. No gimmicks, just good ole pc parts.