AMD: Success of small cores vs. big cores?

[strata8]

I'm curious about this. My anecdotal evidence is that of the 10 AMD laptops my local electronics store sells, 9 of them are based on Kabini.

Also, how do they compare in cost? Kabini was 107mm2 and Richland was 248mm2. Beema and Kaveri are similar, and they're both on GlobalFoundries' 28nm.

 

[frozentundra123456]

How do you define "success"? Total sales, profit, performance, performance per watt, battery life?

Are you talking about desktop, mobile, servers??

Personally, I think performance wise compared to their direct competition, the small cores are most competitive. I would not consider a kabini laptop though, the sacrifice in performance is too much, and at least in the models I have seen, battery life is not outstanding enough to make up for the low performance.

 

[strata8]

How do you define "success"? Total sales, profit, performance, performance per watt, battery life?

Are you talking about desktop, mobile, servers??

Personally, I think performance wise compared to their direct competition, the small cores are most competitive. I would not consider a kabini laptop though, the sacrifice in performance is too much, and at least in the models I have seen, battery life is not outstanding enough to make up for the low performance.

Sales and profit wise, mostly in mobile because that's where they overlap. We already know - roughly - how they stack up against one another in performance.

My impression is that with such a large die and a top bin that only competes with Intel's (cheaper and smaller) low end, AMD won't exactly be raking in the cash with their big core parts. The small cores at least are better positioned in this regard.

 

[mrmt]

My impression is that with such a large die and a top bin that only competes with Intel's (cheaper and smaller) low end, AMD won't exactly be raking in the cash with their big core parts. The small cores at least are better positioned in this regard.

Your perception is mostly correct. The entire big core line is, in the words of an AMD SVP, an unmitigated failure for the company. AMD lost market share in every single segment they were competing with their big core line.

The small core line is the line that brings money to AMD's war chest, the one that has better margins, better cost structure and better value proposition for the supply chain.

 

[raghu78]

Your perception is mostly correct. The entire big core line is, in the words of an AMD SVP, an unmitigated failure for the company. AMD lost market share in every single segment they were competing with their big core line.

The small core line is the line that brings money to AMD's war chest, the one that has better margins, better cost structure and better value proposition for the supply chain.

yeah the big core is the primary reason for AMD's problems in their computing division. Server market share is at an all time low of <3% . AMD has almost zero CPU presence in high end desktops and notebooks. AMD's fortunes now rest on the 2 new high performance cores releasing in 2016.

 

[Ajay]

yeah the big core is the primary reason for AMD's problems in their computing division. Server market share is at an all time low of <3% . AMD has almost zero CPU presence in high end desktops and notebooks. AMD's fortunes now rest on the 2 new high performance cores releasing in 2016.

So, nothing new in Big core in 2015? If so, AMD is clearly pushing most of it's R&D into small core, ARM and GFX.

I would sort of hope that AMD updated Kaveri in 2015 with higher performing SKUs and some sort of solution to their bandwidth problem - but that may have been tough to do on the cheap.

 

[rainy]

So, nothing new in Big core in 2015?

No, we should see in the next year Kaveri successor (Carrizo) with Excavator cores (more precisely modules).

 

[PPB]

What most people still dont get is how one type of core or another is designed with a specific target in mind. It is correct that Cat family is smaller in die size than Dozer family, and they can achieve similar or lower power consumption at normalized clocks and voltages, and that IPC is the same at normalized clocks...


... See a pattern here? Yes, I have to say "at normalized" clocks every time I try to compare them in about anything valuable besides die size, because one product was designed to run at 3++Ghz, while the other at >2ghz clocks.

The problem here is that you are trying to compare them by lowering the Dozer family clocks (justifying a potential mobile scenario), making it a uncomfortable situation for that kind of uarch, because the dozer family intended to reach their IPS target via clocks. Can we compare them by overclocking the cat family to the dozer family range of clocks? Right now, we cant, and that only is enough to tell both designs are suited for different purposes.

That we are able to get mobile offerings from the big core family is only determined because it is a more flexible design, thanks to its higher xtor count and subsequent die size. When you throw more xtors at the problem, you can scale back your performance to reach a wider range of power envelopes, even at a better perf/watt that the lower xtor count, but higher clocked variant. With the small core we just can't, and this is such a limitation that it's not a coincidence that the latest releases of the cat family of cores have all being about squeezing every mhz out of the design without leaving the targeted power envelopes. In other words, trying to find ways to scale upwards.

The problem with the big core family is that they aren't focusing enough in improving it's IPC (after it has been proven time after time that the 6hz target clock is not coming anytime soon, even less from GF) lately, but there has been too many "diversions" in the middle to deal with (making it HSA compilant, steering a little from the pure CMT design to compensate the max ammount of cores being halved in their top end consumer products, etc) to just focus on better IPC. The fact that they are stuck on a bigger node doesnt help at all, either, as a key point into making the Intel Core design what it is today was a succession of incremental width upgrades after each node shrink. The dozer core family is really longing for a doubling in it's width right now (and by judging the TDP of the top Carrizo offering, I just dont really think it will be coming at all).

 

[mrmt]

snip

The "unmitigated failure" statement from Mr. Feldman is quicker to explain the situation. AMD screwed up with their big core, end of history.

 

[Fjodor2001]

The "unmitigated failure" statement from Mr. Feldman is quicker to explain the situation. AMD screwed up with their big core, end of history.

Regardless, that's history. AMD is developing a new 64-bit x86 big core and the design is headed by Jim Keller. Likely they'll use the successful cat cores as base, so it's not unlikely we'll see AMD big cores shine once again in a few years time. Intel better watch out, they've forgotten what Andrew Grove said - "only the paranoid survive".

 

[Arachnotronic]

Intel better watch out, they've forgotten what Andrew Grove said - "only the paranoid survive".

Is that why Intel aggressively went after the low end with Bay Trail-M and further contributed to the collapse of AMD's computing solutions group? Because they're forgotten to be paranoid?

Intel spends 3x in a quarter what AMD spends in a year on R&D. I'm pretty sure you can count on Intel continuing to push the envelope in CPUs and while AMD might be touting an amazing 2016 core, you surely must realize that AMD has nothing for 2015 which is why they are talking up 2016.

2016 will be a lot better for AMD's competitive position but by the time that core rolls out on 20nm BEOL + 14nm transistors from GloFo, Intel will have blazed through two big cores on 14nm and will launch Cannonlake (10nm) in 2016.

Unless Intel screws the pooch big time, AMD will still have a tough time competing with Intel in high end X86 on performance/watt and cost structure.

 

[Fjodor2001]

I'm thinking about Intel delivering ~8% CPU performance increase per generation on desktop CPUs, effectively ~5-6% per year. If AMD designs a new big core x86 uArch based on the cat cores instead of Bulldozer they have a fair chance to catch up and perhaps even surpass Intel. It's happened before.

 

[erunion]

Intel aggressively driving down the BoM for future atoms systems is really going to put pressure on AMD in the low end.

 

[Arachnotronic]

I'm thinking about Intel delivering ~8% CPU performance increase per generation on desktop CPUs, effectively ~5-6% per year. If AMD designs a new big core x86 uArch based on the cat cores instead of Bulldozer they have a fair chance to catch up and perhaps even surpass Intel. It's happened before.

Ah, I see, so you're assuming that Intel is incompetent and doesn't understand the markets that it sells ~$50B worth of product into and is basically spinning its wheels by dropping $10B/year+ on R&D, while AMD - which has failed time and again to execute in PCs - on a relative shoestring R&D budget and with two pretty mediocre performance/watt CPU cores to begin with will simply hit a home-run/slam dunk.

Yeah, seems reasonable.

 

[Arachnotronic]

Intel aggressively driving down the BoM for future atoms systems is really going to put pressure on AMD in the low end.

Yeah, Bay Trail-M even with relatively lackluster GPU performance is taking share from Kabini/Temash, and it only gets worse with Braswell (Cherry Trail/14nm/BoM optimized) and then the Goldmont based part that is surely to follow.

Intel has the cost structure and the scale to make life in the PC market even worse for AMD, and it is already doing so. AMD has made it clear that its plans are to diversify into other markets and is pumping ARM like there's no tomorrow, so any restraint that Intel may have had in place at the low end in order to make sure AMD remains viable has now been removed.

Also, AMD has now telegraphed broadly 2 years in advance what its plans are, so Intel has plenty of time to react appropriately.

 

[Fjodor2001]

Ah, I see, so you're assuming that Intel is incompetent and doesn't understand the markets that it sells ~$50B worth of product into and is basically spinning its wheels by dropping $10B/year+ on R&D, while AMD - which has failed time and again to execute in PCs - on a relative shoestring R&D budget and with two pretty mediocre performance/watt CPU cores to begin with will simply hit a home-run/slam dunk.

Yeah, seems reasonable.

I'm thinking they are getting lazy and optimizing CPU releases and technology advances based on maximizing profit and returns on investment using marketing policies. They can do so thanks to their close to monopoly situation.

They got lazy one time before, and then AMD surpassed them. I'm saying it can happen again. It's time they get a bit paranoid as Mr Grove would say.

 

[Arachnotronic]

I'm thinking they are getting lazy and optimizing CPU releases and technology advances based on maximizing profit and returns on investment using marketing policies. They can do so thanks to their close to monopoly situation.

They got lazy one time before, and then AMD surpassed them. I'm saying it can happen again. It's time they get a bit paranoid as Mr Grove would say.

If you think Haswell is a 'lazy' effort then I don't really know what to tell you.

Plenty of innovation/work that went into Haswell. Please peruse the following slide deck:

https://intel.activeevents.com/sz14...hType=session&tc=0&sortBy=abbreviationSort&p=

 

[jpiniero]

Also, AMD has now telegraphed broadly 2 years in advance what its plans are, so Intel has plenty of time to react appropriately.

I think you underestimate the lead time for processors. 2 years is nothing. Intel's not going to be able to make changes that quickly to be able to react to any possible threat by any of the announcements made by AMD.

 

[AtenRa]

Yeah, Bay Trail-M even with relatively lackluster GPU performance is taking share from Kabini/Temash, and it only gets worse with Braswell (Cherry Trail/14nm/BoM optimized) and then the Goldmont based part that is surely to follow.

Intel has the cost structure and the scale to make life in the PC market even worse for AMD, and it is already doing so. AMD has made it clear that its plans are to diversify into other markets and is pumping ARM like there's no tomorrow, so any restraint that Intel may have had in place at the low end in order to make sure AMD remains viable has now been removed.

Also, AMD has now telegraphed broadly 2 years in advance what its plans are, so Intel has plenty of time to react appropriately.

Intel better start worrying about loosing 5B in mobile this year and even more the next year than AMDs ARM severs.

 

[Arachnotronic]

Intel better start worrying about loosing 5B in mobile this year and even more the next year than AMDs ARM severs.

Intel can afford to lose the $4b in mobile that it likely will this year. This is what strong, powerful companies can afford to do - pay the price to break into a large, established market with powerful incumbents.

Weak players like AMD tout powerpoint slides about their 2016 offerings and name-drop more powerful/successful companies like ARM and point to the fact that there are many competitors developing ARM chips as an "advantage" of the ARM ecosystem.

p.s. here's the best illustration of how poorly managed AMD has been over the last 30+ years relative to Intel:

 

[AtenRa]

Intel can afford to lose the $4b in mobile that it likely will this year.

Add another Billion for the Contra-Revenue

This is what strong, powerful companies can afford to do - pay the price to break into a large, established market with powerful incumbents.

The problem is, continue like this will not make them any profit in that segment.

AMD on the other hand can make 1B profit each year from ARM Server SoCs in 2017 onwards.

 

[Arachnotronic]

Add another Billion for the Contra-Revenue



The problem is, continue like this will not make them any profit in that segment.

AMD on the other hand can make 1B profit each year from ARM Server SoCs in 2017 onwards.

When was the last time AMD ever made $1 billion in profit in a year?

I'll wait.

 

[Arachnotronic]

From AMD's form 10-K filed in 2007...

Even at the height of its competitiveness with Opteron/Athlon, the best AMD managed to do during those halcyon days was generate a measly $232 million in operating income.

Looking a bit more recently, it looks like thanks to ATI and a PC market boom, they were able to generate nearly $500 million...

Even after "contra-revenue", Intel will generate about $12 billion in operating income during 2014.

Just sayin'.

 

[AtenRa]

When was the last time AMD ever made $1 billion in profit in a year?

I'll wait.

When was the last time AMD could be the biggest player in a 5-6B market ??

 

[Headfoot]

Only 20 posts before the thread completely fell apart. Not a record, but still pretty quick.


I think AMD has a chance to pick up some profit if they successfully iterate on their current Seattle/ARM server strategy. There is definitely a niche for many small power efficient cores, or many small cores that need huge amounts of memory and bandwidth but not much processing. AMD theoretically has the know-how to implement ARM's efficient cores with high-end server fabric and uncore, whether they can beat other ARM players or Intel into addressing that market is completely up in the air. But I think its possible so long as they keep their eye on the prize. It won't be like when x86 took over high-end RISC in the server space, but it could be a profitable niche.