[AT] AMD Reorganizes Business Units - no more high performance x86 cores?

[Fjodor2001]

This thread was started by a poster misunderstanding that Anandtech were displaying a slide from over two years ago, not a new "statement by an AMD executive".

Yes, it's been mentioned by someone in this thread that the slide is two years old. But it's not obvious when reading the AT article though since there is no indication of that in the article.

But anyway, isn't the slide still valid? Then how come AT included it in the article just days ago...?

Decisions taken two years ago may not show up in actual products until perhaps 2015-2016 or later, so it could fit with the new x86 uArch design. Just that AMD hasn't mentioned the new uArch until lately. Maybe they didn't want to give away details to the enemy for free...

 

[NTMBK]

Yes, it's been mentioned by someone in this thread that the slide is two years old. But it's not obvious when reading the AT article though since there is no indication of that in the article.

But anyway, isn't the slide still valid? Then how come AT included it in the article just days ago...?

Decisions taken two years ago may not show up in actual products until perhaps 2015+, so it could fit with the new x86 uArch design. Just that AMD hasn't mentioned the new uArch until lately. Maybe they didn't want to give away details to the enemy for free...

The slide says no more high performance, and since then AMD has announced two new high performance CPU architectures. So I would agree with your third paragraph. It's an old slide

 

[mrmt]

But anyway, isn't the slide still valid? Then how come AT included it in the article just days ago...?

The context of that slide is still very current. It was when AMD was shifting from a bleeding edge company and expanding in the lagging edge embedded market. It was all about focusing on cost efficient chips and not on top notch performers. It's about being faster and releasing simpler chips, projects with a smaller, more manageable scope.

And I think this is exactly what we can expect of K12, descent performance, good power consumption and reasonable cost but not a real competitor to Intel in most markets. In a sense this is AMD going back to its roots, offering cheaper competitors to Intel in the bottom market, exactly what they did before the acquisitions of DEC IP and NextIO, and guess what, they grew with this strategy.

 

[Fjodor2001]

The context of that slide is still very current. It was when AMD was shifting from a bleeding edge company and expanding in the lagging edge embedded market. It was all about focusing on cost efficient chips and not on top notch performers. It's about being faster and releasing simpler chips, projects with a smaller, more manageable scope.

And I think this is exactly what we can expect of K12, descent performance, good power consumption and reasonable cost but not a real competitor to Intel in most markets. In a sense this is AMD going back to its roots, offering cheaper competitors to Intel in the bottom market, exactly what they did before the acquisitions of DEC IP and NextIO, and guess what, they grew with this strategy.

I think such an approach has an even greater chance of success nowadays, because most consumers don't crave that much more CPU performance. So AMD does not have to surpass Intel, they just have to come reasonably close. I.e. around 70-80% of Intel's top mainstream desktop CPU performance (4770K). That definitely feels within range, with a new x86 uArch from AMD.

If AMD can achieve 70-80% of Intel's CPU performance, paired with better iGPU and a lower price, I think it can be a recipe for success.

 

[mrmt]

I think such an approach has an even greater chance of success nowadays, because most consumers don't crave that much more CPU performance. So AMD does not have to surpass Intel, they just have to come reasonably close. I.e. around 70-80% of Intel's top mainstream desktop performance (4770K). That definitely feels within range, with a new x86 uArch from AMD.

If AMD can achieve 70-80% of Intel's CPU performance, paired with better iGPU and a lower price, I think it can be a recipe for success.

I think this strategy will not make AMD fans happy, there isn't much fun in water cooled htpc chips after all. but it might ensure that the company will become profitable.

As for AMD gpu advantage, it is a thing of the past. Intel top end skus are already on par or beating AMD top APU in graphics and this technology is going to waterfall to the lower skus. 14nm and 10nm will just reinforce this trend.

 

[Fjodor2001]

I think this strategy will not make AMD fans happy, there isn't much fun in water cooled htpc chips after all. but it might ensure that the company will become profitable.

80% of Intel's top end mainstream CPUs (e.g. 4770K) is not a HTPC chip.

And anyway, it's more than enough for most consumers, so the market potential is huge.

As for the gamers/enthusiasts AMD could make an 8-core version, and sell it at a more reasonable price than what Intel is willing to do currently. That ought to make them happy, as games are transitioning to a more parallelized phase going forward (next gen consoles and all).

As for AMD gpu advantage, it is a thing of the past. Intel top end skus are already on par or beating AMD top APU in graphics and this technology is going to waterfall to the lower skus. 14nm and 10nm will just reinforce this trend.

But the thing is that Intel is brute forcing it's way forward to come close to AMD's iGPU performance. The Intel GPU cores are not as efficient as AMD's. So Intel has to waste more die area to add more cores (and eDRAM too) for the same level of performance. It's an expensive "workaround" strategy.

 

[mrmt]

80% of Intel's top end mainstream CPUs (e.g. 4770K) is not a HTPC chip.

And anyway, it's more than enough for most consumers, so the market potential is huge.

As for the gamers/enthusiasts AMD could make an 8-core version, and sell it at a more reasonable price than what Intel is willing to do currently. That ought to make them happy, as games are transitioning to a more parallelized phase going forward (next gen consoles and all).


But the thing is that Intel is brute forcing it's way forward to come close to AMD's iGPU performance. The Intel GPU cores are not as efficient as AMDs. So Intel has to waste more die area (and add eDRAM) for the same level of performance. It's an expensive "workaround" strategy.

i don't think AMD will aim for 80% of 4770k, that's too much for them as it would place them competing against the entire i5 stack. I think a reasonable goal is 80% of an i3, but costing significantly less. This is what will give then an edge on the bottom market.

intel can afford to "waste" die area and edram because of their node advantage if it is strategically important for them. it is AMD who must be very careful on how they spend their die budget, because of their foundry handicap.

 

[ShintaiDK]

80% of Intel's top end mainstream CPUs (e.g. 4770K) is not a HTPC chip.

Why not? It works fine as a media center and plays all movies. Even more correctly than others.

HTPC=Home Theater PC.

 

[sefsefsefsef]

I don't get how it's unreasonable to assume that AMD can make a quantum leap over its Bulldozer-based big cores. The worst part of those cores (and it's a really bad part) is its bewildering cache hierarchy. If they just straight-up copy Intel's cache sizes they would see a large increase in performance. Very little work to be done there.

Other major contributors to performance are things like branch prediction and prefetching, and there are some very well known old techniques (and some new ones) in each of those fields that AMD is not currently using, that if they were to just implement them would give them significant gains. It's not a very complicated strategy, but it does require solid execution in designing and implementing those structures within required timing parameters, and IMO it's entirely doable.

 

[mrmt]

I don't get how it's unreasonable to assume that AMD can make a quantum leap over its Bulldozer-based big cores. The worst part of those cores (and it's a really bad part) is its bewildering cache hierarchy. If they just straight-up copy Intel's cache sizes they would see a large increase in performance. Very little work to be done there.

Easier said than done. The fact that AMD engineers and managers decided to not go along with these ideas says a lot on how easy they are to implement on actual products.

 

[sefsefsefsef]

Easier said than done. The fact that AMD engineers and managers decided to not go along with these ideas says a lot on how easy they are to implement on actual products.

Changing cache sizes and associativity is trivially easy to do. The harder part would be to get rid of the terrible write-through policy between L1 and L2.

 

[mrmt]

Changing cache sizes and associativity is trivially easy to do. The harder part would be to get rid of the terrible write-through policy between L1 and L2.

then why didn't the engineers pursue that route and instead launched fundamentally crippled products?

 

[NostaSeronx]

The harder part would be to get rid of the terrible write-through policy between L1 and L2.

The write through policy is for global memory and it is possibly the best thing Bulldozer has. The issue is the interconnect between the L1 and L2. Not the write policy of which the L1 and L2 is deriving.

 

[krumme]

Cache latency probably have most to do with proces competences and technology than purely design. If anything with Intels grib on synergy here i would expect them to just as relative strong here in the future because amd declining proces competences will make it even harder to sustain competitive here. No way can they remotely do the same as Intel here.

Amd instead have to rely on arm and eg tsmc knowledge to spare ressources. Plenty of excellent opportunities here instead of wasting ressources on something they can never be compettitive on.

 

[Fjodor2001]

Why not? It works fine as a media center and plays all movies. Even more correctly than others.

HTPC=Home Theater PC.

Yes, just like the Intel Xeon E5-2697 12C/24T 2.7/3.5GHz HTPC chip.

 

[ShintaiDK]

Yes, just like the Intel Xeon E5-2697 12C/24T 2.7/3.5GHz HTPC chip.

I didnt know you considered that a top end mainstream chip. Because thats not how the rest of us sees it.

80% of Intel's top end mainstream CPUs (e.g. 4770K) is not a HTPC chip.

Seems to me you simply had a lack of understanding what HTPC actually is. Its not a gaming machine that can only be handled by IGP in the Richland/Kaveri/Iris Pro league

 

[Fjodor2001]

I didnt know you considered that a top end mainstream chip.

I didn't. But it qualifies according to your only definition/criteria for a HTPC CPU:

"It works fine as a media center and plays all movies."

Seems to me you simply had a lack of understanding what HTPC actually is. Its not a gaming machine that can only be handled by IGP in the Richland/Kaveri/Iris Pro league

Huh...? Where did I say that?

 

[chrisjames61]

ARM has been most of its existence unknown company that made mobile CPUs, they gained more attention only in last 3 years with mass deployment of android phones and tablets. So I assume they had to work very hard and earn shitloads of money over a time to be where they are now.

Back in the early 1990's Apple poured a lot of resources into Acorn to make the processor for the Newton. Acorn of course became ARM.