Fudzilla: New AMD Zen APU boasts up to 16 cores (plus Greenland GPU with HBM)

[Enigmoid]

Bulldozer was created for Throughput, the entire Module has more execution units (Int and FPU) than a single Intel Core with HT.

Just because the BD architecture has a smaller INT execution unit it doesnt mean AMD cannot create a wider INT Core. Star (Llano) had a wider core than BD and yet BD has higher ST performance.

If they will go with wider INT cores (3-4 ALUs vs 2 ALUs in BD mArch) in order to implement SMT within the module then IPC and ST performance will increase tremendously. Throughput and efficiency will also skyrocket but the Module die size will be bigger.
That will be compensated from the higher density and low power consumption of the 14nm FF node.

Kaveri has 4 ALUs per module. Haswell has 4 per core as well. Haswell has much more FP power too with wider units.

That said, adding more units brings diminishing returns

I'm not saying AMD is going to magically pull out 25-30% IPC gain from the new Arch...but personally I'd think of anywhere between 10-15% as realistic...and thus competetive if the price is right. About + 10% on Excavator IPC @ 3.8/4Ghz clock + an iGPU strong enough to master 1080P gaming? To me that would be an all kill.

Sure...Intel will still be ahead heaps in raw CPU power...but for the medium to low ranged gamer (as one example for a consumer base) this is unimportant as their budget wouldn't allow them to buy the fattest Intel CPUs, anyway if they still need a decent GPU.

Agree.

 

[el etro]

How are you adding a 125W GPU and 88W CPU and getting 288W?

Maybe counting the efficiency(85% ???) of the psu into this.

 

[AtenRa]

Apples and oranges. AMD needed a huge power hungry die with 8 cores to even compete with quadcores. 140W and 315mm2 without IGP and northbridge to compete with a 216mm2 SB 95W GT2 quad with integrated PCIe x16. All on 32nm.

Core i7 3820 is a 300mm^2 at 32nm with huge L2+L3 cache, this one should be compared to Bulldozer quad Module SKU and not the APU (Core i7 2600K).

 

[AtenRa]

Kaveri has 4 ALUs per module. Haswell has 4 per core as well. Haswell has much more FP power too with wider units.

That said, adding more units brings diminishing returns

The difference is that Hasswell can use all of them with a single thread for higher IPC, SteamRoller can only use two ALUs per INT core.

Although HW has more FP performance, the Bulldozer arch gives higher throughput per module for different loads. That is, Steamroller can execute one INT and one FP thread simultaneously per Module without sharing the Execution Units, when Hawell can do the same but each thread (INT and FP) fight for the same resources(ALUs etc) through Hyper-Threading.

 

[el etro]

The craving for the server space is AMDs evil nemesis. And its what destroys the company.

So much true. AMD should target HEDT, semicustom and embedded business.

 

[Fjodor2001]

I think they do it because there's a huge profit margin on x86 server CPUs.

Also, if the same CPU core can be used for both desktop and server, then the R&D costs can be shared.

So I don't see why this should be considered AMD's nemesis.

 

[krumme]

I think they do it because there's a huge profit margin on x86 server CPUs.

Also, if the same CPU core can be used for both desktop and server, then the R&D costs can be shared.

So I don't see why this should be considered AMD's nemesis.

There is 2 ways to earn the big money on server.

Either you have highest per core perf hands down because some licensing is per core or you have best perf/watt because thats what matters for tco.
No chance amd gets perf/w leadership and if they fail on another big ass core it will be of no use elsewhere.
The rest is niche but the big money is on beeing number 1 and effectively beeing the only one. Its a monopoly race on the server side. On consumer side its a race for volume and effectively a monopoly race there too. Beeing number 2 is a pain. At all cost avoid it

 

[Fjodor2001]

There is 2 ways to earn the big money on server.

Either you have highest per core perf hands down because some licensing is per core or you have best perf/watt because thats what matters for tco.
No chance amd gets perf/w leadership and if they fail on another big ass core it will be of no use elsewhere.
The rest is niche but the big money is on beeing number 1 and effectively beeing the only one. Its a monopoly race on the server side. On consumer side its a race for volume and effectively a monopoly race there too. Beeing number 2 is a pain. At all cost avoid it

Reading this it sounds like there is only room for one company on the planet in each product segment.

But you forgot one aspect; cost. With the huge profit margin Intel has today on server CPUs, AMDs can drop the price a bit and grab market share. Sure you have to consider the total cost of ownership (including power consumption), but still.

Also, you forgot that AMD's upcoming platform can be used for both Zen and K12. The latter being ARM based and has potential to provide even better perf/watt than Intel's x86 servers. In addition it will be cheaper. So if you say that there is only room for one server CPU manufacturer in the world and that perf/watt is all that matters, then maybe that will be AMD and their K12 ARM based server CPUs. Perhaps it will be Intel that exits the server CPU segment in the future.

 

[el etro]

I think they do it because there's a huge profit margin on x86 server CPUs.

Also, if the same CPU core can be used for both desktop and server, then the R&D costs can be shared.

So I don't see why this should be considered AMD's nemesis.

Yeah, but you can optimize the arch for server or consumer workloads.

 

[ShintaiDK]

Core i7 3820 is a 300mm^2 at 32nm with huge L2+L3 cache, this one should be compared to Bulldozer quad Module SKU and not the APU (Core i7 2600K).

Only AMD got APUs.

And no, that one got 2 cores disabled. And its still faster.

Its about equal to compare it with a FX6xxx.

AMD simply won on nothing. It was a complete failure in all segments.

 

[cbn]

Only AMD got APUs.

And no, that one got 2 cores disabled. And its still faster.

Its about equal to compare it with a FX6xxx.

AMD simply won on nothing. It was a complete failure in all segments.

i7 3820 was based on a native 4C die:

http://www.anandtech.com/show/5276/intel-core-i7-3820-review-285-quadcore-sandy-bridge-e

 

[Phynaz]

But you forgot one aspect; cost. With the huge profit margin Intel has today on server CPUs, AMDs can drop the price a bit and grab market share. Sure you have to consider the total cost of ownership (including power consumption), but still.

How well has this strategy worked for AMD in the past decade?

 

[cbn]

Two things I am wondering about:

1. Will SKUs be offered that are multi-socket capable?

2. Will SKUs be offered with variable GPU sizes? One common usage I can see for something with 16 cores or more is video editing. However according to testing at Puget Systems there can be diminishing returns for increasing video card size depending on how many GPU accelerated effects are on the timeline: http://www.pugetsystems.com/labs/articles/Adobe-Premiere-Pro-CS6-GPU-Acceleration-162/#results , http://www.pugetsystems.com/labs/articles/Adobe-Premiere-Pro-CC-Professional-GPU-Acceleration-502/

 

[krumme]

Reading this it sounds like there is only room for one company on the planet in each product segment.

But you forgot one aspect; cost. With the huge profit margin Intel has today on server CPUs, AMDs can drop the price a bit and grab market share. Sure you have to consider the total cost of ownership (including power consumption), but still.

Also, you forgot that AMD's upcoming platform can be used for both Zen and K12. The latter being ARM based and has potential to provide even better perf/watt than Intel's x86 servers. In addition it will be cheaper. So if you say that there is only room for one server CPU manufacturer in the world and that perf/watt is all that matters, then maybe that will be AMD and their K12 ARM based server CPUs. Perhaps it will be Intel that exits the server CPU segment in the future.

You dont understand servermarket. Compared to some software license cost cpu cost is nada. And eg total cooling cost dwarfs direct cpu cost. There is a natural monopoly situation. Look at the current situation.
Amd is lightyears after Intel and will just get the leftovers.
And if you bet amd can beat Intel here just buy stocks. We quite agree if they succeed they wil sky rocket. But hey perhaps that whats keep them from hitting below 1.

 

[Fjodor2001]

How well has this strategy worked for AMD in the past decade?

Pretty well with their earlier Opterons. Here's a blast from the past:

"AMD on Monday [2006] said it increased its share of the x86 server processor market to 25.9 percent, a number confirmed by Mercury Research's Dean McCarron, who tracks market share figures. [...] AMD has been picking away at Intel's server market share for several years based on the superior performance and power consumption of its Opteron processor."

Not so much lately though, but that's due to the Bulldozer uArch that is affecting desktop too. But now we're talking about AMD's next gen Zen and K12 uArch, so we'll have to wait and see. However as the article indicated, AMD has been successful in the server CPU segment before, so it can happen again.

 

[krumme]

One person IT cost per year is often around 5-15.000 and typically aprox 10.000 a year including wasted time waiting for systems to respond, downtime ...
Server hardware cost is marginal here. And cpu cost is marginal of the marginal. Trying to sell something here because its 20% cheaper makes no sense for anyone. Its about TOTAL cost.

 

[Fjodor2001]

You dont understand servermarket. Compared to some software license cost cpu cost is nada. And eg total cooling cost dwarfs direct cpu cost. There is a natural monopoly situation. Look at the current situation.
Amd is lightyears after Intel and will just get the leftovers.
And if you bet amd can beat Intel here just buy stocks. We quite agree if they succeed they wil sky rocket. But hey perhaps that whats keep them from hitting below 1.

One person IT cost per year is often around 5-15.000 and typically aprox 10.000 a year including wasted time waiting for systems to respond, downtime ...
Server hardware cost is marginal here. And cpu cost is marginal of the marginal. Trying to sell something here because its 20% cheaper makes no sense for anyone. Its about TOTAL cost.

Then how did AMD manage to get 25.9% server CPU market share in the past...?

 

[ShintaiDK]

Then how did AMD manage to get 25.9% server CPU market share in the past...?

K8. Faster, better performance/watt etc.

It was a one hit the company could never follow up on.

 

[DrMrLordX]

By having the best performance/watt in the commodity x86 market, that's how. Opterons were the best until Intel finally ditched Netburst in the server room.

edit: damn Shintai beat me to it.

 

[Fjodor2001]

K8. Faster, better performance/watt etc.

It was a one hit the company could never follow up on.

If it happened once, it can happen again. We'll not know for sure until K12 & Zen have been released. Until then it's just speculation.

 

[krumme]

Then how did AMD manage to get 26% market share in the past?

The DEC. Ibm. P4 song.

The intersting is why didnt they get higher marketshare when p4 was so off and their product was so superior?

For a few years intel just paid eg Dell not to use amd. Plain and simple. Money.
That just shows how long a way it is.

Even if Intel 2 consistently good designer teams got sick at the same time or all died by a meteor and process development got hit by a bug. Then a string of telephonecalls and loads of cash can make up for it.

The problem is if amd have anything competitive how will Intel respond? Will they lower price of that segment? Otellini for sure would do that. Imo thats the most effective way to handle it. So if amd get some server parts that is interesting Intel will just rebrand something expensive into that segment and kill it. Loss for Intel but relatively bigger loss for amd.

 

[Fjodor2001]

K8. Faster, better performance/watt etc.

It was a one hit the company could never follow up on.

Ok, so then let's turn it around: Why did Intel still have 75% market share if AMD back then? I thought the server CPU segment was a natural monopoly with only room for one player.

 

[Phynaz]

Pretty well with their earlier Opterons. Here's a blast from the past:

"AMD on Monday [2006] said it increased its share of the x86 server processor market to 25.9 percent, a number confirmed by Mercury Research's Dean McCarron, who tracks market share figures. [...] AMD has been picking away at Intel's server market share for several years based on the superior performance and power consumption of its Opteron processor."

Not so much lately though, but that's due to the Bulldozer uArch that is affecting desktop too. But now we're talking about AMD's next gen Zen and K12 uArch, so we'll have to wait and see. However as the article indicated, AMD has been successful in the server CPU segment before, so it can happen again.

At the time of this article AMD was not the low cost option. You even bolded where the article says it was due to performance, not price.

In other words despite AMD lowering prices they have continued to lose market share. Their server market share is one tenth of what it was a decade ago.

This shows a couple of things:

The number one rule of business - watch your margins.
The number 2 rule of sales - he who lives by price dies by price.

I know you want to try to discredit what I write, but jeez, at least find an reference that supports what you are arguing about. Talk about self pownage.

 

[Fjodor2001]

The problem is if amd have anything competitive how will Intel respond? Will they lower price of that segment? Otellini for sure would do that. Imo thats the most effective way to handle it. So if amd get some server parts that is interesting Intel will just rebrand something expensive into that segment and kill it. Loss for Intel but relatively bigger loss for amd.

Great strategy. Why don't all companies do that in all product segments? Just lower the price and kill the competition.

There only ought to be one company in each product segment. Strange reality doesn't look that way. Must be something wrong with it.

 

[krumme]

Intel actually have a big professional organisation with eg a far superior salesforce. That accounts for something. And it goes for eg packaging support marketing whatever. Everything is something you get when you buy from Intel.