AMD Raven Ridge 'Zen APU' Thread

[sirmo]

I'm not sure GF is going to abandon FinFETS. 12FDX is probably going to be low-power stuff only.

Definitely not.. AMD used some of the share dilution proceeds to extend the agreement for dual sourcing 14nm FF with Samsung and the 7nm FinFet future. I really don't see them going with anything else.

 

[Sweepr]

Not even a quarter/semester defined for the Raven Ridge's launch yet.

 

[itsmydamnation]

Given that everyone expects H2 for Raven then maybe there is a slight chance of hitting a Late Q2

 

[Arachnotronic]

Not even a quarter/semester defined for the Raven Ridge's launch yet.

Notice the different language used for the APU versus the CPUs. Summit/Naples are "on track" with specific quarters given. The Zen APU is given language such as "introduce" (not "volume shipments") and "planning" (not "on track"), and "in 2017" (no specific timeframe given).

AMD's top priority with Zen was clearly the server die/platforms, which is why we are seeing concrete schedules/time-frames for those, while not so much with the APU.

When you are resource constrained, you need to pick and choose your battles. AMD has made its bet, now we will see over the next year or so if it pays off.

 

[unseenmorbidity]

AMD's top priority with Zen was clearly the server die/platforms, which is why we are seeing concrete schedules/time-frames for those, while not so much with the APU.
.

I am fairly certain the CEO said they planned on releasing the consumer FX CPU's first.

 

[Arachnotronic]

I am fairly certain the CEO said they planned on releasing the consumer FX CPU's first.

Which are based on the Zeppelin die which is used for the server parts.

 

[Atari2600]

I am fairly certain the CEO said they planned on releasing the consumer FX CPU's first.

It buys them more time for server/workstation verification* on pretty much the same hardware.

*the depth and coverage of such verification is a world apart from consumer grade.


It also explains why there is so little info of Zen performance; the traditional vendors that would have it for fairly wide verification testing do not yet have it (to my knowledge). These would be a historical source of informed pub-talk.

 

[raghu78]

Notice the different language used for the APU versus the CPUs. Summit/Naples are "on track" with specific quarters given. The Zen APU is given language such as "introduce" (not "volume shipments") and "planning" (not "on track"), and "in 2017" (no specific timeframe given).

AMD's top priority with Zen was clearly the server die/platforms, which is why we are seeing concrete schedules/time-frames for those, while not so much with the APU.

When you are resource constrained, you need to pick and choose your battles. AMD has made its bet, now we will see over the next year or so if it pays off.

Its a well known fact that AMD is resource constrained as they are still bleeding cash. Therefore its not a surprise that AMD prioritized consumer/enthusiast desktop and server CPUs over APUs. AMD product margins are definitely going to be better on these CPUs than APUs. I think Zen APUs are likely to launch in Q4 2017 (with the real volume coming in Q1 2018).

 

[Arachnotronic]

Its a well known fact that AMD is resource constrained as they are still bleeding cash. Therefore its not a surprise that AMD prioritized consumer/enthusiast desktop and server CPUs over APUs. AMD product margins are definitely going to be better on these CPUs than APUs. I think Zen APUs are likely to launch in Q4 2017 (with the real volume coming in Q1 2018).

Sounds right to me, Zen APUs will have to go up against CFL, CNL-Y/U, and KBL-R U.

 

[Abwx]

Sounds right to me, Zen APUs will have to go up against CFL, CNL-Y/U, and KBL-R U.

Most of the mainstream PC market is squatted by laptops and they have a strong enough offering with Bristol Ridge for this segment, hence they can calmly wait for Intel to launch their next gens before releasing a replacement, if the schedule is the same as the previous gens they should release it roughly one year after BR, wich get us to Q3 2017.

 

[krumme]

Notice the different language used for the APU versus the CPUs. Summit/Naples are "on track" with specific quarters given. The Zen APU is given language such as "introduce" (not "volume shipments") and "planning" (not "on track"), and "in 2017" (no specific timeframe given).

AMD's top priority with Zen was clearly the server die/platforms, which is why we are seeing concrete schedules/time-frames for those, while not so much with the APU.

When you are resource constrained, you need to pick and choose your battles. AMD has made its bet, now we will see over the next year or so if it pays off.

Good catch. I think its simply a question of beeing sure to have enough capacity for servers if it takes off big time.
Frankly i dont really understand why they are in such a hurry with such a low margin apu product. Granted there migh be some good margins in high end laptops but that could be covered by binned 8c 16t and vega for the high end. That would be really good for the brand also. And extremely needed to say the least.
As usual i think amd makes their portfolio to wide for no good business reason. Stupid engineers....lol

 

[TheELF]

Frankly i dont really understand why they are in such a hurry with such a low margin apu product.

ZEN apu will be the replacement for the consoles,it's the only market where they will have ensured high volume sales...

 

[krumme]

ZEN apu will be the replacement for the consoles,it's the only market where they will have ensured high volume sales...

I doubt consoles will be 14nm but instead straight to gf 7nm or tsmc 10/7nm. How they will fit a zen in that mm2 envelope as jaguar 28nm remains to be seen but perhaps a slimmed down l3 or even sans l3 will do the job? They cant go 4c because then the jump from 8c jag is to little for 6 years. And we have a lot of code already running 6 threads. Makes sense to use 8c x86 cores again for lessening software dev cost and then keep the 128bit double fpu in zen. So the codebase remains similar.

If dev cost is less then using more mm2 on cpu side is possible its kind of distributing cost. But with consoles running 8c zen it will put a lot of stress on pc ports. Then all the crap ports will have to go. But thats about time to.

 

[TheELF]

I see you still believe that ZEN will have intel HT...might still be but is very doubtful.
I believe that Zscaling will be splitting cores up into two so 4c/8t will actually run just as 4m/8t jaguar,a 4c zen at 1-1,5Ghz will be a huge boost and run at very low power ratings (for the cpu part) compared to jaguar without changing anything to how games run.

 

[inf64]

I see you still believe that ZEN will have intel HT...might still be but is very doubtful.
I believe that Zscaling will be splitting cores up into two so 4c/8t will actually run just as 4m/8t jaguar,a 4c zen at 1-1,5Ghz will be a huge boost and run at very low power ratings (for the cpu part) compared to jaguar without changing anything to how games run.

"ZScaling" was just a name that came from a rumor, it never came up in any slide deck or presentation. Zen will have SMT that is very much akin to what intel is using in their CPU lineup since Haswell and what AMD used in Bulldozer's FP part of the core. There are murmurings that it is bringing higher performance jump in MT code than intel's implementation but those are just unconfirmed rumors. Looking at what AMD presented at HotChips it is a traditional SMT implementation with various levels/ways of sharing of parts of the core depending on the complexity of shared parts. So there is no magic pixie dust in there and therefore performance benefits should be in line with what intel's cores see from their own implementation of SMT( aka HT).

 

[TheELF]

Even if it is HT a 4c/8t ZEN will still function just the same way as a 8 core CPU so it's still a very viable possibility for the jaguar replacement.

 

[jpiniero]

Even if it is HT a 4c/8t ZEN will still function just the same way as a 8 core CPU so it's still a very viable possibility for the jaguar replacement.

It's possible that Scorpio is a 4C8T Zen die plus a discrete Vega GPU. IOW more or less a PC.

 

[Sweepr]

They clearly mentioned 8-core design for Scorpio, so I wouldn't count on it.

 

[Arachnotronic]

It's possible that Scorpio is a 4C8T Zen die plus a discrete Vega GPU. IOW more or less a PC.

Uh, no.

 

[krumme]

Even if it is HT a 4c/8t ZEN will still function just the same way as a 8 core CPU so it's still a very viable possibility for the jaguar replacement.

Its way to expensive for the power it gives. 8c Jaguar gave tremendous power for the invested mm2.

An 2 wide arm a73 is 0.65mm2 incl. L1 on 10nm.

We are at a point where going for more ipc makes mm2 explode. Using more cores is far cheaper.

Clearly HT is going to be okey at zen given its wide design and ressources available. It makes no sense to have that without at least reasonably HT. Probably it will be just fine.

But a new console gen needs to be faster than what 4c 8t zen can bring imo. And why not use new engines can use more cores and have been for 6 years? If i was sony /ms i would use zen to get an edge to pc gaming. 8c 8t zen is an insane amount of power it "just" got an mm2 problem even at 7nm unless devs pays more so to speak.

 

[Arachnotronic]

Its way to expensive for the power it gives. 8c Jaguar gave tremendous power for the invested mm2.

An 2 wide arm a73 is 0.65mm2 incl. L1 on 10nm.

We are at a point where going for more ipc makes mm2 explode. Using more cores is far cheaper.

Clearly HT is going to be okey at zen given its wide design and ressources available. It makes no sense to have that without at least reasonably HT. Probably it will be just fine.

But a new console gen needs to be faster than what 4c 8t zen can bring imo. And why not use new engines can use more cores and have been for 6 years? If i was sony /ms i would use zen to get an edge to pc gaming. 8c 8t zen is an insane amount of power it "just" got an mm2 problem.

IMO the consoles might be better served with future ARM cores rather than a Zen derivative.

 

[krumme]

IMO the consoles might be better served with future ARM cores rather than a Zen derivative.

Perhaps. But now we are at x86. It also makes porting cheaper. No more Sony cell hell. I think focus is in keeping dev cost low and staying at x86 does that. I think developing new engines becomes more and more expensive and that will decide focus.

Another issue is simply if mobile phones is gaining track so much consoles and pc takes more hits? Using my note 5 with a vr gear its incredible what a portable device can do.

 

[krumme]

If new engines can all master the concept of no main thread then surely there is perspective above 8c and then arm solutions become more competitive. But what comming games is without a main thread? I think its 9 years out.

 

[TheELF]

Clearly HT is going to be okey at zen given its wide design and ressources available. It makes no sense to have that without at least reasonably HT. Probably it will be just fine.

Well that's why I believe that AMDs SMT will be more like AMDs CMT and not like HT.
Sandy/ivy had HT on just 5 ports, zen has 10 that's too much even for HT makes sense though if you can cut up the core into two separate parts.

 

[NostaSeronx]

Sandy/ivy had HT on just 5 ports

There are five matrix ports and one discrete port(Store Data) versus eleven discrete ports in Sandy/Ivy. To be exact there are eleven units in Sandy Bridge tied to six ports. (Port 0(1 port) to Port 5(6 ports))

It does take an extra cycles for the same port, but in most cases there is a wait period. 1 cycle latency, 5 cycle latency, 3 cycle latency, 4 cycle latency, etc. So really, ignore the ports just use the generalized units.

Technically, port 5 has three units because of AVX. So, that would up the count to 12 units to Zen's 10 units. +2 units are discrete VI/FP Misc unit and Store Data.

Zen = ALU, ALU, ALU, ALU, AGU, AGU, FPU, FPU, FPU, FPU (+3 VI ALUs but share the same exact units as FPUs, so -3)
SB = ALU, ALU, ALU, VI ALU, VI ALU, VI ALU, FPU, FPU, FPU, AGU, AGU, AGU (AVX is bridged so negating VI ALUs would also negate Zen's FPUs to only two FPUs; 8/9 mix versus 10/12 mix)