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

[cytg111]

40% increase in IPC is great, but I bet it's marketing speak. It's probably 40% on one synthetic benchmark they manipulated.

Don't get me wrong, I hope I'm wrong. An AMD CPU with Ivy Bridge or even Haswell IPC with 8 cores would attract a lot of attention, but I'm pessimistic But let's hope so because it will probably force Intel to release a 6 core mainstream CPU.

Maybe, yet there is precedence for the opposite, look back at the mantle introduction, their estimates very pretty conservative, no where near 'optimal' levels.
Anyway, at 40% you gotta wonder what kind of rabbit they expect to pull from that hat that they cannot pull today. I suppose the two next-biggest unknownes is clocks and fabbing, will it be a performance node /power node /?

 

[inf64]

Seems like I was right there in the ballpark when I said 50% better IPC than Piledriver . AMD's 40% over XV, which is probably an average from commonly used set of benchmarks, is even higher since SR brought ~7.5% higher IPC and XV supposedly brings 5% over SR (according to AMD).

If it clocks mid to high 3Ghz range (3.5 base-3.8Ghz ST Turbo) then 8C/16T Zen could end up performing like ~ 2x8350 in MT and ~1.42x 8350 in ST, both at stock.

I'm a bit worried the APU lineup is pushed for 2017, AMD could use Zen APUs since those would be i5/i7 competitors on x86 side and would be way better in iGPU. Let's hope we get HBM supporting APUs in 2017.

 

[erunion]

. I honestly am blown away by this roadmap, and not in a good way.

I was too.

That roadmap seems like a disaster. 2016 is supposed to be the turn around year. Almost everything they talked about in 2014 is either cancelled or pushed back to 2017.

What will the reaction be if AMD tries to pull the "wait for 2017" card next year?

 

[Abwx]

I wonder if Zen is also in simulation as well, or if they've got actual silicon in hand?

They need real silicon no later than this quarter given the time needed for validation in both client and servers products, 15 months is not too much for such complexe product, for the record Carrizo silicon was functional as early as feb 2014 according to some Sisoftware submissions, and that is a chip that has lot in common with Kaveri.

 

[inf64]

Yeah they got Zen eng samples by now for sure. Probably an early version running at low clocks but they can use it to extrapolate perf. numbers easily.

 

[Arachnotronic]

Some back of the envelope math...

A10-7850K @ 4.0GHz (max single-core turbo) hits 2458 in GB3. An i5-4590 does about 3401, with max single core frequency at 3.7GHz.

If we normalize the Haswell to 4GHz, this should imply single-core score of approximately 3676. 40% If we assume a 5% IPC improvement from SR to XV, and then add another 40% on top of that, then we're at 3613 -- or roughly on par with Haswell in Geekbench 3, clock-for-clock.

Intel should still have a perf/clock advantage with Skylake, which is ~12.3% faster in GB3 per clock than Haswell.

It'll really come down to the kinds of clocks AMD can deliver with these parts and its relative performance-per-watt.

 

[ShintaiDK]

Lets not hope the IPC is based with the usage of SMT. Or some extreme case vs XV.

But we have to wait 1½ year to see the result.

 

[Dresdenboy]

Lets not hope the IPC is based with the usage of SMT. Or some extreme case vs XV.

But we have to wait 1½ year to see the result.

We still don't have clock frequencies, power efficiency, information about avg/max. IPC (max. would be useless anyway), normalization base. So there are enough degrees of freedom in the game.

But a SMT based IPC would be compared against what - CMT? Then it would be even better than 1T vs. 1T w/o SMT/CMT-penalties.

BTW this SMT implementation might take care of thread priorities...

 

[cytg111]

I was too.

That roadmap seems like a disaster. 2016 is supposed to be the turn around year. Almost everything they talked about in 2014 is either cancelled or pushed back to 2017.

What will the reaction be if AMD tries to pull the "wait for 2017" card next year?

I think it is pretty clear that they are betting the farm on this one, make it or break it.

 

[Abwx]

Some back of the envelope math...

A10-7850K @ 4.0GHz (max single-core turbo) hits 2458 in GB3. An i5-4590 does about 3401, with max single core frequency at 3.7GHz.

If we normalize the Haswell to 4GHz, this should imply single-core score of approximately 3676. 40% If we assume a 5% IPC improvement from SR to XV, and then add another 40% on top of that, then we're at 3613 -- or roughly on par with Haswell in Geekbench 3, clock-for-clock.

Intel should still have a perf/clock advantage with Skylake, which is ~12.3% faster in GB3 per clock than Haswell.

It'll really come down to the kinds of clocks AMD can deliver with these parts and its relative performance-per-watt.

I would say, based on a lot of integer based benches, that 40% in Integer is a best case scenario for Haswell, the Geekbench 40% include the FP perf where Haswell has generaly a significantly bigger advantage than in Integer code.

I wouldnt count FP in the numbers as this is a too favourable case, most of Intel s alleged IPC gains are in FP not in Integer where progress has been limited, and increasing throughput in FP will be much more easier than in Integer for AMD, so i wouldnt expect 40% in FP but more something like 60% better perf/Mhz, wich is not IPC as i already stated but throughput.

Indeed it always amaze me that people are using FP benches to compare CPUs that will be used at 99.999% with integer code based applications.

 

[Phynaz]

I would rather think that your post is not probably but most surely what one can call deffamation...

The rest of us call it a fact.

 

[Abwx]

The rest of us call it a fact.

You are perfectly right since the "us" seems to apply only to the usual Intel fanatics...

Other than that, since how much time didnt you manage to post something that has the slightest technical value..?.

Must be a record on this forum for the continual sterility technicaly wise.

That roadmap seems like a disaster. 2016 is supposed to be the turn around year. Almost everything they talked about in 2014 is either cancelled or pushed back to 2017.

What will the reaction be if AMD tries to pull the "wait for 2017" card next year?

Seriously.?.Getting back to more profitable market is a disaster..?.

Disaster could be eventualy for the competition stock holders, but for AMD..?.

 

[imported_ats]

How many people would be interested in a SKL-E with 12 cores clocked at 2.5Ghz and costing $1,000?

likely less than a 10c clocked at 2.8, which would be less than an 8c clocked at 3+, which would be less than a 6c clocked at 3.5+ which would be less than a 4c closed at 4+.

 

[frozentundra123456]

40% increase in IPC is great, but I bet it's marketing speak. It's probably 40% on one synthetic benchmark they manipulated.

Don't get me wrong, I hope I'm wrong. An AMD CPU with Ivy Bridge or even Haswell IPC with 8 cores would attract a lot of attention, but I'm pessimistic But let's hope so because it will probably force Intel to release a 6 core mainstream CPU.

According to anand's article there are no benchmarks yet:

"We do not have performance estimates for Zen, but from an architectural standpoint AMD is making it clear that they are shooting for much higher IPC with Zen than Carrizo/Excavator, touting a 40% increase in IPC."

Just like they were shooting for higher clocks and better ipc with Bulldozer. So we will see. However, one only needs to look at the thread title and predictions in this thread and compare them to what is on the actual roadmap to see that a healthy dose of skepticism is warranted.

 

[CHADBOGA]

According to anand's article there are no benchmarks yet:

"We do not have performance estimates for Zen, but from an architectural standpoint AMD is making it clear that they are shooting for much higher IPC with Zen than Carrizo/Excavator, touting a 40% increase in IPC."

Just like they were shooting for higher clocks and better ipc with Bulldozer. So we will see. However, one only needs to look at the thread title and predictions in this thread and compare them to what is on the actual roadmap to see that a healthy dose of skepticism is warranted.

We'll have a better idea on how Zen will perform if we later hear that Charlie DeMargerine is dancing in the aisles.:awe:

 

[Shehriazad]

So +40% IPC... and other performance metrics like new low latency cache design and FinFET process or SMT are not yet taken into account.


If they manage to push this ranging from 3.5Ghz to 4 Ghz base...it's a rock solid CPU that will at least be able to compete with whatever Intel has in store in 2016.

I obviously won't claim that it can easily beat Intel...because that would be way too ballsy of a claim right now.

But in the price + performance Area we could have a new winner.

Now if it actually was 4Ghz base + 40% IPC compared to Excavator + more boost from FinFET + low latency Cache (the absolute best case scenario) then it would be the ultimate "comeback" AMD will deliver ever since they started making CPUs.
But their own slides are pessimistic toward CPU marketshare...that makes me think that it won't even end up being close to the best case scenario.


Either way...now I can sit back and "not regret" my choice of waiting until 2016 until I upgrade all of my platforms (to see if AMD has an ace or if all my PCs are going Intel only)

 

[Arachnotronic]

So +40% IPC... and other performance metrics like new low latency cache design and FinFET process or SMT are not yet taken into account.

Huh? "low latency cache design" absolutely impacts the number of instructions that can actually be executed per clock. SMT impacts multithreaded performance, and FinFETs impact power consumption and maximum achievable clocks.

 

[JDG1980]

It's encouraging to know that K12 is being delayed to 2017; AMD is really focusing all their resources on Zen, which is where they need to be. Hopefully K12 will be quietly axed; unless they have a specific large customer in mind, I don't see this as being a productive use of AMD's limited R&D budget.

It's even more encouraging to see that Zen will be specifically focusing on servers and HEDT and going down from there. AMD is back in the big game again. And it looks like we might see Sandy Bridge level IPC after all.

Bowing out of ultramobile makes sense. Intel's Atoms are a massive profit drain; they're literally giving them away to the tune of billions of dollars in contra-revenue per year. AMD has no reason to do this, and can't afford it. Better to focus on high-margin areas. AMD has made it clear in the analyst roadmap that they want out of the low end.

 

[Arachnotronic]

Bowing out of ultramobile makes sense. Intel's Atoms are a massive profit drain; they're literally giving them away to the tune of billions of dollars in contra-revenue per year. AMD has no reason to do this, and can't afford it. Better to focus on high-margin areas. AMD has made it clear in the analyst roadmap that they want out of the low end.

"Billions" in contra-revenue per year? It was ~$1 billion in 2014, should come down a lot in 2015, and should be gone by the second half of 2016.

 

[Abwx]

OK, listening to the Q&A...definitely coming to APUs in 2017.

Page 14 of Lisa Su s CG Segment overivew is the presentation page of "Future Innovation", that is plan for later than 2015.

The second page of this part present the guidelines for those innovations, and theses are :

Simple plateform infrastructure
"Zen" Increased CPU/APU performance
Scalable plateform design points.

Next page is the roadmap for current as well as thoses products, among others the 7th gen APU, i dont think that this APU is Excavator based, it would be a waste of time given that they ll have a fully functional Zen design at this time and that Carrizo already provide all others parts, all they have to do is to replace the cores.

 

[podspi]

That doesn't make sense. It means that all of the APUs are complete garbage. Who will want an APU when they're claiming ~40% increase for Zen, much less considering the already huge gap between Intel and AMD?

Sure it does. Performance bar for laptops is much lower than desktop/workstation. Heck, where I work people are still using laptops with C2D in them - Luckily I'm not one of them though :whiste:

As long as they can hit acceptable performance, decent power usage, and a low enough price, they will be fine. What is really hurting them now is power efficiency, and the multitude of 4200Us in mid-range machines.

Carrizo should be interesting. If the power-efficiency gains are there, it should do well (at least on a technical level). That + a node-shrink (to whichever FF process they end up with) + HBM should make a decent gaming laptop and/or HTPC. Heck even now they do alright - I was going to pick one up for that exact purpose but the sweet siren call of Steam in-house streaming + the 4200U pulled me away...

According to anand's article there are no benchmarks yet:

"We do not have performance estimates for Zen, but from an architectural standpoint AMD is making it clear that they are shooting for much higher IPC with Zen than Carrizo/Excavator, touting a 40% increase in IPC."

Just like they were shooting for higher clocks and better ipc with Bulldozer. So we will see. However, one only needs to look at the thread title and predictions in this thread and compare them to what is on the actual roadmap to see that a healthy dose of skepticism is warranted.

If there are no internal benchmarks, yikes. My guess is they haven't nailed down clockspeed yet. As we've gone over in this forum time and time again, IPC and clockspeed matter. Even assuming we take 40% as gospel across the board, we gotta know clockspeeds to know final performance.

How funny would it be if someone from AMD joined these forums and started to insisting that clockspeed wasn't going to go down?

 

[Abwx]

"Billions" in contra-revenue per year? It was ~$1 billion in 2014, should come down a lot in 2015, and should be gone by the second half of 2016.

It was already discussed, if there was 3bn of RD that would be enough to have 15 000 people in RD for BTrail, at a rate of 200K$/year cost per employee, that s impossible and truth is that all thoses losses are a little RD not recouped with most losses being subsides + free chips.

So it s rather the other way around, eventualy 1bn of internal costs and 3 bn in checks...

 

[DrMrLordX]

The only reason AMD prices its FX chips at low prices is because they won't sell at higher prices.

If AMD has an "Intel killer" as you say, then expect them to charge for it.

Actually, AMD's first time taking the performance crown from Intel was back in the days of the 1.4 ghz Tbird. Those chips were dirt cheap, faster than any PIII Intel had at the time, and were also faster than P4 1.8a in a lot of things. So no, AMD hasn't always charged a premium while selling the fastest chip available. They certainly did it with k8, though.

$350 desktop and notebook market, type of people who go down to wally world and just buy whatever is on sale with the shiniest plastic everything. you know, the current APU market.

I'm not sure that the "7th gen APU" mentioned in the slide is aimed at those $400-and-under AiO units. It would be sweet if they bumped the cat cores out of those things and replaced them all with Excavator quads with 8 CUs each. That would be a huge jump in computing power for the low-end crowd.

 

[soresu]

Something to bear in mind, that AMD are currently facing a lawsuit over Llano, and inflated projections - so any projections and roadmaps they made for the FAD will likely be on the most conservative side.

Edit: it also means if they execute better than projected for the FAD that they can say so, and make themselves look better (if unlikely to happen).

 

[frozentundra123456]

It's encouraging to know that K12 is being delayed to 2017; AMD is really focusing all their resources on Zen, which is where they need to be. Hopefully K12 will be quietly axed; unless they have a specific large customer in mind, I don't see this as being a productive use of AMD's limited R&D budget.

It's even more encouraging to see that Zen will be specifically focusing on servers and HEDT and going down from there. AMD is back in the big game again. And it looks like we might see Sandy Bridge level IPC after all.

Bowing out of ultramobile makes sense. Intel's Atoms are a massive profit drain; they're literally giving them away to the tune of billions of dollars in contra-revenue per year. AMD has no reason to do this, and can't afford it. Better to focus on high-margin areas. AMD has made it clear in the analyst roadmap that they want out of the low end.

They arent just bowing out of ultramobile. They are pretty much abandoning the entire cpu market to intel except for HEDT and servers, which is the only place I see for gpu-less Zen. Does anybody seriously think amd can keep refreshing the same APUs that are getting slaughtered already by intel, while intel brings out Broadwell, Skylake, and probably Cannonlake before any Zen apus come out, and still maintain any semblance of competitiveness. Honestly this roadmap is bad, bad news, almost worst possible case bad news, and it is still being spun into a positive.