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

[Fjodor2001]

How likely is it that Zen will be available within the first 3 quarters of 2016?

Or, how long would you wait in 2016 for Zen to be available?

Is it absolutely crucial to know if it's 2016Q3 or Q4?

 

[Fjodor2001]

Intel's HEDT offerings could be better? They could have more cores, better performance per core, or even better performance/$ -- you really don't know what Intel's plans are for the 2016 HEDT platform, so it's hard to make a hypothetical 2016 buying decision right now.

Also, I'd wager that if you want to build your system between Q1 2016 and Q3 2016, your best choice will be Intel.

Problem is that 8 core Intel HEDT is $1000, so nearly no gamers buy it. I'd expect AMD Zen 8 core to be $300-400 or so. They're saving die area by not having any iGPU, and they also need to price it competitively to sell. That is of course just a guesstimate, but still.

 

[Fjodor2001]

Maybe because AMD is not likely to reach "top" levels of performance as far as gaming is concerned?

IB/Haswell level of performance and 8 cores is what most in this forum has been estimating. That puts it into $1000 8 core Haswell-E territory. Is that not considered "top" performance for you?

 

[Valantar]

Exactly. We know Zen will be high throughput with better IPC and better efficiency then BD. But we don't know that it will be a good gaming CPU.
This is a product targetted at servers, after all.

Just to clarify, BD is ... Bulldozer?

And yeah, it's aimed at HPC and other high margin markets, but it would be very, very odd for AMD not to simultaneously launch a "simplified" version for the consumer market. And besides, isn't the FX brand a purely consumer oriented brand? Not that AMD really has anything else these days, but I'd still expect them to whip out the Opteron name at some point if these chips were meant solely for professional use. I'm expecting something along the lines of Intel's i7 ~= Xeon strategy, unless AMD want to play the value card for all it's worth and sell FX chips at consumer-level prices for professional use as well.

 

[Fjodor2001]

You're still referencing those fake slides.

Where did AMD say how many cores this Zen FX would have?

Would it make any sense to release a 4 core Zen with no iGPU, if they're also targeting the server market?

http://hexus.net/tech/news/cpu/82909-amd-outlines-product-roadmap-financial-analyst-day/

"AMD didn't specify exact core counts but Zen will be available in up to 8-core, 16-thread designs, at the very minimum, with the potential for higher core counts to target datacentre markets."

 

[Valantar]

Would it make any sense to realease a 4 core Zen with no iGPU, if they're also targeting the server market?

http://hexus.net/tech/news/cpu/82909-amd-outlines-product-roadmap-financial-analyst-day/

"AMD didn't specify exact core counts but Zen will be available in up to 8-core, 16-thread designs, at the very minimum, with the potential for higher core counts to target datacentre markets."

He's not saying there won't be an 8-core Zen FX, just that there might also be (far) higher core counts. See the exchange between me and erunion above.

 

[Valantar]

By the way, does anyone have any decent numbers on how Zen's 40% IPC increase might compare to Haswell/Broadwell/Skylake? I know pure IPC is near impossible to benchmark, but we know AMD these days is way behind Intel. Will a 40% increase be enough to close the gap (or get close enough that simulatneous multithreading might put them ahead)?

 

[AtenRa]

Ok, fair enough. But assuming Zen will be in IB performance range, have 8 cores, and priced around $300-400 or so. What's your take?

Actually as i have said before, with DX-12 games you dont really need that much CPU performance. I really dont know what to say as of now because information is scarce about the performance and price.
But since ZEN will be a high performance core, i do consider them for a Gaming PC in 2016 onwards. If they have a competitive SKU at a right price i already consider a ZEN SKU to retire my Core i7 3770K.

 

[Fjodor2001]

He's not saying there won't be an 8-core Zen FX, just that there might also be (far) higher core counts. See the exchange between me and erunion above.

Ok, I agree with that. I'd say an 8 core variant makes sense in the performance consumer segment, and higher core counts (16-32?) in the server segment. Would you agree?

Anyway, I have hard time seeing that more than 8 cores will be the lowest consumer SKU, and also a hard time seeing how less than 8 cores will be the highest consumer SKU. I.e. there ought to be at least one SKU with 8 cores, then there might be others that have less or more cores.

 

[erunion]

Just to clarify, BD is ... Bulldozer?

And yeah, it's aimed at HPC and other high margin markets, but it would be very, very odd for AMD not to simultaneously launch a "simplified" version for the consumer market. And besides, isn't the FX brand a purely consumer oriented brand? Not that AMD really has anything else these days, but I'd still expect them to whip out the Opteron name at some point if these chips were meant solely for professional use. I'm expecting something along the lines of Intel's i7 ~= Xeon strategy, unless AMD want to play the value card for all it's worth and sell FX chips at consumer-level prices for professional use as well.

Yes, BD is bulldozer. Though I was using the term to include the entire construction core lineup. (piledriver, steamroller, etc)

Intel and AMD both sell the same die under different brands to both the consumer and enterprise markets. Bulldozer for servers was called Opteron, and bulldozer for consumers was FX. I predict the same for zen. The server Zen cpu will be branded FX and sold to consumers.

 

[coercitiv]

IB/Haswell level of performance and 8 cores is what most in this forum has been estimating. That puts it into $1000 8 core Haswell-E territory. Is that not considered "top" performance for you?

Most of this forum guesstimated SB level IPC with clocks somewhere under 4Ghz. (and this was the optimistic scenario)

 

[Valantar]

Problem is that 8 core Intel HEDT is $1000, so nearly no gamers buy it. I'd expect AMD Zen 8 core to be $300-400 or so. They're saving die area by not having any iGPU, and they also need to price it competitively to sell. That is of course just a guesstimate, but still.

No. The problem is that 8 core Intel HEDT doesn't actually gain you anything in >99% of all gaming scenarios, which is why (even as mind-numbingly MOAR POWAH-minded as gamers tend to be) they steer away from these chips. Or at least, the price in combination with the realization that they might gain 10% performance IF they run tri-SLI/CF. The fact is that until DX12 comes along (and presumably a good while after that as well), 8 threads will be PLENTY for gaming. Clock speeds and IPC matter more.

 

[Fjodor2001]

Actually as i have said before, with DX-12 games you dont really need that much CPU performance.

Ok. But don't you see that changing with ever more powerful GPUs? Doesn't the CPU performance have to keep up with that?

The 14 nm GPUs with HBM are right around the corner. They will bump up the GPU performance quite a lot...

 

[AtenRa]

Yes, BD is bulldozer. Though I was using the term to include the entire construction core lineup. (piledriver, steamroller, etc)

Intel and AMD both sell the same die under different brands to both the consumer and enterprise markets. Bulldozer for servers was called Opteron, and bulldozer for consumers was FX. I predict the same for zen. The server Zen cpu will be branded FX and sold to consumers.

yeap, but this time i dont expect a 4 core 8 thread ZEN FX since that will be made on the 8th Gen APUs in 2017 on the same socket.
So im expecting a 2-4 core 2-4-8 Threads APUs and 6-8 core 12-16 thread FX SKUs.

 

[Fjodor2001]

Clock speeds and IPC matter more.

So you're saying some games are CPU bottlenecked after all?

 

[Phynaz]

IB/Haswell level of performance and 8 cores is what most in this forum has been estimating.

More like wishing for. Estimates have at least some basis in fact. Since AMD has released zero facts about performance there can be no estimates.

 

[Valantar]

So you're saying some games are CPU bottlenecked after all?

Uuuuhh? No. Or, they are if they are pairing, say an AMD FX or Ax with an R9 290x, GTX 970, 980, Titan X, or anything truly high end these days. What I'm saying is that CPUs have been quick enough for todays gaming loads for years. An i7-4790 doesn't gain you much over an i7-2600K. AMD, however, seriously needs to catch up when it comes to IPC and perf/watt as their current ~4GHz (4C4T) chips perform significantly worse than Intel's ~4GHz chips, all the while using far more power.

 

[AtenRa]

Ok. But don't you see that changing with ever more powerful GPUs? Doesn't the CPU performance have to keep up with that?

The 14 nm GPUs with HBM are right around the corner. They will bump up the GPU performance quite a lot...

2016 onward 4K monitors will become cheaper than today, those 14nm High-End GPUs will have way more power than needed for 1080p gaming. Thus those High-End GPUs will be used in UltraWide 3440x1440 and 4K monitors and games will be GPU limited than CPU limited at those resolutions.

For 1080p a middle-end 14nm 2016 $200-300 GPU will be fine in DX-12 games even with a FX8350.

 

[LTC8K6]

Is it absolutely crucial to know if it's 2016Q3 or Q4?

Crucial, no? But I'm not an early adopter. For me, late 3Q might as well be 2017.

As I said, how long are you willing to wait for parts to build your new computer "in 2016"?

To me, a new computer for 2016 is one I build early in 2016.

Also, are you willing to be an early adopter when spending your hard earned money?

I'm generally not. I usually wait and see if the early adopters report any problems in the first months.

 

[LTC8K6]

Ok. But don't you see that changing with ever more powerful GPUs? Doesn't the CPU performance have to keep up with that?

The 14 nm GPUs with HBM are right around the corner. They will bump up the GPU performance quite a lot...

Yes, it could well be a mistake to think your current CPU will be able to keep up with the coming GPUs.

 

[Fjodor2001]

Quite a lot of different opinions here on what kind of CPU performance will be needed for a top gaming computer in 2016. Anything from a 2500K or FX8350 to an 8 core Intel HEDT or Zen... that's a wide range...

 

[Shehriazad]

By the way, does anyone have any decent numbers on how Zen's 40% IPC increase might compare to Haswell/Broadwell/Skylake? I know pure IPC is near impossible to benchmark, but we know AMD these days is way behind Intel. Will a 40% increase be enough to close the gap (or get close enough that simulatneous multithreading might put them ahead)?

Take this with a HUGE grain of salt what I'm about to say now(Since I'm not all that savvy, pros can feel free to correct me):

+40% compared to excavator
Excavator should be anywhere between 10-15% above FX in IPC.

Let's say 50% total vs FX Vishera.

I'd say non biased benchmarks and real world tools put Ivy-E 50% above FX Vishera in IPC.

This means it'd even out clock for clock with Ivy-E. Then there's Haswell-E...I'm rather inconclusive about the results of this chip since it varies so much and DDR4 might play a role as well.

But let's say after possible DDR4 benefits (which are Zen as well) there is 10% gains in performance.

AMD is notorious for punching up their base clocks as much as possible. Seeing Zens at 4 Ghz base is a possibility.

This means non-OC Zen could also beat Haswell-E (Has-E's highest base is 3.5 right now, 3.85 on Zen would mean it's even again)

So Zen would actually compete and lose against whatever Intel brings out after the current Haswell-E (Broadwell-E?). But then there's pricing and core count.
IF AMD dishes out a 4Ghz base, 8core Zen that beats Has-E and Ivy-E due to clocks and costs 400$ as opposed to some 1000$ 2016 Intel CPU that delivers 10% more....then you got a product that might actually get AMD to gain a market share that approaches 10% again...yay relevancy.



Realistically speaking though...this is all just a bunch of wild guesses.

 

[Ajay]

Too many ifs to say. A 40% IPC gain in a vacuum means nothing. The architectural details will limit issue rate, cache performance and core execution performance (unless Zen started b/4 Keller arrived, enough time will not have passed for hand tuned critical path optimization).

Then there is the who issue of the GFs implementation of Samsung's 14FF process. Are they able to tune it for higher clock rates and even if so, will they put in the extra time and money to do so for AMD. So we have no idea what the max clock rate may be.

On top of that the implementation of the micro-architecture on the 14FF will limit raw performance to something well below that achievable by a simple test sled (like an tiny old ARM 7 core). This can be improved with respins, which are fixes to xtor layout that require new masks. These masks appear to be ever increasing in complexity and cost - so AMD would likely be loathe to do more than one respin b/4 release). Many companies today just wait till a shrink or major architectural improvement (like Zen+) to make new masks and rely heavily on computerized simulation and verification of their designs before 'tape-out' (ordering the masks and then ordering a test run of wafers from their fabrication partner).

 

[DrMrLordX]

Excavator should be anywhere between 10-15% above FX in IPC.

You're shooting a bit low there. Kaveri can already trump Vishera by 15% or more depending on the workload, at least on a per-module basis at the same clockspeed. AMD's own literature shows Excavator being 5% faster than that.

 

[inf64]

I'm sticking with my prediction for 8C/16T Zen parts (made it before FAD held on 6th, I assumed 30% higher IPC vs XV core, was bashed by everyone from blue camp ). I adjusted the ST a bit up to correlate with new information from FAD.
ST vs 8350 : ~42-45% higher performance, stock vs stock; Zen at 3.5Ghz Base-3.8Ghz Turbo 95W; ST IPC vs PD with L3 => 50-60% higher
MT vs 8350 : ~2x-2.1x higher performance, stock vs stock; Zen at 3.5Ghz Base-3.8Ghz Turbo,95W