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

[podspi]

There did not used to be "K" SKUs - they were all unlocked. If Intel cared about chip enthusiasts, they wouldn't have locked almost all of their chips.

FYI, back in the day Intel had a waterfall product portfolio, basically new products would be released, and older products would be discounted. This was alright because the differentiation between products was clockspeed and that was it (for the same line, anyway). Before they locked the processor, unscrupulous OEMs (think smaller local shops - before Dell/HP/etc owned the world) would OC the CPUs and claim to be selling something they weren't.

That's why they and AMD locked their CPUs.

 

[DrMrLordX]

MS is not abandoning the Xbox One or any possible successors. What they are abandoning is Windows Media Center. Xbox is meant to replace that.

 

[cbn]

MS is not abandoning the Xbox One or any possible successors. What they are abandoning is Windows Media Center. Xbox is meant to replace that.

According to numerous rumors last year MS was close to selling off Xbox.

So they have been on the fence already.

And considering all that is happening currently I have a hard time believing in a successor for Xbox One.

 

[frozentundra123456]

I'm sure the standardization helps, but just look how well even a four+ year old Sandy Bridge quad core does for gaming.

Then factor in DX12 and Vulkan low level optimizations further enhancing those used CPUs potential for gaming in the future.

Then remember Microsoft loses money on consoles.

Do you see where this is going?

Do you have any documentation that MS is losing money on the current generation of consoles? I thought they were breaking even or coming close on the hardware, and made money from licensing and xbox live.

 

[Burpo]

OT I know.. but
"According to a recent SEC report, Microsoft's MSFT +0.02% Xbox division revenue is now up to $1.7B annually, because of the Xbox One, but its costs are up to $2.1B, also because of the Xbox One. That’s yielded a collection of startled headlines about how Microsoft has dropped the ball with their new console by losing $400M already.
The $400M figure itself was not the point of this article, but many are pointing out the math leading to all these headlines is misguided. These numbers reflect gross margin, not profit and loss, and are increases, not totals. While we can use that data to approximate $7.1B in revenues compared to $5B in costs, without further information from Microsoft, it’s hard to pinpoint an exact profit or loss on the Xbox One. Though if they have lost money on the system so far, it’s for the reasons I go into below, and would not be cause for alarm.

it costs a lot of money to develop, market and release a new console. As such, it’s fairly standard to expect to lose money on a system at first, given that Microsoft is playing the long game here."

http://www.forbes.com/sites/insertc...-fine-if-microsoft-has-lost-400m-on-xbox-one/

 

[cbn]

Do you have any documentation that MS is losing money on the current generation of consoles? I thought they were breaking even or coming close on the hardware, and made money from licensing and xbox live.

I am still investigating the most current results, but so far I did notice Anandtech report Xbox revenue decrease 24% for Q3 2105;

http://www.anandtech.com/show/9187/microsoft-releases-fy-2015-q3-financial-results

Xbox revenue dropped 24%, and Microsoft has not been able to keep up the post launch numbers from last year.

I'm sure a price drop will be on the way soon.

Meanwhile I expect the price of a Windows 7 licensed (upgradeable to Windows 10 for free) Sandy Bridge quad core SFF Pre-built to also drop in price.

 

[imported_ats]

So, if they save money by running 10nm wafers through 50% faster, why not do this for 14nm as well? Oh, because then it would become apparent that 10nm is more expensive.

The cost of multiple patterning techniques according to ASML:

Anyone else find it humorous that the graph and numbers don't match?

 

[Phynaz]

Intel generally doesn't get into price wars. They didn't do so even when they were getting beaten outright in both raw performance and perf/watt by AMD back in the Netburst days. Intel's culture is 60%+ gross margins on everything - with the notable exception of ultramobile Atoms, which they're willing to literally give away if that's what it takes to score design wins.

What you're suggesting here is AMD releasing a part with similar market positioning and appeal to Thuban. Remember when that was released back in 2011? You had to pay $999 for a six-core Intel SKU (the i7-980X), but AMD provided six full cores for $199-$295. The most popular Intel chip at that time for enthusiasts was the i5-750 (Lynnfield Nehalem). Thuban fell behind i5-750 in single-threaded performance, while defeating it in most multi-threaded benchmarks. You can see Anandtech's review results here. The important thing is that even though it did fall behind Intel's competing products in single-thread, it didn't do so by anything like the ridiculously embarrassing margins we've seen with Bulldozer and its progeny. Thuban was also only 904 sq. mm., less than a billion transistors, and was far more energy-efficient than Bulldozer (though not quite as much so as Intel's Nehalem-based products). I think we're going to see similar market appeal with Summit Ridge as we did with Thuban. For $299 or so, you'll be able to get 8 full cores, with each core having similar IPC to Sandy Bridge: basically the equivalent of a Xeon E5-2687W, though with less cache and memory bandwidth. In benchmarks using four or fewer threads, a traditional i7-4770K will beat it (though it will still provide reasonably competitive performance), but in multi-threaded applications, Summit Ridge will have a substantial edge.

Hopefully it clocks high enough that this comes true.

 

[cbn]

On the topic of Bristol Ridge (or whatever the Zen based APU gets called), one AMD tradition I hope we don't see continue is the disabling of much (or most) of the iGPU on the lowest end SKUs (eg, A4-5300, A4-6300, A6-7400K).

Instead, I think at least 75% iGPU should be left active for even the lowest end Zen based APUs.

With that mentioned, this doesn't mean I believe in large iGPUs for desktop.....but if AMD is going to design a large iGPU I feel they might as well let people use it.

 

[frozentundra123456]

Hopefully it clocks high enough that this comes true.

Yea, suddenly IPC has become the new buzzword for AMD fans. Even if Zen does substantially improve IPC, we still have to remember clockspeed matters too. We have hyperthreaded intel quads with 4ghz base, and hex and octa cores that fairly reliably overclock to 4ghz or better with a good air cooler. So lets see how samsung 14nm clocks. Personally I expect SB range IPC and maybe mid 3ghz base clocks, perhaps lower on the 8 cores if they want to maintain 95 watt TDP.

 

[coercitiv]

Yea, suddenly IPC has become the new buzzword for AMD fans. Even if Zen does substantially improve IPC, we still have to remember clockspeed matters too. We have hyperthreaded intel quads with 4ghz base, and hex and octa cores that fairly reliably overclock to 4ghz or better with a good air cooler.

Suddenly frequency has become a new buzzword. The irony.

For a meager 20-30% increase in IPC over a core arch considered as failed by every expert out there, some people on this forum sure seem scared of something. And the sad thing is... it's not even real, just a few pixels on a ppt slide. The odds are stacked against AMD mile high, and even with the most optimistic results they would barely get SB level of performance. At best we may get a cheaper solution with enough perf/watt for some users to consider it a better fit for their needs, while also providing AMD with a solid foundation for (hopefully) continued existence.

But that's it. No Intel killer, no Dragon slayer. Just a more competitive arch to review, compare, overclock, undervolt. In other words, more toys. Both Intel and AMD know how to do high frequency design, and for that we have to thank the oddball cousins, P4 and BD.

PS: hype will always be a problem in these discussions, but it's the same on both sides: remember when first geekbench benches for Skylake ES appeared and some forum members considered reported base frequencies as max clocks so IPC was literally Sky high?

 

[carop]

I didn't see any explicit fixed cost/invest in the slide. Is it hidden in "cost"?

I don't think fixed cost or investment is in that slide. However, cost seems to be a sensitive issue so they are using # of litho steps as well as # OVL (overlay) metrology in recent slides:

 

[Dresdenboy]

I don't think fixed cost or investment is in that slide. However, cost seems to be a sensitive issue so they are using # of litho steps as well as # OVL (overlay) metrology in recent slides:

Could that be throughput/risk related. This also lowers wafer yield.

Do you have more recent tool prices incl. EUV?

 

[jhu]

Yea, suddenly IPC has become the new buzzword for AMD fans. Even if Zen does substantially improve IPC, we still have to remember clockspeed matters too. We have hyperthreaded intel quads with 4ghz base, and hex and octa cores that fairly reliably overclock to 4ghz or better with a good air cooler. So lets see how samsung 14nm clocks. Personally I expect SB range IPC and maybe mid 3ghz base clocks, perhaps lower on the 8 cores if they want to maintain 95 watt TDP.

If AMD makes an 8 core that is able to get at least Sandy Bridge IPC at the 3-4 GHz range, I'd totally replace my FX 8350 (which actually doesn't really need to be replaced, and which replaced a 1090T that also didn't actually need to be replaced).

 

[JDG1980]

Personally I expect SB range IPC and maybe mid 3ghz base clocks, perhaps lower on the 8 cores if they want to maintain 95 watt TDP.

That sounds about right to me. But I think you might be a bit pessimistic about the 8-core clocks; Intel's current Xeon D-1540 has eight Broadwell cores that can turbo to 2.6 GHz, plus two power-hungry 10GbE controllers, all on a TDP budget of 45W. Double that TDP (rumored 95W for the top Summit Ridge parts), and you should easily be able to get above 3 GHz, even when taking into account the fact that Samsung/GloFo 14nm FinFET may not be quite as good as Intel's corresponding process.

 

[Exophase]

That sounds about right to me. But I think you might be a bit pessimistic about the 8-core clocks; Intel's current Xeon D-1540 has eight Broadwell cores that can turbo to 2.6 GHz, plus two power-hungry 10GbE controllers, all on a TDP budget of 45W. Double that TDP (rumored 95W for the top Summit Ridge parts), and you should easily be able to get above 3 GHz, even when taking into account the fact that Samsung/GloFo 14nm FinFET may not be quite as good as Intel's corresponding process.

Turbo clock is meaningless for determining thermal clocking limits. It's turbo precisely because it can't be sustained with everything running. It does however look like Intel is either short-changing the turbo specification (should be able to reach much higher than 30% more with only one core and a limited amount of peripherals active) or something else is limiting clocks.

 

[cbn]

Turbo clock is meaningless for determining thermal clocking limits. It's turbo precisely because it can't be sustained with everything running. It does however look like Intel is either short-changing the turbo specification (should be able to reach much higher than 30% more with only one core and a limited amount of peripherals active) or something else is limiting clocks.

I've been wondering if Intel is purposely limiting turbo clock on Xeon-D for product segmentation reasons. (re: Intel Ark does list several low power E 5 2600 v3 SKUs like this 55 watt Haswell E5 octocore with 2.9 Ghz turbo)

 

[jpiniero]

It's a server chip. Turbo boost is pretty meaningless, the companies that buy this will have the thing running at full load 24/7. Maybe the 1540 could cheat up to 2.6 for a brief moment occasionally if you don't use some of the features like only use one of the ethernet ports.

 

[ShintaiDK]

Turbo is very useful for .net type installations

But else depends on workload type. For something Like D series it does look relatively useless.

 

[AtenRa]

For the Data Center and Cloud, the work load is not 100% 24/7. What you need for those two is core count density, core count power consumption and core count perf/$.

 

[witeken]

AMD has confirmed that those recent slides are fake, FYI.

 

[Exophase]

It's a server chip. Turbo boost is pretty meaningless, the companies that buy this will have the thing running at full load 24/7. Maybe the 1540 could cheat up to 2.6 for a brief moment occasionally if you don't use some of the features like only use one of the ethernet ports.

You mean the companies are going to expect to have a constant load for all times of day? Or do they have load balancing that favors migrating activity to fewer sockets so they can completely deactivate some? I wouldn't expect it to work that way.

Then again, they may also prefer to maximize power consumption over latency when loads are lower.

 

[Snafuh]

AMD has confirmed that those recent slides are fake, FYI.

Source: http://www.planet3dnow.de/cms/14885-details-der-vermeintlichen-roadmap-genauer-beleuchtet/

„These slides contain fabricated content and were not generated by or on behalf of AMD” – AMD Spokesperson

 

[cbn]

Deleted.

 

[erunion]

Source: http://www.planet3dnow.de/cms/14885-details-der-vermeintlichen-roadmap-genauer-beleuchtet/

I'm not convinced.

I will be very surprised if those slides turn out fake. Not due to the content, but because the style looks very genuine AMD.

Obviously that earlier Fudzilla slide was fake, though.

We'll know in 2 days.