New Zen microarchitecture details

[ShintaiDK]

I hope Zen is on par with Intel, or at least close enough to wage a price war. CPUs are commodity chips, and it would be good to see them priced accordingly.

Its very unlikely you see a price war. If Zen performs like X, it will be priced like X.

 

[SAAA]

It can't be a 4 modules CPU, it's way to rectangular. Maybe an opteron with 8 modules if you really want to push it:

It still looks too high to me tough. And the memory controller/cores aren't similar at all.

 

[turtile]

I hope Zen is on par with Intel, or at least close enough to wage a price war. CPUs are commodity chips, and it would be good to see them priced accordingly.

Neither company can wage a price war. Intel's pricing is lower now to maximize profit since the market is dying. Most server customers are going to stick with Intel initially too so there's no pressure where the market can actually move.

 

[senseamp]

Its very unlikely you see a price war. If Zen performs like X, it will be priced like X.

That may be true, but X is a variable dependent on the competitive environment. If Zen performs like Intel Core, is Intel just going to give up market share to AMD? I doubt it. They may keep the MSRP same to dupe their investors into thinking they still have 60% GM, but there will be stuff like counter-revenue to maintain share, so effective prices will drop.

 

[senseamp]

Neither company can wage a price war. Intel's pricing is lower now to maximize profit since the market is dying. Most server customers are going to stick with Intel initially too so there's no pressure where the market can actually move.

I wouldn't bank on that. Server customers are becoming larger and more hands on, they aren't just buying whatever Dell has, like the old days. They are building reference designs, and in some cases their own chips, and telling SuperMicro and Quanta what they want built. Once there is a competitive Zen, unless Intel drops the price to get in line, it will be designed out of customer's next compute node. If someone like AWS drops Intel, that is a big financial hit and a black eye for them, so they will play ball.

 

[ShintaiDK]

That may be true, but X is a variable dependent on the competitive environment. If Zen performs like Intel Core, is Intel just going to give up market share to AMD? I doubt it. They may keep the MSRP same to dupe their investors into thinking they still have 60% GM, but there will be stuff like counter-revenue to maintain share, so effective prices will drop.

That would be illegal. So no...

 

[frozentundra123456]

That may be true, but X is a variable dependent on the competitive environment. If Zen performs like Intel Core, is Intel just going to give up market share to AMD? I doubt it. They may keep the MSRP same to dupe their investors into thinking they still have 60% GM, but there will be stuff like counter-revenue to maintain share, so effective prices will drop.

Well obviously what he is saying is that if AMD has a competitive product they will price it close to intel prices, and I tend to agree, especially initially. 8 core zen will at best appeal to a what, 5 or 10 percent of the market, and i expect supplies will be tight and if the performance is competitive the can sell all they make. They would be fools not to charge a premium price for a competitive, limited supply, high end product.

 

[The Stilt]

they are not taking pictures of a brand new wafer to show it.

To me it looks like they just did :sneaky:

 

[sirmo]

It's been a while since we had performance competitive CPUs from AMD, so some may have forgotten, but usually when it comes to CPUs, when they are close in performance they trade blows. Neither one of them is an absolute winner.

But ultimately, the most important thing for Zen is perf/watt. They don't even need the absolute performance crown, as long as it's close enough. Haswell level on single thread would be sufficient.

If Zen is performance per watt competitive with Intel, there is no way it's not taking some marketshare from Intel in server space. And even 10% of the server marketshare taken would be huge for AMD. While 10% of the server marketshare may not mean much to Intel, for AMD that would double their earnings.

Intel might be ready to cut prices, but there is plenty of profit margin there for AMD to slot in.

Also Zen itself on its own isn't particularly dangerous in the consumer space. Let's remind ourselves that the only reason AMD's APUs haven't taken off as they should have is because they were held back by poor CPU cores, and an old fab process.

We know AMD is capable of delivering a power efficient CPU even on 28nm, I mean look what they have been able to do with Carrizo, and that's starting from an inherently inefficient arch. If Zen is performance competitive to Intel, Zen based APU will absolutely dominate x86 APU market.. and that's not even counting HBM based APUs. Which would be a no brainier for OEMs like Apple etc.. Heck they can even offer custom-semi designed SoCs with all the features an OEM would want to differentiate their product. Intel can't/won't compete with that.

It's a big if, but if AMD executes well on Zen, they are absolutely on the track of reclaiming some of their old market share. And as I said, AMD doesn't even need that much market share to turn things around, because they are a much smaller company today than they were before.

Intel perhaps saw the writing on the wall, when they decided to lay off 12% of their workforce. In either case, these next 18 months might be really interesting.

 

[Azuma Hazuki]

And Intel has worried about legality since precisely when, Shintai...?

 

[Arachnotronic]

If Zen is performance per watt competitive with Intel, there is no way it's not taking some marketshare from Intel in server space. And even 10% of the server marketshare taken would be huge for AMD. While 10% of the server marketshare may not mean much to Intel, for AMD that would double their earnings.

I am not sure if this is intended to be serious. If you think that Intel will just "let" AMD have 10% server market share, or that 10% server market share "doesn't mean much to Intel," then you must think that Intel is run by the biggest idiots in the semiconductor industry.

Losing ~10% share on ~$16B is $1.6 billion in revenue and ~$800 million in operating income. That's huge for a company that does ~$14 billion in operating income per year, especially when a big chunk of that operating income (PC related) is shrinking.

I guess using this logic, Intel should be cool if other competitors get little "slivers" of the server CPU pie too. Hey, what's 10% lost to Qualcomm, maybe 5% lost to Applied Micro, and heck, how about Cavium gets 5% too? Don't forget IBM and the OpenPower Alliance, let them have another 10% too!

Under this mentality, Intel's dominant (90%+ share) gets reduced to a mere 50% -- and that would hurt a lot .

 

[The Stilt]

Zen must provide 82% of the IPS of Broadwell-EX to match it's performance-per-watt figures in the high(est)-end server CPU segment. Can it do that? Your guess is just as good as mine.

 

[deasd]

For the price, I could only assume the AM4 boards would be cheap enough, not as expensive as X79/X99, but I have no idea about Zen, the chip itself, because there's no logical that make sense between price and performance.

edit: looks like some experts seize some hint from those slides:thumbsup:

 

[majord]

I actually think the L2 is further back, as per Jeebee's guess

http://semiaccurate.com/forums/showpost.php?p=263135&postcount=2746

 

[monstercameron]

Excellent post. May I add that I am extremely impressed with Dr. SU and her professional demeanor ( I wonder if she had some words "in private" with Joe Macri about his over the top overclocker's dream comments about Fiji).

I just looked at the top end AMD desktop processor, 9590 and it is $225 hot and uses an old platform.

ZEN, especially the top end one won't be cheap but I suspect will be a worthy and relevant successor.

I thought the ZEN brand was to cover the spectrum of AMD cpu needs from servers to laptops. The divergence will be apu vs desktop/server.

I know this discussion happened months back but the whole overclockers dream thing was always about the cooling and power delivery. They never stated that it would overclock high. I just think it is dishonest to bring up that quote and misconstrue as him saying the fury x would overclock high.

 

[senseamp]

Well obviously what he is saying is that if AMD has a competitive product they will price it close to intel prices, and I tend to agree, especially initially. 8 core zen will at best appeal to a what, 5 or 10 percent of the market, and i expect supplies will be tight and if the performance is competitive the can sell all they make. They would be fools not to charge a premium price for a competitive, limited supply, high end product.

I don't disagree that it will be priced relative to Intel prices, based on competitiveness. My argument is that Intel's own prices will be significantly lower than what they would have been if there was no competition from Zen.
Intel is not going to let them build market share without putting up one hell of a fight. They don't like low margins and counter-revenue, but with $15B of cash, they can survive a price war, it's an open question whether AMD can.

 

[itsmydamnation]

I don't disagree that it will be priced relative to Intel prices, based on competitiveness. My argument is that Intel's own prices will be significantly lower than what they would have been if there was no competition from Zen.
Intel is not going to let them build market share without putting up one hell of a fight. They don't like low margins and counter-revenue, but with $15B of cash, they can survive a price war, it's an open question whether AMD can.

Your ignoring what that would do to intels market cap. Especially because other area's are having negative growth intel will need to set prices that generate them the most gross profit/net profit/whatever.

Also the world has changed from 2005 you think google /facebook and co will just watch Intel destroy a competitive AMD with predatory pricing? These companies are big enough and forward thinking enough to know keeping two health x86 companies benefits them in the long run.

 

[Azuma Hazuki]

At this point I'd think they'd be more interested in propping up some ARM vendor or other. Intel might prefer that in the short run, too, since it would mean they can go "See? SEE? Not a monopoly!" while strangling AMD as per usual.

Of course that'll backfire horrendously if/when ARM manages to beat them out for the important server metrics, but since when has any C-suite given a damn about anything but the next quarter's results?

...I really hope AMD puts a big dent in Intel with Zen. If nothing else it'll light a fire under them and, for example, force the issue of 6+ cores in mainstream.

 

[itsmydamnation]

At this point I'd think they'd be more interested in propping up some ARM vendor or other. Intel might prefer that in the short run, too, since it would mean they can go "See? SEE? Not a monopoly!" while strangling AMD as per usual.

What you mean is like AMD .........assuming K12 makes 2017. AMD is likely of all the ARM server vendors to have the most complete and robust platform. Seattle got panned here (ignoring its a development chip) but organisations like Red Hat said as a platform it was much better then other arm platforms.

 

[sirmo]

I am not sure if this is intended to be serious. If you think that Intel will just "let" AMD have 10% server market share, or that 10% server market share "doesn't mean much to Intel," then you must think that Intel is run by the biggest idiots in the semiconductor industry.

Losing ~10% share on ~$16B is $1.6 billion in revenue and ~$800 million in operating income. That's huge for a company that does ~$14 billion in operating income per year, especially when a big chunk of that operating income (PC related) is shrinking.

I guess using this logic, Intel should be cool if other competitors get little "slivers" of the server CPU pie too. Hey, what's 10% lost to Qualcomm, maybe 5% lost to Applied Micro, and heck, how about Cavium gets 5% too? Don't forget IBM and the OpenPower Alliance, let them have another 10% too!

Under this mentality, Intel's dominant (90%+ share) gets reduced to a mere 50% -- and that would hurt a lot .

As I said if Zen is perf/watt competitive, there is nothing Intel can do to stop AMD from taking back some of that share. What's Intel going to do? Bribe OEMs again? Wouldn't be the first time. But they can't bribe everyone.

Not to mention China is the fastest growing server market, and the THATIC JV with Chinese government for server CPUs guarantees marketshare.

 

[majord]

Things not adding up Regarding L2 still.

Whilst i'm having a few issues estimating what density L2 cache should be - even with large amt of uncertainty, we can say L2 Cache should be somewhere between 1.1 and worse case ~1.4mm/MB (the A9 cache is around 1.1-1.2 mm2 for L3 and L2 respectivly for the record)

with that in mind, If the area shown was 512kb L2, that would then make each zen core only between 1.8 and 2.4mm2

Which is smaller than a single Apple Twister core , and frankly at least 3 times smaller than reasonaably expected . So hugely out of whack.

TLr I think L2 is hidden away there somewhere - (possibly not spanning the width of a core, hence difficult to identify) and the L3 is indeed 8MB x 2

 

[sirmo]

Zen must provide 82% of the IPS of Broadwell-EX to match it's performance-per-watt figures in the high(est)-end server CPU segment. Can it do that? Your guess is just as good as mine.

Seeing what they were able to do with Carrizo and the fact that Jim Keller was responsible for one of the most power efficient smartphone chips on the market (Apple's Ax chips).. I think it's not only possible, but probable.

 

[The Stilt]

I know this discussion happened months back but the whole overclockers dream thing was always about the cooling and power delivery. They never stated that it would overclock high. I just think it is dishonest to bring up that quote and misconstrue as him saying the fury x would overclock high.

If Fiji was called as an "overclockers dream" due the VRM (which isn't even that great) or the water cooling, then you could just as well call a rusty 1970 Ford F-100 pickup equipped with LS7 engine as a track racers dream. Both are equally faulty statements.

What I meant with outright false statements is the marketing figures used for Kaveri and newer APUs for example.

AMD has stated that the Steamroller based A10-7850K APU has 856GFLOPS of combined computing power. They've also stated (IIRC) that A10-7890K has 1.02TFLOPS of combined computing power. While based on the specs it should be the case, practically it is not. It is impossible for these products to reach the quoted figures.

That's because as soon as the iGPU is utilized, the CPU frequency will be locked to 3.0GHz (instead of the 3.7 / 4.1GHz) AMD used for the advertized GFLOP / TFLOP figure. It is technically impossible to reach the advertized figures, unless the default configuration is altered. I've released an unofficial fix (GeApm) to alter the default behavior, but without it these numbers are not real. The issue has been observer in various reviews.

856GFLOPs advertized for A10-7850K is in reality 833.28GFLOPs and 982.784GLOPs instead of the advertized 1018GFLOPs for the A10-7890K.

Not a big deal, but I don't like stuff like this. No matter which company does it.

 

[sirmo]

GFLOPs numbers are short for peak GFLOPs usually. At least that's the case with GPU specs. They often omit the "peak" from the marketing material though.

 

[The Stilt]

GFLOPs numbers are short for peak GFLOPs usually. At least that's the case with GPU specs. They often omit the "peak" from the marketing material though.

The thing is that it is not possible to achieve it even as a peak. Unless a peak is defined as few microseconds. That's around the time it takes for the power management to lock the CPU frequency to 3.00GHz after the iGPU is utilized.