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

[R0H1T]

I don't think you are recalling the things correctly.

- Bulldozer didn't compete with Sandy Bridge at the time of its launch, but with Westmere. Only two quarters after Bulldozer was launched SNB-EP arrived, and AMD still lost share to Westmere-EP with Bulldozer.

- Bulldozer wasn't launched in 2009 because the 45nm Bulldozer sucked so hard that it couldn't even beat Thuban, so AMD canned the thing and went to the 32nm shrink. We can only wonder how bad that chip was.

- Raw performance wasn't the real issue when Bulldozer was compared to Sandy Bridge, it wasn't an order of magnitude different like it is today. In fact, Intel sold plenty of server SKUs with the same raw performance levels of Bulldozer SKUs, but with *much* improved perf/watt.

Yes, I did't get the timeline right.

That was because BD performed worse than Phenom, at least in terms of perf/$ IIRC wasn't it?

I know that but if you reread that post, what I said was that had BD launched a lot earlier & had time to be spun into something better AMD wouldn't have lost the marketshare that they did. BD was just way ahead of it's time & wasn't refined enough to do the tasks that it was supposed to, all within a reasonable TDP.

It's pure conjecture, yet again, but Excavator on 20nm would sell would't it? I'd like to believe there's still a market out there where perf/watt isn't the final (or the only) talking point before buying or building a server, even for business use.

 

[R0H1T]

If your position was correct, AMD would have greater than the 1.5% market share in the server space they currently have.

No that's because atm AMD is notoriously bad at perf/watt & only marginally better/worse in case of perf/$ depending on the total server cost.

An 80~100% more efficient chip (as compared to the original BD) that also performs a lot better is going to sell, maybe not in the truckloads that everyone here is asking it to but it will sell, & depending on the pricing it can be a hit that'll hit or hurt Intel for sure. Don't underestimate the lead Intel's fab have given it over AMD in all the segments where they operate, especially servers.

 

[mrmt]

It's pure conjecture, yet again, but Excavator on 20nm would still sell would't it?

No. Bulldozer and its descendants are fundamentally flawed, it's not worth to die shrink the thing.

I'd like to believe there's still a market out there where perf/watt isn't the final (or the only) talking point before buying or building a server, even for business use.

Usually when you are buying a server you don't believe anything, you run a TCO analysis of the proposed scenarios and pick the best offer, and guess what, perf/watt is *the* main metric on this analysis if you are talking about a datacenter, even the most pedestrian ones. Now if you are talking about a SMB with a server under a desk.... well, AMD still retains 1.5% of share despite selling 2012 trash silicon, there's a market for that. Whether this is a viable market in the long run that's another matter, especially because the cloud providers and virtualization tools are eating heavily on this market.

 

[AtenRa]

perf/$ is irrelevant, at least when the $ part is restricted to acquisition costs. That's why AMD didn't even bother to launch excavator on the server market. The CMT family was a real unmitigated failure on this one.

Since Servers are more about throughput and especially Integer, let me tell you a little secret.

28nm A10-7850K 3.7GHz Quad Core CMT is ~5 minutes faster (56,36minutes vs 61,51minutes) in x264 than 22nm Haswell core i3 4330 3.5GHz SMT. And that is without any L3.

A 20nm SOI Excavator would be faster than Haswell and with higher perf/watt. Intel was lucky GloFo and the rest of the Fab industry were after the Mobile market and 16/14nm FF made 20nm planar obsolete before it even started.

SteamRoller and Excavator were not released in the Server because of the luck of manufacturing process and thus they wouldnt be able to compete in perf/watt against 22nm FF Intel products not because CMT was not suited for the server market. In fact the CMT is way better for the Data center and the cloud than the SMT when both are using the same fab process.

 

[R0H1T]

No. Bulldozer and its descendants are fundamentally flawed, it's not worth to die shrink the thing.



Usually when you are buying a server you don't believe anything, you run a TCO analysis of the proposed scenarios and pick the best offer, and guess what, perf/watt is *the* main metric on this analysis if you are talking about a datacenter, even the most pedestrian ones. Now if you are talking about a SMB with a server under a desk.... well, AMD still retains 1.5% of share despite selling 2012 trash silicon, there's a market for that. Whether this is a viable market in the long run that's another matter, especially because the cloud providers and virtualization tools are eating heavily on this market.

Sure to each his own :thumbsup:

Why are you so hung up on perf/watt when we have no data on Zen, BD yes but anything else that follows is a cesspit of guesswork!

 

[Phynaz]

Well Amur is a 20nm soc in 2015, lets see when it will be release exactly.

Never.

Lisa Su noted that while AMD had started multiple 20nm designs, those projects aren’t going to come to market, due to minimal profitability on that silicon.

 

[AtenRa]

Maybe you should follow your own advice. That deal is dead, Seamicro no longer exists.

Im sure you have a link to prove that the deal is dead correct ??

 

[mrmt]

An 80~100% more efficient chip (as compared to the original BD) that also performs a lot better is going to sell, maybe not in the truckloads that everyone here is asking it to but it will sell, & depending on the pricing it can be a hit that'll hit or hurt Intel for sure. Don't underestimate the lead Intel's fab have given it over AMD in all the segments where they operate, especially servers.

AMD is one node behind Intel in servers, in 2016 they will still be one node behind Intel in servers and two in 2017. What are you expecting?

 

[AtenRa]

Never.

yeap, that was from the fake roadmap

 

[mrmt]

Why are you so hung up on perf/watt when we have no data on Zen, BD yes but anything else that follows is a cesspit of guesswork!

AMD failed to give K12 better performance/watt than Intel chips, and ARM is much cheaper to develop than x86 chips per AMD's own admitance. Now that AMD moved K12 to the backburner, we are supposed to believe that Zen will have better perf/watt, or at least good enough performance/watt?

I don't think Zen will have much chance on the server market, and I expect the trash silicon that didn't make into server SKUs to inundate the desktop market at very cheap prices. It's Bulldozer all over again, but at least this time AMD isn't embracing failed concepts like CMT.

 

[R0H1T]

AMD is one node behind Intel in servers, in 2016 they will still be one node behind Intel in servers and two in 2017. What are you expecting?

Try that again, same node in 2016 & one node behind in 2017. I don't believe every presentation that Intel gives & I do expect 10nm to be delayed just like 14nm was, as for whether GF/Samsung 14nm is equal to (or better) than Intel's 14nm I'd just say one word - mobile.

AMD failed to give K12 better performance/watt than Intel chips, and ARM is much cheaper to develop than x86 chips per AMD's own admitance. Now that AMD moved K12 to the backburner, we are supposed to believe that Zen will have better perf/watt, or at least good enough performance/watt?

I don't think Zen will have much chance on the server market, and I expect the trash silicon that didn't make into server SKUs to inundate the desktop market at very cheap prices. It's Bulldozer all over again, but at least this time AMD isn't embracing failed concepts like CMT.

Better than BD (at launch) relative to Intel at that time, surely that's not asking much is it?

We'll see whether the silicon was trash or everything else around it wasn't up to the mark.

 

[AtenRa]

AMD is one node behind Intel in servers, in 2016 they will still be one node behind Intel in servers and two in 2017. What are you expecting?

2016 both Intel and AMD will have 14nm in servers.

Also,

If you are expecting 10nm Intel servers in 2017 you are badly informed. Intel just released 22nm Haswell based E7 XEONs, they will introduce 14nm Broadwell XEONs in 2016 and 14nm Skylake in 2017.

So for 2016 and 2017 Both AMD and Intel Server SKUs will be on the 14nm FF process. Yes Intel may still have a small advantage but it will be nothing like the one they had with 32nm vs 22nm FF.

 

[Phynaz]

Im sure you have a link to prove that the deal is dead correct ??

Yes.

http://www.seamicro.com/

 

[mrmt]

Try that again, same node in 2016 & one node behind in 2017.

Samsung 14nm doesn't really match Intel 14nm specs. Just because they are both called 14nm doesn't mean that they are remotely close in terms of parameters. Samsung 14nm will be much more comparable to Intel 22nm than with the 14nm process.

whether GF/Samsung 14nm is equal to (or better) than Intel's 14nm I'd just say one word - mobile.

So you are extrapolating the performance of 150W SKUs from the results Samsung gets with 1-6W SKUs?

 

[AtenRa]

Yes.

http://www.seamicro.com/

I dont see anything about the Verizon deal being dead in this link.

 

[ShintaiDK]

I dont see anything about the Verizon deal being dead in this link.

SeaMicro was to deliver the servers. SeaMicro is dead.

 

[R0H1T]

Samsung 14nm doesn't really match Intel 14nm specs. Just because they are both called 14nm doesn't mean that they are remotely close in terms of parameters. Samsung 14nm will be much more comparable to Intel 22nm than with the 14nm process.



So you are extrapolating the performance of 150W SKUs from the results Samsung gets with 1-6W SKUs?

The only real way to determine if one fab is superior than the other is to have the same chip being fabbed at both of'em. I doubt Intel is going to GF or Samsung for any of their pieces of Silicon & neither is any other chipmaker that I know of, so you can't decisively close this debate till you remove every other outlier - a single chip or unique piece of silicon as such.

Parameters don't mean nothing if you don't have the performance to back it up, the mobile segment proves that.

 

[mrmt]

Parameters don't mean nothing if you don't have the performance to back it up, the mobile segment proves that.

How so? Did Intel say that its 22nm process was superior to TSMC or Samsung on the mobile market? I remember them making claims about 14nm or 10nm, but not 22nm.

 

[R0H1T]

How so? Did Intel say that its 22nm process was superior to TSMC or Samsung on the mobile market? I remember them making claims about 14nm or 10nm, but not 22nm.

No but wasn't Atom on 22nm supposed to be the best performing SoC, in terms of perf/watt ? Maybe it wasn't officially proclaimed (by Intel) at that time but I do recall vaguely that some forum members here claimed the same.

Of course since then ARM has moved onto A72, from A7, whilst still being on 28nm (majority of chips) but Intel also has moved two nodes ahead to 14nm.

 

[mrmt]

No, but wasn't Atom on 22nm supposed to be the best performing SoC, in terms of perf/watt ? Maybe it wasn't officially proclaimed (by Intel) at that time but I do recall vaguely that some forum members here claimed the same.

Forum members aren't part of Intel PR team as far as we are concerned. Only AMD disclosed a paid fanboy program for internet forums.

 

[MrTeal]

Yes, I did't get the timeline right.

That was because BD performed worse than Phenom, at least in terms of perf/$ IIRC wasn't it?

I know that but if you reread that post, what I said was that had BD launched a lot earlier & had time to be spun into something better AMD wouldn't have lost the marketshare that they did. BD was just way ahead of it's time & wasn't refined enough to do the tasks that it was supposed to, all within a reasonable TDP.

It's pure conjecture, yet again, but Excavator on 20nm would sell would't it? I'd like to believe there's still a market out there where perf/watt isn't the final (or the only) talking point before buying or building a server, even for business use.

If 20nm and the HDL actually provided a large reduction in power consumption, and if they could keep the clocks close to the same high speeds as PD and get the reasonable IPC improvement from PD to XV, and if it were priced reasonably, I think it would be a nice upgrade path for someone with an AM3+ board. ~FX-9590 performance at reasonable clocks with a non-insane 100W TDP.

That's a lot of ifs though, and I don't imagine the number of people looking to upgrade from a FX-6200 or FX-4300 would even come close to paying off the NRE. I can't imagine many people building a new DDR3-based AMD rig in late 2015 when Skylake will be the competition.

 

[Enigmoid]

So paying ~70% more for ~25% more perf, btw only ~12% more efficient, is good economics for you? Glad you don't advise small businesses cause that is an unmitigated disaster of epic proportions. Also I mentioned performance relative to a single core Xeon (obviously) & we don't know how many cores the top Zen SKU will be equipped with. Even after disregarding my other worldly predictions for the Zen there's no way any small or medium business is going to dole out even 50% more for a chip that performs only 25% better, that doesn't work & at least not in my part of the world.Is that irrelevant for everyone that buys a Xeon or only those firms where the cooling costs exceed the hardware (server) costs? If that's a market AMD can exploit then that is where they can succeed & surely that must equal something like ~20% of the current market where Xeon operates.

The CPU is a fraction of the cost of the computer. Its great economics when you consider the popular haswell 10-12 core chips are around $1700 (probably a good bit cheaper in bulk). Now lets take that CPU and add 64 GB RAM (64 @ $8 GB), a $300 mobo, $400 SSD, $100 PSU, $500 racks enclosures, floorspace, and assembly to get a total cost of ~$3500 (this is an extremely rough estimate but basically other system components are $1800 - perhaps there is a GPU). Now lets say we have another CPU that costs only 60% the price = $1020 with system cost ~$2800.

Cost of 100 Xeon servers = $3.5M
Cost of 125 opteron servers = $3.5M

But the Xeon's use less power (the CPU power is lower AND you don't have to power the equipment in 25 nodes), they require 25 fewer software licenses (if applicable), and requires 20% less manpower for tech support.

28nm A10-7850K 3.7GHz Quad Core CMT is ~5 minutes faster (56,36minutes vs 61,51minutes) in x264 than 22nm Haswell core i3 4330 3.5GHz SMT. And that is without any L3.

Not sure why you keep obsessing about L3. Kaveri has 4 MB L2 + L1. The i3 has 4 MB cache as well buts its inclusive (<3.5 MB total).

 

[Arachnotronic]

How so? Did Intel say that its 22nm process was superior to TSMC or Samsung on the mobile market? I remember them making claims about 14nm or 10nm, but not 22nm.

Intel said its 22nm SoC process was superior to TSMC 20nm for mobile in terms of xtor performance. 20nm is denser thanks to the fact that you can use tighter metal pitches (64nm for minumum metal vs 80nm for Intel 22nm).

I'd buy the xtor comparison, but at the same time, I would say that Intel's architectural/timing failures made it so that this comparison was moot.

At 14nm, Intel appears to have better xtor performance and superior density. In Cherry Trail we're still seeing architectural weakness relative to top ARM chips. We'll see if the 14nm SoFIAs and Broxton will help Intel compete better on this front.

 

[R0H1T]

Forum members aren't part of Intel PR team as far as we are concerned. Only AMD disclosed a paid fanboy program for internet forums.

I guess I'll have to dig up some of the review sites then, but since it's not official PR from Intel it'd just be a waste of time. So now how did anyone come to the conclusion that Intel's 14nm is better than the competition, other than Intel proving that beyond contention?

You can paint all the electrical parameters, from the competition, in bad light but since the underlying Silicon is different (to the point of them being designed for various markets as well as the underlying uarch & ISA) you cannot decisively award victory to one or the other unless you have removed every possible anomaly or difference that makes each chip unique.

 

[Abwx]

Samsung 14nm doesn't really match Intel 14nm specs. Just because they are both called 14nm doesn't mean that they are remotely close in terms of parameters. Samsung 14nm will be much more comparable to Intel 22nm than with the 14nm process.

Could you point us to published numbers or official statements that support this boat load of blank statements ..?.

I say that Samsungs Finfets are much more efficient, it s enough for this debate to keep going its course, why substantiate any saying after all..??.

You can paint all the electrical parameters, from the competition, in bad light but since the underlying Silicon is different (to the point of them being designed for various markets as well as the underlying uarch & ISA) you cannot decisively award victory to one or the other unless you have removed every possible anomaly or difference that makes each chip unique.

Best way to paint electrical parameters in bad light is to be clueless about what they are about or what they are actualy, i guess that it s not exactly the domain of accountants...

Intel said its 22nm SoC process was superior to TSMC 20nm for mobile in terms of xtor performance. 20nm is denser thanks to the fact that you can use tighter metal pitches (64nm for minumum metal vs 80nm for Intel 22nm).

I'd buy the xtor comparison, but at the same time, I would say that Intel's architectural/timing failures made it so that this comparison was moot.

At 14nm, Intel appears to have better xtor performance and superior density. In Cherry Trail we're still seeing architectural weakness relative to top ARM chips. We'll see if the 14nm SoFIAs and Broxton will help Intel compete better on this front.

That s not difficult, TSMC processes perfs are not that good, for instance GF 28nm efficency is almost a node apart of TSMC so called HP 28nm.

As for Intel s 14nm its not that fantastic and comparison of notebooks using same parts point to mediocre improvement over 22nm, once and for all a Core M has 10-15% better perf/Watt than previous gen 22nm Haswell.