AMD CPUs, SoC Rumors and Speculations Temp. thread 2

[cdrkf]

To be honest I think that is sensible. All I was trying to point out to some is that the chance something might flop is not the same as will flop.

If we approach Zen looking for issues I'm sure we'll find something. Similarly if we expect it to be the next k8 then we're likely to be disappointed as circumstances are very different (Intel is firing on all cylinders atm).

All I'd ask is can we try and give Zen a fair chance. It doesn't have to beat everything to be a good product for amd. I mean people are berating fury yet from my view is a great card that keeps up with the best from the competition. Those who are determined to be disappointed with fury are inevitably going to be disappointed with Zen, and that's the wrong way to view things imo.

 

[gamerk316]

4. CUDAs advantage is that they help in writing the software to take advantage of it and is guaranteed to be updated. Problem with most open source software is it depends on someone having an interest in it to update and write for it and if AMD is not going to put the foot work to do so, no one is going to do it for free. It is the same with Gameworks. While I hate the proprietary nature of things like these, I can't hate that nVidia pushes to get their ideas and tech into games a bit more than AMD.

Pretty much this. NVIDIA develops/purchases tech, but they push it as well. They make demos, release code, help developers, and get it adopted. OSS...not so much.

 

[8350rocks]

4. CUDAs advantage is that they help in writing the software to take advantage of it and is guaranteed to be updated. Problem with most open source software is it depends on someone having an interest in it to update and write for it and if AMD is not going to put the foot work to do so, no one is going to do it for free. It is the same with Gameworks. While I hate the proprietary nature of things like these, I can't hate that nVidia pushes to get their ideas and tech into games a bit more than AMD.

Pretty much this. NVIDIA develops/purchases tech, but they push it as well. They make demos, release code, help developers, and get it adopted. OSS...not so much.

That was partially my point.

If AMD could get something productive and moving in OCL and made it cost less than CUDA support/software much at all...they could conceivably convert a large chunk of the population using that stuff.

 

[jimmysmitty]

1. We will have to see. I don't think AMD can offer more cores at the same price if Zen is to be a normal CPU. The reason they could with BD/PD was because the extra costs for the secondary cores was not that much. IT cost more than SMT but not more than a actual second core. Remember that the 14nm they will be using will be new and more cores means lower yields so I expect them to match Intel in cores.

2. Will have to wait and see. AMDs 65nm was not a good process node, was one of the big issues with K10 as it had leakage issues. Their 45nm was better but Intels was still better thanks to the HK/MG. We also have to consider this is going to be the first time AMD uses FinFETs while 14nm is Intels second generation so they have a bit of an advantage there.

3. Considering the rumors I see it around Haswell at best. Which will not allow for better pricing but the same as usual, just under Intel probably top end matching top end i5/i7.

4. Anything they do can be better than BD/PD. As for DX12, from what we have also seen games using something like Mantle/DX12 also do not benefit as much on higher end systems. Low end CPUs with top of the line GPUs do as do low end GPUs with low end CPUs. But when it was tested, Mantle offered little to no performance gains for people with i5/i7s and top end AMD GPUs. I tested it myself and it was within 5-10FPS on any Mantle game. Of course we have to wait and see how DX12 will handle it in the end but we have yet to get an official DX12 game worth testing. Maybe in 3-6 months.

---------------------------------
The S9170 is nice but they still haven't countered CUDA and that is major.Having software support means a lot more than having the highest SP/DP out there. If anything AMD has always had higher TFLOPS capabilities than nVidia but I have always said that there is more to it than that and that does not always translate to performance, hence why a R9 290X has equal TFLOPS to a 980Ti yet doesn't perform the same.

How do we know they are in high demand? I have never seen a launch, short of the Bitcoin craze period, where they were sold out or out of stock most of the time. I think yields are just really low.

I think Nano will be interesting.

As for Fury X2, I will wait and see. They say the radiator can handle 500W but dual Fury GPUs is a lot of power and heat. I think it iwll be pushing that radiator close to its limits if they even do it.

Here is the thing regarding (1): AMD is shooting for top SKU to be 16 cores if they can keep the die size manageable...if not 16 then 12, if not 12 then 8. Nothing less than 8 physical cores will be the flagship from my information...but it will be 8C/16T versus 8C/8T now.

Regarding (2): This may be AMDs first attempt with FinFETs, however, I expect this to be less an issue with design, and more an issue with "fine tuning" so to speak. I would imagine they get everything they want to do in, and have no issues there...I just expect that the second generation (Zen+) will likely see some pretty good gains from low hanging fruit, plus optimizing current uarch after observation/R&D.

As far as pricing goes...sure...you may potentially have less IPC...however...if you can get an 8C/16T AMD for the same money as an i7 intel...and AMD catches up to haswell...that is a might powerful processor...especially if power consumption is inline. If they get SMT nailed on the first run, they may even see a large uptick in server penetration and market share.

CUDA is a farce, you can do all that with OCL, and AMD would benefit hugely having someone to write a lot of these programs that are broadly applicable to specific markets with heavy CUDA saturation.

1. Right now it is 4M/8T not 8C/8T. I don't count their CPUs as 8 cores. And depending on the yields of the new 14nm will determine pricing as will more cores. 16 cores would have to be the very top end and again low yields would make it harder to put a lower price if true. Still waiting to see official specs.

2. That is what I am getting at, again Intel has already had time to fine tune the use of FinFETS and by the time Zen is out they will have "tuned" their 14nm for better yields/lower voltages. However without seeing the design we wont know this for sure. Remember, BD looked good on paper but is a leaky bastard.

3. Yes it would be but again more cores does not always mean better. AMDs Phenom II X6 wasn't better and still is not. It all depends on a lot of factors. This is why I hate rumors, they never bring anything good but hype for something to never live up to.

4. CUDAs advantage is that they help in writing the software to take advantage of it and is guaranteed to be updated. Problem with most open source software is it depends on someone having an interest in it to update and write for it and if AMD is not going to put the foot work to do so, no one is going to do it for free. It is the same with Gameworks. While I hate the proprietary nature of things like these, I can't hate that nVidia pushes to get their ideas and tech into games a bit more than AMD.

That said, Zen is nothing but rumors and hope. If we let the hype build up it could just turn into another K10 if it has an issue or falls short. I say we just hope it is decent and not some magical cure all for AMD because it is very certain that it could flop.

MOAR COREZ can be better depending on what you use it for...CPU rendering is a great example...so is encryption, multi-tasking, running virtual machines, any number of things can benefit from more cores.

As far as tuning goes...if that is where Intel came up with their 5% gains from Ivy to Haswell...not really sweating that one too much. Especially since the murmurs coming from team blue are talking about having sub 5% gains from Haswell to Skylake at this point...

At that point...if AMD ties haswell...and Intel is only ~10% ahead...what a race it is...

No Devils Canyon was the more mature 22nm. Ivy to Haswell had a lot of improvements in areas we wont see, especially in encoding/decoding.

As for Skylake to Haswell, have yet to see anything official on official hardware. Rumors are rumors. I doubt it is going to be amazing, then again we have hit a wall and the only reason Zen will look so good is because PD is nothing amazing to start with.

And within 10% wont win them jack in the server market. They need to beat Intel there. Not so much in the consumer desktop market but they need to win in the much more lucrative server market to make Zen worth anything.

4. CUDAs advantage is that they help in writing the software to take advantage of it and is guaranteed to be updated. Problem with most open source software is it depends on someone having an interest in it to update and write for it and if AMD is not going to put the foot work to do so, no one is going to do it for free. It is the same with Gameworks. While I hate the proprietary nature of things like these, I can't hate that nVidia pushes to get their ideas and tech into games a bit more than AMD.

Pretty much this. NVIDIA develops/purchases tech, but they push it as well. They make demos, release code, help developers, and get it adopted. OSS...not so much.

That was partially my point.

If AMD could get something productive and moving in OCL and made it cost less than CUDA support/software much at all...they could conceivably convert a large chunk of the population using that stuff.

But AMD never has nor do I doubt they will spend the money and time to write and code for it.

If they ever do then good. But it is the same as with Intel. Intel works with software developers and other hardware manufactures to get things working for what they have. It costs a lot of money to have people doing that. AMD doesn't have the spare money. I would rather at this point see them focus on improving their driver support for their existing platforms.

 

[bharathbigb]

To me,Zen should touch Sandy Bridge levels of Performance,UNDER 95w TDP.But if They could get to Haswell Levels,which is like 20% Faster (Not Sure.Please Correct me,If I am Wrong) than Sandy Bridge (Yet Still an OC'ed 2500K can make a 4690K sweat)It is surely an Engineering Feat with literally 10x LOWER Budget,Also some Fine Tuning will make it a Proper Competetion to Intels..This will be a Win for me.AMD should truly stop Hyping their Products.I mean look at AMD,They Hyped the Fury X as a Titan Killer in which a Single Fury X with 4 GB Limitation can NEVER Be.But Crossfire it and it will smoke any Nvidia Card out there even if Nvidia made their Dual-GPU Variant.I don't really care about Power Consumption because we are talking about Desktops and NOT Portable Nor Mobile.If the Fury X is rated at 500w WHO THE HELL Cares ? Because to just NOT Bottleneck that Card,You need an i5 4690K.These are Enthusiasist Grade Parts and will definitly need Good Power and Cooling.

A True Gamer or an Enthusiasist will never want any competetion to go and become a Monopoly.First off I will NEVER Trust Nvidia/Intel as a Monopoly because If AMD Dies,they can price a Geforce 210 and a Celeron 1000$ EACH which can NEVER Play anygame today.If this Happens,I really wanna see the face of Nvidia/Intel Fanboys who generally Bash AMD.

Also,They need to Introduce many products like Athlon X4 860K,A10 7870K.etc In India because it is a Huge Market.With Good Advertising,They will surely gain some Money/

 

[jdwii]

I got us all in trouble before(sorry guys) but i really don't think Amd is getting near haswell IPC and i expect similar or lower clock speeds with the first launch of Zen. Not to sound like a Amd troll but this is a new design based on old designs and with improvements its not going to beat a design by Intel that is getting pushed to its limits.

 

[bharathbigb]

I got us all in trouble before(sorry guys) but i really don't think Amd is getting near haswell IPC and i expect similar or lower clock speeds with the first launch of Zen. Not to sound like a Amd troll but this is a new design based on old designs and with improvements its not going to beat a design by Intel that is getting pushed to its limits.

Please...Zen IS NOT some Old Design with Improvements.Its a FULLY GROUND-UP NEW DESIGN LEAD BY JIM KELLER.But you are right about the Performance Part.Nobody CAN Expect Zen to be on Par with Haswells because with that levels of Budget,Its Practically Not Possible.Reaching Haswell Levels is sure Very Challenging for Jim Keller and the Whole Company with so much less money.But as an Enthusiasist,I believe in 'Nothing is Impossible' .If Keller and Team could pull it off,I mean Man....That surely is one helluva Work put into it.

 

[cdrkf]

Again I think we must clarify in what context we expect certain performance metrics. In pure IPC then they probably won't catch haswell. In multi thread they will likely be as fast or faster thanks to more cores on the high end parts without the module penalty.

In terms of performance per watt, thats the killer question and the hardest to predict. If they can for example be within 10% of Intel in raw performance, but match or marginally exceed them in perf per watt, then they have a strong case for servers.

As always it's a complex topic. I predict they will be very efficiency biased based on they're recent work on carrizo and the fact Keller has come from developing for mobile.

 

[juanrga]

(1): AMD is shooting for top SKU to be 16 cores if they can keep the die size manageable...if not 16 then 12, if not 12 then 8. Nothing less than 8 physical cores will be the flagship from my information...but it will be 8C/16T versus 8C/8T now.

Regarding (2): This may be AMDs first attempt with FinFETs, however, I expect this to be less an issue with design, and more an issue with "fine tuning" so to speak. I would imagine they get everything they want to do in, and have no issues there...I just expect that the second generation (Zen+) will likely see some pretty good gains from low hanging fruit, plus optimizing current uarch after observation/R&D.

As far as pricing goes...sure...you may potentially have less IPC...however...if you can get an 8C/16T AMD for the same money as an i7 Quad intel...and AMD catches up to haswell...that is a might powerful processor...especially if power consumption is inline. If they get SMT nailed on the first run, they may even see a large uptick in server penetration and market share.

CUDA is a farce, you can do all that with OCL, and AMD would benefit hugely having someone to write a lot of these programs that are broadly applicable to specific markets with heavy CUDA saturation.

(1) Are you talking about desktop or server?

(2) According to some sources Zen was first designed for 20nm planar, then ported to 16FF. It remains to be seen like good is the porting. Zen+ will be a minor improvement over Zen.

(3) No gonna happen AMD 8C/16T will be priced as an i7 Quad Intel whereas catching Haswell IPC. Do I need to recall what happened with Fury X HYPE with some people waiting Nvidia performance at half the cost?

(4) The penetration of Zen in the server market will be in the single digit percent even if AMD manages to offer a Haswell IPC processor.

(5) You can do all that C does on assembly, that doesn't mean that developers will massively move to assembly.

 

[jimmysmitty]

@8350rocks one thing I do agree with Juan on is that the concept of "80-90% performance for 20% of the price" is very unlikely. If AMD have a chip at 90% of the performance of the equivalent Intel, I'd expect it to be priced accordingly (I do think AMD usually offer a bit more for your money, so if it's 90% performance, maybe 80% the cost so a nice but more modest saving).

I mean when you look at the launch prices of their most competative parts (and espetially the prices of the original FX parts when they had a definative lead) and, whilst not as insane as some of Intel's 'EE' parts they definately changed good money for them.

That said I'm all for that really- AMD need to make better margins for them to remain viable long term. As I said before as well, I anticipate that some of the harvested sku's (e.g. hex core parts built from octa core dies) will probably wind up being the real value leaders, just how the Radeon Fury is a considerably better deal from perf / $ perspective compared to the top end part due to the fact they'll invariably have quite a lot of silicon with small faults so it makes sense to get something for them whilst charging a premium for the best silicon.

The FX series cost more than the EE series when it was king. Hell AMDs solution to the Core 2 Quad was the QuadFX system which at the time the CPUs alone cost $1400 bucks ($700 each) and at 3GHz (these were the FX-74s) couldn't beat the Q6600 which was $850 bucks at launch. And that was after Core 2 took back the performance crown.

The FX-60 launched at $1031. Intel didn't even have a EE out at the time so I don't have anything equal.

The Fury X is priced equal to the GPU it performs with. If it beat the GTX 980Ti it would have been priced higher.

AMD will most certainly take advantage of a performance lead. Any company will, there are no such things as charitable for profit companies.

I just don't get why people forget this. People seem to think AMD is a company that is giving when the only reason for it right now is because they don't have the performance. If Zen does what they want it to do then it will be priced higher than Intels offerings.

 

[con635]

^amd mid range is always the great value, they're not as stingy with unlocking/overclocking. I think that's often forgot about by enthusiasts these days.

 

[jimmysmitty]

^amd mid range is always the great value, they're not as stingy with unlocking/overclocking. I think that's often forgot about by enthusiasts these days.

Back then there was no locked/unlocked with overclocking. Both worked the same, mostly by FSB.

Sure now it seems that way but again most of the time there is very little benefit these days to overclocking a low-mid range CPU as most are already clocked high enough and the gains are not as good as they used to be.

Used to be able to get a Pentium D 805 and overclock it to insane levels and it would beat both the Pentium EE and AMD FX chip. But even the unlocked anniversary edition Pentium G3258 overclocked to insane levels wont beat top end i5/i7s.

 

[8350rocks]

@8350rocks one thing I do agree with Juan on is that the concept of "80-90% performance for 20% of the price" is very unlikely. If AMD have a chip at 90% of the performance of the equivalent Intel, I'd expect it to be priced accordingly (I do think AMD usually offer a bit more for your money, so if it's 90% performance, maybe 80% the cost so a nice but more modest saving).

I mean when you look at the launch prices of their most competative parts (and espetially the prices of the original FX parts when they had a definative lead) and, whilst not as insane as some of Intel's 'EE' parts they definately changed good money for them.

That said I'm all for that really- AMD need to make better margins for them to remain viable long term. As I said before as well, I anticipate that some of the harvested sku's (e.g. hex core parts built from octa core dies) will probably wind up being the real value leaders, just how the Radeon Fury is a considerably better deal from perf / $ perspective compared to the top end part due to the fact they'll invariably have quite a lot of silicon with small faults so it makes sense to get something for them whilst charging a premium for the best silicon.

The FX series cost more than the EE series when it was king. Hell AMDs solution to the Core 2 Quad was the QuadFX system which at the time the CPUs alone cost $1400 bucks ($700 each) and at 3GHz (these were the FX-74s) couldn't beat the Q6600 which was $850 bucks at launch. And that was after Core 2 took back the performance crown.

The FX-60 launched at $1031. Intel didn't even have a EE out at the time so I don't have anything equal.

The Fury X is priced equal to the GPU it performs with. If it beat the GTX 980Ti it would have been priced higher.

AMD will most certainly take advantage of a performance lead. Any company will, there are no such things as charitable for profit companies.

I just don't get why people forget this. People seem to think AMD is a company that is giving when the only reason for it right now is because they don't have the performance. If Zen does what they want it to do then it will be priced higher than Intels offerings.

While you are true in regards to the pricing when they outright had the performance crown. I expect that AMD with 90% performance would take the slightly cheaper route still...because having high volume with slightly less margins is much better than ridiculous margins with extremely low volume.

I do think an 8 core of theirs would likely be priced competitive to something like a 4820k/5820k...not a 4970X/5970X

 

[jimmysmitty]

^Of course they will. Thats my whole point. I just get tired of people forgetting that AMD prices their products accordingly. The HD7970 launched at $550 when the GTX 580 was $500. Made perfect sense as it performed better than the GTX 580. They didn't price it cheaper or lower because for the time between its launch and the GTX 680s launch they had the better performance.

 

[-Fran-]

I think pricing wise, it will be depending on how bad AMD wants to take back market share. 8350Rocks might be just right saying 90% for 40% of the price if AMD can still make a profit (a big *if* I might add). You know how it is in business unfortunately. There's always the money concern for the financial arks before what the market actually wants.

Zen has to be a solid tie. And answering that, I mean for every important metric the Server business uses, they have to perform the same as Intel at the same power envelope. I don't know how much margin of difference they will have in terms of price (that 90% at 40% price talk), but I will not assume the 14nm process node will come in cheap for AMD. They're bleeding money and that 2020 prediction might be right on the money (geddit? xD) if Zen does not create market penetration for them.

So, in short there are broadly 2 fronts of attack for AMD where we can actually speculate: how bad they want market share (I'd say pretty bad) and the classic performance talk (tie with Haswell? better? worse? etc).

Cheers!

 

[de5_Roy]

if they can't match haswell they can compete against broadwell-de (xeon-d) with zen-opterons - a much more viable target and sorta intel's weak spot in it's server offerings. glofo's 14nm won't deliver the power efficiency but it will raise the performance level.

 

[jimmysmitty]

I think pricing wise, it will be depending on how bad AMD wants to take back market share. 8350Rocks might be just right saying 90% for 40% of the price if AMD can still make a profit (a big *if* I might add). You know how it is in business unfortunately. There's always the money concern for the financial arks before what the market actually wants.

Zen has to be a solid tie. And answering that, I mean for every important metric the Server business uses, they have to perform the same as Intel at the same power envelope. I don't know how much margin of difference they will have in terms of price (that 90% at 40% price talk), but I will not assume the 14nm process node will come in cheap for AMD. They're bleeding money and that 2020 prediction might be right on the money (geddit? xD) if Zen does not create market penetration for them.

So, in short there are broadly 2 fronts of attack for AMD where we can actually speculate: how bad they want market share (I'd say pretty bad) and the classic performance talk (tie with Haswell? better? worse? etc).

Cheers!

Performance sells. That's why even the $1K CPUs from Intel still sell.

 

[-Fran-]

if they can't match haswell they can compete against broadwell-de (xeon-d) with zen-opterons - a much more viable target and sorta intel's weak spot in it's server offerings. glofo's 14nm won't deliver the power efficiency but it will raise the performance level.

That is something I don't know, to be honest: Intel's server line up in terms of performance. I haven't been keeping up with their releases 😛

That will shed some light to the possible price scheme of Zen if we assume/project certain performance targets.

Performance sells. That's why even the $1K CPUs from Intel still sell.

Yes, definitely. If perf/watt is not something clients are looking for and just want sheer performance, AMD might still be in for some action. They'll be getting into IBM territory and Intel's higher offerings though, so Zen would have to be really good to play there. If not, all they have is the perf/watt battle.

I read this in the Intel thread, but I don't know how much of a FAD is for big companies to "reduce the carbon footprint", but some could make an argument based on that where Intel would be playing alone, but in every other scenario, AMD might be able to position Zen, as long as it's on par with Intel performance wise.

So many questions, but little answers yet. Anyone knows anything about GloFo's 14nm? What they'll be doing or something?

Cheers!

 

[jimmysmitty]

The only thing we know about GloFlos 14nm is that it is based on Samsungs which has proven decent in their Galaxy S6 CPU.

 

[juanrga]

(4) Nobody cares about HPC. As for being 20%, you do not have any idea that skylake will launch at 4 GHz base clock, and I honestly doubt that is the case to be honest.

Skylake slides have been just leaked and they report 10--20% increase in performance over Broadwell. 😀

I am sorry to say this but Zen CPU looks DOA, with each new leak.

 

[jimmysmitty]

(4) Nobody cares about HPC. As for being 20%, you do not have any idea that skylake will launch at 4 GHz base clock, and I honestly doubt that is the case to be honest.

Skylake slides have been just leaked and they report 10--20% increase in performance over Broadwell. 😀

I am sorry to say this but Zen CPU looks DOA, with each new leak.

I think for me the temps are what amaze me. Then again they removed the VRMs from the CPU and that helps a ton.

 

[juanrga]

Spoiler

1.) I was talking about HEDT with 8C/16T, servers will likely be shooting for a 16C/32T SKU and if they get more out of it...then they end up something in the neighborhood of maybe 24C/48T or something along those lines...32 cores...who knows.

2.) AMD is going to do something to differentiate themselves. Recently, that has been more cores...the last word I have on the subject was that the more cores trend will continue. As I said before...they are shooting for 8 cores in HEDT.

3.) Running AVX2 code that no one even uses yet because you are using floating point operations is not what I am talking about. I am talking about on your average 2006 C++ x86 library extensions that run in every computer on the planet using windows/linux/mac, they got 5%. Sugar coat special cases all you want...that does not impress me.

4.) Skylake will not be 20-25% ahead if AMD ties haswell...I agree it is optimism to expect them to do so...however, skylake will be about 5% ahead of haswell in things that matter...like not specially paid for by intel benchmarks that show a best case small program running only floating point operations to showcase AVX512. Nobody cares about that stuff...because the real world has very little, if any, use for it. So the "100% faster with...blah, blah, blah" is a lot of blue sky, and nothing more.

(1) Info I have is that Zen-based server CPUs will get up to 32 cores, using a multi-die configuration (four dies on interposer). That will introduce latency and power disadvantages compared to Skylake Xeons, which will bring up to 28 cores on a monolithic die. Therefore AMD will be at clear disadvantage here: less performance and less efficiency.

(2) Intel has been selling 8 cores for HEDT since the past year. AMD will not differentiate anything by providing 8 cores for HEDT in late 2016...

(3) Haswell got a ~5% IPC for legacy x86 binaries plus ~15% higher frequency, which gives about 20% higher performance not 5%.

The reason why Intel gets 5--10% IPC gains per gen is because the old x86 ISA is near its practical limits and the only way to extract significant performance is by developing a new modern ISA. That is what Intel has been doing. The reasons why AVX2 is not massively popular (there are however more AVX2 software than HSA software) are that AVX2 is not supported on all Haswell chips (e.g. Haswell Pentium doesn't support AVX2) and that Phi Xeons use a different 512bit ISA. Skylake Xeon will introduce AVX 512 ISA convergence with KNL Phi and it seems evident low-level chips will receive the ISA when was mature.

When x86-64 appeared it wasn't instantaneously adopted by developers. Why would Intel new ISAs be different?

(4) Zen will hit ~3.5GHz and according to data provided by AMD, the IPC will be behind Haswell. Let us assume that AMD ties Haswell IPC for legacy x86 binaries. Skylake will launch next month with 4GHz base clock and somewhat between 5% and 10% higher IPC for legacy x86 binaries. That gives roughly 20--25% higher performance. Now, when Zen launches Intel would have ready Skylake refresh with higher clocks, maybe 10% higher.

And of course, the performance with new binaries will be very very ahead of AMD, maybe 2x. That explains why no customer has chosen Zen CPUs for building the fastest supercomputers in the Earth. That people uses optimized binaries, not slow binaries compiled 10 years ago. We also know that

certain customers are canceling their orders of current-generation Intel Xeon-powered supercomputers in anticipation of machines based on “Skylake” chips. For example, the U.K.’s Met-office wants “Purley”-based supercomputers because of higher performance and energy efficiency.

(1) Source? You never post sources...just speculation without a reliable source document for people to view and interpret independently.

(2) Sure, pay $1k for an 8 core Intel, I will take an AMD for $200 and get 80-90% performance for 20% of the price.

(3) x86-64 actually was adopted almost instantly...it was a huge hit and Microsoft themselves came out endorsing how wonderful it was. Meanwhile Itanic (Itanium) sat by the way side while everyone cheered AMD on for doing the smart thing and making a legit 64 bit ISA that could allow the legacy programs MS had us all stuck on to continue to flourish. Remember...the windows everyone was running in the 90s was mostly 16 bit or towards the end of the 90s, 32 bit.

(4) Nobody cares about HPC. As for being 20%, you do not have any idea that skylake will launch at 4 GHz base clock, and I honestly doubt that is the case to be honest. Especially considering that the clocks on their chips have mostly hit a plateau, or even declined in some cases. You have the devil's canyon stuff...but the devils canyon chips consumed significantly more power according to all the reviews, and that goes against your perf/watt arguments...so what is most important? Perf/Watt, or outright performance?

(2) Performance/cost is a misleading metric for CPUs, because the relationship between ST performance and complexity is highly non-linear. Introducing 2x more transistors on the design of a core doesn't improve IPC by 2x, but a lot less, which hurts the performance/cost metric. Providing 80% of performance by about 60% of cost is not any technological merit, but just a consequence of laws of physics.

Add to that of above that AMD has been using a cheap node and selling CPUs under cost --the CPU division has been in red numbers during years--. AMD will not sell Zen CPU at prices under cost. They no longer have the money and market for that pricing.

Increasing the IPC by 40% is going to cost AMD about something between 2x and 3x more per CPU, considering both the extra transistors that each Zen core will need and the extra design checks (e.g. checking there is not a critical path bottleneck affecting cycle time due to design complexity increase). Add that 14nm process node is much more expensive than 32nm node used in Piledriver FX CPUs and the hype that Zen CPUs will provide Haswell-like performance by half the cost is gone. Zen CPU will be an enthusiast-class oriented product aimed at competing with Intel eXtreme series and pricing will be similar. Mainstream user will be served by the cheap 28nm made Excavator-based APUs.

We are here on similar situation to what happened on GPU, with some people waiting Nvidia-like performance at half of price and being disappointed by both performance and pricing of fury X flagship product. I warned before launch. I am warning again.

(3) x86-64 wasn't adopted almost instantly. In fact AMD had to actively fight chicken and egg problems

http://arstechnica.com/features/2008/09/x86-64/5/

AMD x86-64 was a dumb move. Intel and HP tried to develop a new, clean, legacy-free, scalable ISA abandoning the limits and weirdness of the x86 ISA. Intel and HP failed, but they tried. Probably they could have been successful on a second or third try.

AMD did the easy stuff: develop a 64bit extension of x86-32. 'Thanks' to AMD, now we have one of the weirdest 64bit ISA in the market: unscalable, highly inefficient, and costly. If at least they had introduced AMD64 as a separate ISA instead like an extension to x86-32... AMD myopia solved the short term problem and generated long-term problems: "bread for today and hunger for tomorrow".

(4) AMD cares about HPC and cares a lot of...

http://www.amd.com/en-us/solutions/servers/hpc
http://www.amd.com/en-us/solutions/professional/hpc
http://www.fudzilla.com/news/processors/37395-amd-announces-glorious-five-year-plan
http://www.hpcwire.com/2015/05/07/amd-refreshes-roadmap-transitions-back-to-hpc/
http://www.theplatform.net/2015/03/10/amd-hints-at-future-hpc-push/
http://www.bit-tech.net/news/hardware/2015/03/31/amd-hpc-roadmap/1

The 4GHz base for i7 skylake is well-known.

With your #2, you are trying to create correlation between things that have no correlation. You can create more complex CPUs at a similar cost to a less complex CPU depending on multiple variables that you cannot possibly have intimate knowledge of for either AMD or Intel. So, barking up that tree and assuming more complexity adds to cost is a fool's errand. In some instances, a more complex CPU can even be cheaper to produce...depending on the factored costs invovled. So this argument has no leg to stand on.

As for the #3 point, x86-64 has been credited by multiple sources, like forbes and others, as being the single most significant ISA change to x86 in the last 30 years. You are claiming it was "dumb"?

#4 AMD is interested in HPC like any major player in PC components should be...especially considering their desire to drive innovation. Now, as far as being in HPCs goes...they would love to have their CPUs and GPUs in every PC on the top 500. However, they would rather have their CPUs and GPUs in half the households in the world more....

(2) The small details can vary in function of those unknown variables, but not the general picture:

(2.i) For Zen architecture, AMD is starting from scratch. The cost of developing a new architecture from a "clean piece of paper" as Keller put it, is higher than tweaking/evolving an existent architecture like in the transition Bulldozer--> Piledriver.

(2.ii) Zen will use many more transistors. The cost of developing, validating, and fabricating will be increased. A 200mm² die cost more than a 100mm² die. Precisely die cost was the main reason for the CMT architecture. It saved for AMD very precious mm².

(2.iii) Zen will be made on a 14FF node, which is much more expensive than the 32PL node used for Piledriver FX CPUs. To get things in perspective it costs about three times more to design a chip with FinFET transistors than to design a chip with planar transistors.

(2.iv) AMD has been selling Bulldozer/Piledriver FX CPUs under cost, due to their lack of competitiveness. This is the reason why AMD CPU division has been in red numbers. AMD solicited bank credits to stay afloat and continue running the company. You cannot pay bank debts by staying on red numbers. Thus AMD will not sell Zen CPUs under cost.

I don't know the fine details and cannot say that octo-core Zen CPU will cost exactly $594.23. But I can analyze the whole picture and claim that people waiting Haswell-E performance at Piledriver prices will be disappointed.

In fact, I warned people that Fury X was going to be costly. I couldn't compute the exact launch price of the card, because lacking those fine-detail variables, but I knew that people waiting a 980Ti killer at half the price was clearly wrong. And time has given me the reason. In fact people was disappointed by Fury X pricing.

(3) I have claimed that it was a dumb move by several reasons, including the fact that it is an extension instead a separate ISA. Those dumb moves are the reason why x86-64 is almost universally considered one the worst 64bit ISAs in the market., with half-dozen of better alternatives.

(4) The lack of public design wins for Zen is saying us is that Zen is not competitive in performance/efficiency. We also know UK consumers that have canceled Xeon purchases because are awaiting for Skylake Xeons. No customer has claimed it is waiting for Zen Opterons, which suggest that is not a competitor for Skylake Xeons.

Desktop/laptop chips will reuse technology from servers. Therefore, there is no reason to believe that Zen based CPUs/APUs will take "half the households in the world".

2.) They were not truly starting from scratch. Want to bet that under all that are some fragments from K10, and pieces of piledriver/excavator, and innovations from other areas as well?

As for process comparisons. 14nm FF *ON BULK* is probably about equal in cost to what 32nm PD-SOI ran when they implemented it. Considering that SOI is a dramatically more complex substrate than just running FFs on bulk wafers...you likely underestimate how little the cost difference actually is. To put it to you another way...Intel conceded that if they had tried to change from bulk to SOI, instead of just tooling for FF on bulk...their fab would likely have nearly bankrupted their cash reserves with R&D and tooling costs

For the cost portion...you do realize that wafer agreements were the predominant reason the CPU division was in the red right? The agreement between AMD/GF has actually put them in more debt because of volume guarantees. AMD overestimated demand when they negotiated those, and they would be further into the red now if GF did not renegotiate a bit last year. CPU margins are still positive where they are selling chips right now...however...failing to meet the minimums on wafer agreements costs them hundreds of millions. In fact, the CPU division is likely running a small profit (read: barely better than breaking even) without those agreement issues...to be completely honest.

3.) Half dozen "better alternatives" that do not get used...and will never unseat it. Primarily because the corporate world runs on windows, and windows requires x86. That will not change anytime soon.

4.) Lack of design wins for a new uarch from a company needing to prove itself does not mean much of anything...wait until roughly around launch time and see what design wins we see. You must also remember...HPC design wins mean very little to AMD right now...server design wins, OEM design wins, and other things are crucial to them.

(2) I agree. They don' need to design an ALU from scratch. They don't need to design a branch predictor from scratch. They don't need to design a x86 decoder from scratch... However, the (micro)architecture Zen is entirely new, from scratch. And will cost much more money to develop Zen than it did cost to develop Piledriver, which was Bulldozer with minor tweaks/fixes.

SOI wafers didn't cost three times more than bulk wafers. Than 14 FinFETs is more costly than SOI is even admitted by Michael Mendicino, a product manager from GlobalFoundries:

Some cost-sensitive mobile chips will avoid the 14 and 10 nm FinFET nodes due to their costs, perhaps for as long as four to six years. SOI may offer an alternative that delivers the performance of 20 nm bulk at a price closer to 28 nm poly, but he noted pressures are driving down all bulk prices, too.

Waffer agreements weren't the cause for the CPU division being in the red numbers. In the first place the GF agreement applies to AMD as a whole not to a division inside a group inside the company. In the second place, the underproduction wafer penalty can be counted as another development/fabrication cost and added to the retail price. The problem is that AMD couldn't increase the price of its products because are not competitive. In the third part now not only the CPU division is in bad shape but the GPU division is as well.

(3) That half dozen of "better alternatives" are used everyday: from phones and printers to supercomputers. Today x86 is a minority ISA.

(4) Lack of design wins for a new uarch from a company needing to prove itself means that they don't have anything competitive to offer. Otherwise, we would hear dozens of customers praising the tech and why selected it.

 

[Ranth]

(4) Nobody cares about HPC. As for being 20%, you do not have any idea that skylake will launch at 4 GHz base clock, and I honestly doubt that is the case to be honest.

Skylake slides have been just leaked and they report 10--20% increase in performance over Broadwell. 😀

I am sorry to say this but Zen CPU looks DOA, with each new leak.

Considering that Broadwell is slower than DevilsCanyon in most benchmarks the % is probably more in the 5-15% (If not lower) so I am not sure that Zen is anymore DOA than before.

Edit: http://www.anandtech.com/show/9320/intel-broadwell-review-i7-5775c-i5-5675c/4 more apparent in professional in the next 2 pages

 

[jimmysmitty]

(4) Nobody cares about HPC. As for being 20%, you do not have any idea that skylake will launch at 4 GHz base clock, and I honestly doubt that is the case to be honest.

Skylake slides have been just leaked and they report 10--20% increase in performance over Broadwell. 😀

I am sorry to say this but Zen CPU looks DOA, with each new leak.

Considering that Broadwell is slower than DevilsCanyon in most benchmarks the % is probably more in the 5-15% (If not lower) so I am not sure that Zen is anymore DOA than before.

Edit: http://www.anandtech.com/show/9320/intel-broadwell-review-i7-5775c-i5-5675c/4 more apparent in professional in the next 2 pages

Yes but you are also looking at a 3.3GHz base with a 3.7GHz boost vs a 4GHz base and 4.4GHz boost (i7 5775C vs i7 4790K). That is a pretty big clock speed difference.

http://www.tomshardware.com/reviews/intel-core-i7-5775c-i5-5675c-broadwell,4169.html

IPC is said to be 5% for Broadwell over Haswell at the same clock speed. So if Skylake is looking at 10-20% IPC over broadwell then that means 15-25% over Haswell. Broadwell was never to be a massive performance jump over Haswell, it is a die shrink after all.

That means Zen has to deal with that same performance gap over Haswell and if the rumors are true and Haswell is the closest we will see it get then they are in the same boat as they currently are, 1-2 generations of Intel performance behind.

 

[Ranth]

But Juan said performance, not IPC?