"Hey, you can still buy twice the processing throughput for half the cost so why would you ever buy Intel again?"

 Its pre-installed on products one frequently purchases.

Eegore wrote on 07/08/19 at 14:49:30:





SLIDE YOUR MONITOR VIEW OVER TO THE RIGHT TO SEE MORE CURRENT MONTHS OF DATA

Please note:   This sales graph ranges back to 3 years ago on the previous generations of AMD and Intel processors.  We expect the relative volumes of new latest and greatest stuff to swing more strongly towards AMD going forward.

You spec what you purchase -- your new cost and the processing ability of what you are providing will be affected by what your specify when you go buy computers.

Intel will have to keep on lowering their part prices to keep selling their stuff at the volumes they are used to moving --- so perhaps you can afford to keep using Intel out into the future ....... mebbe.


Supplying CPU and GPU processors to the entire industry is now a "conglomerated effort" from several vendors now, and we can say that right now AMD is attempting to supply enough customer desirable chipsets to help Intel recover their "processor shortage" due to Intel's "not keeping up with industry lithography standards".

Intel begins squealing and thrashing about madly due to all the AMD abuse they are taking from all sides.   Intel is simply way beyond it's pain threshold from all the picador barbs it has been getting stuck with from everybody on all sides and Intel is just behaving exactly like a desperate pain maddened bull in the bull ring at the moment.

Intel cannot move down to a lower lithography, because as has been stated many times Intel's 10nm isn't anything real at this point in time, being composed of FAILED 10nm lines and some limping 7nm process lines that are only good for putting out large scrap rates and making very little in the way of any progress.

So, Intel now says it is now going to do what it said it couldn't afford to do any further ----- more 14nm cores, 10 big fat Xeon cores to be more exact.   10 Xeon cores with older hyper threading enabled from the get go.   Wonderful, huh?   Yep, it is a mainframe Xeon chipset from 5 years ago -- ain't Intel grand?

Yep, yet another old Intel Xeon mainframe design is coming out of mothballs to join in with Intel's Consumer HPE line ---- with 10 each 14nm Xeon cores swinging the older style Intel Hyper threading.    

What is sad is this old resurrected stuff isn't going to be all that really blazing fast because it never was (although Intel has claimed 5.2 ghz for it before even building any samples) and the required security mitigations for the 5 different sorts of predictive security issues these old Xeon chipsets will be riddled with throughput processing delays and the mitigations may indeed slow the older Xeon mainframe stuff down to a functional net loss in performance instead of offering any real improvement.

But that is assuming the new / old 10 core Intel processor ever actually reach any real production ...... with Intel any of that "far far away in the future" stuff is always suspect.   Many things promised by Intel simply get overcome by events and the next really neat new promised by Intel thing overlaps them and they never seem to arrive.

Intel is faced with 12 core real 7nm AMD chip sets that are shipping right now, and with 16 core real 7nm AMD chip sets that are coming by the end of September of this year.

Let's be generous, since this promised Intel solution doesn't really match up as any form of real competition with AMD's current stuff  ......  let's just say that all that Intel has really just done is to put in an urgent process improvement request for Lisa Su's very best chiplets to get up rated to 5.2 ghz before going into full production on building the next generation of 12 and 16 banger AMD processors.

Lisa Su responds "CAN DO, old buddy" to this Intel urgent process improvement request, and Lisa makes a phone call to the binning department to start saving up the really really good un's in bulk for the Intel requested 12 and 16 core AMD specials.





          Sush now,  you don't know this --- it's a deep deep secret that AMD just got in a 5nm trial sample run of special deep layer burn chiplets in from Samsung just this past week.

Supply.  

Chip supply.

7nm chiplets and 7nm graphics cores.  

5nm chiplets and 5nm graphics cores.

AMD has some potential issues meeting their long term demand for chiplets and graphics cores.

AMD is planning on using both TSMC and Samsung production to meet all these supply demands.    Having enough supply is where Intel fell down and AMD has some well laid out plans not to repeat that issue on their part.

AMD can use Global Foundry to augment for their center chiplet business at 12nm if TSMC can't meet that 12nm center chiplet demand completely (assuming Global actually remains in business long enough to do this as their current owners are busy liquidating their old lines as soon as they can find a buyer).

AMD can use Samsung if TSMC can't meet the total demand for chiplets and for graphics cores.

AMD will task TSMC for whatever 7nm chipsets TSMC can fulfill, as they have volume contracts with TSMC to keep and they will do so.   Just as Apple did, there is a back door clause to these contracts that releases AMD to buy from someone else if TSMC can't make what is needed, but if TSMC is production volume capable, they own that business.

7nm and 6nm are pretty much a known thing now --- it is 5nm and 3nm that is going to get interesting pretty quickly once Apple finishes up their first big 2020 production year at 5nm and TSMC becomes free to fill their mounting general orders at 5nm.

People wondered how AMD was able to abruptly double their on die gaming chipset memory so suddenly for the July Ryzen 3000 roll out.  In fact it was AMD's first released use of the multi-layer direct burn 7nm production machines making two full layers of memory on top of the existing AMD lithography system (a total of 3 layers of direct burned stuff).   3 layers is the best the new modified 7nm scanner lines can do, but ASML's 5nm scanner equipment was built from the get go to be able to burn 14 layers deep.

Apple now has year 2020 5nm designs out now that use 5nm multi-layer (14 layers are available but fewer are actually needed right now for current Apple designs) and you can count on Apple and TSMC working out all those layer bugs during their locked down first year of production.

Just think of how many layers of memory can get dropped on top of a 5nm AMD chipset with 14 direct burn lithography layers available to be used at 5nm .......

https://www.techradar.com/news/intel-admits-it-wont-catch-up-with-amds-7nm-ch...



I predict stuff based on a lot of information from a lot of places --- and I am generally pretty conservative, actually.

Here are some confirmations of my predictions of late about Intel 10nm and it really being the first failed 10nm combined with the first failed 7nm process lines (Intel would never admit to these things, of course).

Swan admits that the 2.7x scaling for 10nm was both too ambitious and too complicated. He also explains how Intel made an error when it “prioritized performance at a time when predictability was really important”.

However, as Swan notes, “The short story is we learned from it, we'll get our 10nm node out this year. Our 7nm node will be out in two years and it will be a 2.0X scaling so back to the historical Moore's Law curve.”

So it looks like we’ll finally see some sorted/partially disabled 10nm chips from Intel this year, while Intel 7nm will have to wait until 2021. That’s a long time considering both AMD and Qualcomm have 7nm chips out in full volume right now.

But will Intel be happy with ceding all the performance advantages and "extra memory on top" that 7nm-5nm brings to its two biggest competitors for that long?
Does Intel really have any choice but to jest suffer through this?

Well, words count I guess.   Bob Swan has admitted publicly now that Intel 10nm will never meet its original goals.   He also admits that Intel 7nm will not fully meet the original Intel 10nm goals either.  

Does it smell like Intel's newest planned stuff is actually a continuation of the old 7nm process path but running looser specs and lower process yield goals that came originally from the "cut back" 10nm stuff?

Does this mean that in 3 years from now (when 14 layer 5nm is shipping in volume from TSMC to AMD and then to you) that Intel will just then be "getting around" to shipping their cut back version of 7nm?

Yep.   It sure does.   Now add a partial year for the inevitable Intel delay over something or other and you got another ~3 years worth~ of slow Intel dog years staring you in the face while you listen to the spew of Intel "product announcements" and "improvement announcements" and solvent rag based "new product lines".

Meanwhile, AMD will improve something for real every six months during that 3 year period that you will spend waiting for Intel to arrive, some will be small AMD tuning improvements and at least one very large improvement will be an across the line lithography shift that will be a really really really big one .....


    AMD 14 layer 5nm is coming inside the next 2 years


I also now predict that 14 layer 5nm will birth some ARM based laptop/Chromebook chipsets earlier than that, with these coming from the phone boys as they are planning on going to 14 layer 5nm for phone uses in 2020, which is a year earlier than AMD currently is thought to be planning on doing it for AMD'd chiplets (just to better thump on Intel, that is).

Please remember, Lisa Su is not slow on meeting all of her competition's moves, and Lisa Su will tend to try to meet/beat the phone boys ARM based chipset efforts no matter that Intel is still stuck in its slowly hardening tar pit and Intel isn't moving anywhere at all.  

Over time, Intel will wind up meaning less and less to the general worldwide competitive actions and general progress levels.

Also realize that NVIDIA is not nearly as pokey as Intel, so AMD will need to shift to 5nm early enough to meet that threat as well.

This does fall in line with rumors that Samsung is sending occasional 5nm and 3nm trial runs of chiplets to AMD for buildup and testing as part of the technical consortium efforts, leading to a new round of Samsung production tune ups and perhaps eventually leading to some possible Samsung early production runs for a new premium line of AMD CPUs.  

Remember, Samsung wishes to become a major TSMC contract foundry competitor again, this has been stated by the CEO of Samsung several times over the last year .......   also remember that Samsung does not hold that there is a whole lot of difference between 5nm and 3nm as it all runs on the same equipment.  

If Samsung or TSMC can work out a FinFet design at 3nm then that is all to the good as that simply means 3nm just gets here in full volume quicker .......  otherwise, Samsung's more complex gate all around nano-tube design has already been sampled both at 5nm and at 3nm sizes and we already know it works.

For the first part of next year TSMC will focus all of their attention on Apple and its needs, leaving a free time segment for some Samsung/AMD research activity to create some positive changes inside that Apple delay period.   And please remember, the Samsung/AMD-Global/IBM consortium really does need a working partner to build up completed finished chipsets off their trial run stuff so they can see if their improvements really work out in the real world and have any real competitive merit or not.

As Global Foundry slowly loses all of their production relevance, Samsung wants to gain that relevance and take some real market share away from TSMC to boot.

A new wave of Interesting Stuff is coming out of Apple at the moment ....... Apple just spent a billion dollars buying all of Intel's modem IP (includes the completed FCC certifications on existing modem designs and the actual modem designs and masks and any dedicated production equipment used to build them).

Apple is intending to build their own modems going forward, and the billion they spent sweeping up the leavings at Intel is dirt cheap compared to the yearly cost of doing business with Qualcomm.

HOWEVER, Apple will likely discover that phone modems are a fussy, tricky thing to do and Intel had failed repeatedly to do a 5G modem for a reason or two, or five or 6 good reasons.

5G IS REALLY REALLY HARD TO DO if you have to avoid Qualcomm patents.   Now, in response to the 5G mess a consortium of companies has proposed new simpler "6G" modem standard that is now being proposed to completely leap frog the troublesome 5G issues by going four times faster using a simpler, much less complex methodology.   A new system, not patented by Qualcomm or by Huawei.

Apple is one of the supporters of this new system, and has gone and hired laid off engineers from Qualcomm and from Intel to support their new Intel modem purchase and the new "6G" modem system that the consortium is proposing.

Trump supports the building of a new USA based "6G" standard as long as all the hardware and software used by the standard is all built by and controlled from inside the USA.

Note that Apple also hired several Qualcomm and ARM chipset architects to go along with this purchase of Intel's old modem capabilities.

AMD has discovered one of the harsh realities that forced Intel to rate their chipsets only off their fastest core ---- in process variation exists, and a new wrinkle has come up about AMD chiplet sorting while using multiple vendors.   Sort and rank some same-same-same chiplets and then build them up into a CPU, test it again and find that the results changed due to the solder dips, heating, etc of the assembly process having changed a few of the chiplet's performance.  

Simply stated, using multiple vendors adds variation as the different producer's chiplets react slightly differently to AMD's assembly processes.

AMD has to deal with this now since they made a big thing out of predictability, and now they are learning that they cannot completely predict the performance of a finished CPU assembly 100% off of knowing what the chiplets did, not when using stocks of chiplets from various vendors and various lots mixed together.    Hey, things CHANGE during chipset assembly, and things also likely shift some over time and heating & use as well.

First solution (and it may be all that is needed) is for AMD's CPU monitoring software (the stuff that auto-overclocks the better AMD processors) to get refined a little bit to continuously ID the slowest cores and simply schedule them dead last when assigning tasks.   This would also cover any ongoing aging changes that take place (and you know this likely happens, right?)

Second solution is to adjust your advertising to reflect what you now know is the real reality.  Your stuff has natural chiplet to chiplet variation, so tell people that and tell them what the allowable pass limits at the factory are.   Offer to take back and replace any chip found exceeding those limits.   And mean it, and do it damme fast.   And possibly offer a discount or a freebie to make it up to the customer who got mislead by your early advertising.

AMD has lots & lots of cores, and some of these many cores really do go slacking all the time anyway due to real workloads that never really use all the cores   --   so, purposefully pick the known slowest cores in each chipset to be the slack takers.

The 'slow core' item seen now mostly only affects the tester folks who are wringing the chipsets out hard 100% on all the cores,  I doubt many real users would see or be affected much at all by a single slower core.

Issue at hand really is your AMD image, you have been honest and forthright all along and that has proven to be a winning image, so this is no time to stop doing that -- not even for a second ........

https://wccftech.com/amd-epyc-rome-7nm-server-cpu-official-launch-64-core-128...



Why is this worth reading?    Because all the hard work that is done on AMD Epyc IS the source of the chiplets for the Consumer chipsets that are doing such severe damage to Intel at the moment.

AMD's improvement cycles start right here as all the throughput improvements and all the neat little tricks that will roll out across the AMD line always start right here at the Epyc chiplet level.  

Yep, oggle this set of Epyc improvements very carefully as they will be showing up in the consumer Ryzen chiplets inside 6 months.  

Mind you, this particular Epyc release is a simple "tuning" release, not a major major lithography shrink like 5nm could be next year .......  assuming Intel actually goes and does anything on their side of the fence "as being competitive" so as to prompt Lisa Su to pull the trigger on that next round of AMD goodies, that is.

AMD has officially announced the launch of their 7nm EPYC Rome processors today, offering higher core count, best in class performance, value & efficiency. Being the first high-performance data center, AI, & HPC centric chips based on TSMC’s bleeding-edge 7nm process node, the Rome lineup takes AMD one step ahead of their Xeon rivals that still utilize 14nm technology.

AMD EPYC Rome Officially Launched – 7nm, 64 Cores, 128 Threads, Higher Clocks, Best-In-Class Performance, Efficiency, and Value
AMD’s 2nd Generation EPYC Rome series is the successor to the first Generation EPYC Naples line of processors which launched two years back. Based on the 7nm Zen 2 core technology which has offered a 15% IPC uplift over the original Zen core, the AMD EPYC Rome CPUs are designed to offer higher performance and better efficiency than their predecessors.


As Lisa likes to do it, several customers and partners joined AMD on the stage to discuss their new AMD EPYC processor offerings:

Google announced it has deployed 2nd Gen AMD EPYC processors in its internal infrastructure production datacenter environment and in late 2019 will support new general-purpose machines powered by 2nd Gen AMD EPYC processors on Google Cloud Compute Engine as well;

Twitter announced it will deploy 2nd Gen AMD EPYC processors across its datacenter infrastructure later this year, reducing TCO by 25%;

Microsoft announced the preview of new Azure virtual machines for general purpose applications, as well as limited previews of cloud-based remote desktops and HPC workloads based on 2nd Gen AMD EPYC processors today;

Hewlett Packard announced the continued support of the AMD EPYC processor family with plans to triple their AMD-based portfolio with a broad range of 2nd Gen AMD EPYC processor-based systems, including the HPE ProLiant DL385 and HPE ProLiant DL325 servers;

Cray announced the Air Force Weather Agency will be using a Cray Shasta system with 2ndGen AMD EPYC processors to provide comprehensive terrestrial and space weather information to the U.S. Air Force and Army;

Lenovo announced new solutions that are specifically built to take advantage of the full range of enhanced capabilities found in the 2nd Gen AMD EPYC processors. Available today, the ThinkSystem SR655 and SR635 are ideal solutions for use cases such as video infrastructure, virtualization, software-defined storage and more, with exceptional energy efficiency;

Dell announced the upcoming availability of newly designed servers optimized for 2nd Gen AMD EPYC processors;

VMware and AMD announced a close collaboration to deliver support for new security and other features of the high-performance 2ndGen AMD EPYC processors within VMware vSphere.

Zen 2 doesn’t only offer higher performance but due to a smaller manufacturing process, the resultant die size has allowed AMD to cram twice the number of cores and threads on the EPYC 7002 CPUs while retaining higher out of box clock speeds.

Following are some of the salient "industry firsts" features of the 7nm EPYC Rome processors:

1)  Built on 7nm advanced process technology – the best the industry currently has to offer, allowing for denser compute capabilities with lower power consumption.  

2)  The world’s first 64 core data center CPU, built using Zen 2 high-performance cores and AMD’s innovative Chiplet architecture

3)  The world’s first mainstream PCIe Gen 4.0 data center CPU with a bandwidth of up to 64GB/s, twice of PCIe Gen 3.0

4)  Embedded hardware security protection to help defend your CPU, applications, and data that requires no mitigations in the OS system. 
(means a 25% throughput upper when compared to Intel processors)

AMD has made significant changes to their CPU architecture which help deliver twice the throughput of their first-generation Zen architecture. The major points include an entirely redesigned execution pipeline, major floating-point advances which doubled the floating-point registers to 256-bit and double bandwidth for load/store units. One of the key upgrades for Zen 2 is the doubling of the core density which means we are now looking at 2x the core count for each core complex (CCX).

Improved Execution Pipeline
Doubled Floating Point (256-bit) and Load/Store (Doubled Bandwidth)
Doubled Core Density
Half the Energy Per Operation
Improved Branch Prediction
Better Instruction Pre-Fetching
Re-Optimized Instruction Cache
Larger Op Cache
Increased Dispatch / Retire Bandwidth
Maintaining High Throughput for All Modes



This also serves as a long laundry list of major companies that are NOT BUYING NEAR AS MUCH STUFF FROM INTEL ANY LONGER for their data center uses.

Nobody simply throws away all their old mainframe stuff and simply replaces all their servers at a lick, but for AMD to pick up 20% in a year's time says they are selling in the vast majority of the new expansion stuff (out of that 25% of the grand total yearly potential expansion mainframe ball court --- that is the playing field that AMD just picked up their 20% market share out of --- so you can say AMD is picking up roughly 80% of the new items out of the mainframe/rackspace expansion new installation sales).

Should Intel be worried?       Yep, Intel should be worried.

If AMD can keep showing improvement levels anywhere near this, even the most loyal of Intel "big iron" fans will eventually have to start buying AMD new products because it is simply much better stuff and it is a lot cheaper to buy and a lot cheaper to own year on year.