Intel Server Technology Falls Behind AMD as 10nm Struggles Continue

[cageymaru]

Charlie over at Semiaccurate has an interesting article about how Intel's server future relies on 10nm technology that is now 4 years delayed, raising server CPU prices by $7,000 for what is essentially the same chip; just clocked higher and drawing more power, and how it will combat AMD's 7nm Rome processor and successors. The article is chock full of interesting tidbits such as the estimated speeds of the new AMD and Intel server chips and musings about internal documents from Intel.

It won't be a fair fight. Why? Rome will beat Cascade by more than 50% in per-socket performance, likely tie or win on a single threaded basis, and more than double the Cascade's core count. Please note that by more than 50% we don't mean a little more, we mean a lot more, think abusive rather than hair's width margins.

 

[socK]

Around that point in time, we'll be in the market for a hardware refresh.

Should make things a bit more interesting.

 

[juanrga]

Game over!

 

[Time2Kill]

Nice that AMD is competing again. Haven't ran AMD since my K7 Thunderbird. I remember using a pencil to close the traces so I could overclock it.

Can't imagine where we would be if AMD remained competitive this entire time. Hopefully they can start competing with Nvidia too. Been too long since I felt the need to upgrade.

 

[smokin9]

Im a little behind on CPUS... actually i usually only really start to research when I doing a new build but is AMD already on 7nm, and Intel is still figuring out 10nm??

 

[IdiotInCharge]

Im a little behind on CPUS... actually i usually only really start to research when I doing a new build but is AMD already on 7nm, and Intel is still figuring out 10nm??

Neither are shipping in volume, but TSMC/GloFo/Samsung appear to be ahead with 7nm. Samsung is the real clincher here; they're the only one that can really compete with Intel in terms of technology and volume.

As for 10nm at Intel... I have no clue. Gotta be kicking themselves for taking the route they did, they're something like four years behind, and supposedly have had a new, higher-IPC architecture waiting for most of that.

 

[nutzo]

I can get a few more years out of the current servers at the office with some more Ram and SSD upgrades.

Hopefully by then Intel will have their mess figured out, or AMD will be well supported due to the large volume of server CPUs they are shipping.
Even better they will be in price war and we'll have cheap 16/24 core CPU's

The one issue I have with AMD is that they need more cores to match Intel's current CPU performance
SQL is licensed per core, so needed 20% more cores for the same performance becomes an expensive issue.

 

[Iratus]

I'm bullish and hopeful for amd at the moment but it's Charlie... He's an anti-intel loon. Famously.

 

[knowom]

Between the node difference Intel's struggle's with their own at 10nm and AMD's multi chip glued together solution leapfrogging Intel Charlie could be close to accurate. Intel has some competition on it's hands not even factoring in the nightmare it has with Nvidia as well.

 

[LBJM]

The number one reason Intel has continued support over AMD is that Intel chips retain instruction set extensions while AMD with their chips does not.


"AMD has a different way of dealing with instruction set extensions than Intel. AMD keeps adding new instructions and remove them again if they fail to gain popularity, while Intel keeps supporting even the most obscure and useless undocumented instructions dating back to the first 8086. AMD introduced the FMA4 and XOP instruction set extensions with Bulldozer, and some not very useful extensions called TBM with Piledriver. Now they are dropping all these again. XOP and TBM are no longer supported in Ryzen. FMA4 is not officially supported on Ryzen, but I found that the FMA4 instructions actually work correctly on Ryzen, even though the CPUID instruction says that FMA4 is not supported."

https://www.agner.org/optimize/blog/read.php?i=838

 

[naib]

FMA4 doesn't work on Ryzen... Us Gentoo users who jumped onto Ryzen had a field day figuring out segfaults while using gcc5.x (didnt including specific flags)

 

[chithanh]

FMA4 appears to work correctly on Ryzen. But it is unknown whether there are any bugs or side effects that are triggered by this instruction. About the other extensions like 3dnow, XOP, TBM, etc. these were almost always used in code that could detect the presence of CPU support at runtime, so nothing of value was lost. Gentoo folks had to recompile their system on CPU upgrade, which is something they need to do occasionally anyway (or when switching between Intel and AMD, or going from Intel Atom to Intel Core, or ...).

Also I believe that at least in the case of 3DNow!, these instruction were trapped / emulated (in microcode?) or somesuch. That means software which didn't check of the presence of the instruction set would still run, just slower.

 

[juanrga]

Im a little behind on CPUS... actually i usually only really start to research when I doing a new build but is AMD already on 7nm, and Intel is still figuring out 10nm??

Those days node names are more marketing labels than anything else. For logic cells:

For HD SRAM cells:

Intel 10nm = 0.0312 um2

Glofo 7nm = 0.0269 um2

TSMC 7 nm = 0.027 um2

 

[R_Type]

One day it'll all come out the actual nitty-gritty what went wrong with 10nm and it'll be a stonking read. Especially this year when they thought it was ready, started production and some giant showstopper didn't reveal itself until the veeeery last second.

 

[OFaceSIG]

Around that point in time, we'll be in the market for a hardware refresh.

Should make things a bit more interesting.

The thing to remember is even if the processors come out it'll be months before Dell, HP, Supermicro and all the whitebox OEM/ODMs actually have servers running the chips.

My company works exclusively with Dell and it takes months before new chips actually make it into volume production servers.

 

[akaliel]

No point in reading the article past "Charlie at Semi Accurate".

 

[OutOfPhase]

One day it'll all come out the actual nitty-gritty what went wrong with 10nm and it'll be a stonking read. Especially this year when they thought it was ready, started production and some giant showstopper didn't reveal itself until the veeeery last second.

For most people, it would really not be a great read. Device physics is already a bit nutty and complicated, and adding on all the manufacturing challenges to the theoretical challenges results in one of the most boring and obtuse episodes of Murder She Wrote ever.

 

[vertices]

Nice that AMD is competing again. Haven't ran AMD since my K7 Thunderbird. I remember using a pencil to close the traces so I could overclock it.

Can't imagine where we would be if AMD remained competitive this entire time. Hopefully they can start competing with Nvidia too. Been too long since I felt the need to upgrade.

Haha, nice! I haven't ran anything AMD in years. I do miss those old hardware modding days. I remember soldering those SMD resistors on the original Athlon to go from 500mhz up to 800mhz. Like this: https://www.tomshardware.com/reviews/overclocking-amd,126.html

Had to go to a local metal working shop to have them fabricate a waterblock for me, couldn't buy those then! Then of course mounting all the potentiometers and such to early Radeons. Good times. Miss those days. It'd be cool to see AMD competitive again.

 

[R_Type]

For most people, it would really not be a great read. Device physics is already a bit nutty and complicated, and adding on all the manufacturing challenges to the theoretical challenges results in one of the most boring and obtuse episodes of Murder She Wrote ever.

Yes well there is that haha!

Alright alright for those interested it will be highly beard strokingly illuminating.