Register a SA Forums Account here!
JOINING THE SA FORUMS WILL REMOVE THIS BIG AD, THE ANNOYING UNDERLINED ADS, AND STUPID INTERSTITIAL ADS!!!

You can: log in, read the tech support FAQ, or request your lost password. This dumb message (and those ads) will appear on every screen until you register! Get rid of this crap by registering your own SA Forums Account and joining roughly 150,000 Goons, for the one-time price of $9.95! We charge money because it costs us money per month for bills, and since we don't believe in showing ads to our users, we try to make the money back through forum registrations.
 
  • Post
  • Reply
GRINDCORE MEGGIDO
Feb 28, 1985


Cannonlake: pinch of salt wccftech point to 256KB / core l2. Which isn't really a big surprise if a bit disappointing.

It might be interesting to compare the 6c HEDT to 6c clake. I don't care about frames except frametime.

Adbot
ADBOT LOVES YOU

Paul MaudDib
May 2, 2006

"Tell me of your home world, Usul"


DrDork posted:

Overall cache hasn't really gone down by too much, though. A 5820k, for example, had 6x256k L2 + 15MB L3 = 16.5MB overall to back 6 cores. The new 7800X has 6x1MB L2 + 8.25MB L3 = 14.25MB to back 6 cores. Similar for the 5960k at 22MB total vs the 7820k at 19MB for 8 cores each. So it's more of a shift from L3 to L2 cache than a huge drop in cache. Considering that L2 undoubtedly takes up more die space than L3, the drop of 2.25MB/3MB might be the sacrifice needed to keep the total cache die sizes similar.

It also looks like The 6/8/10/12c's will be sticking with quad-channel 2666 memory, while the 4c chips are stuck with dual-channel. Considering X99 already defacto supports quad-channel upwards of 3200, it doesn't seem like much of an improvement.

All things equal, L2 cache may be more of a "bang for your buck" in terms of performance than L3 is, such that it makes sense to go for an overall smaller amount of cache but more heavily tilted towards L2. After all, the higher-tier caches are usually faster and physically closer.

Or maybe they read this thread and they're putting a whole bunch of eDRAM on as L4 on that isn't listed in the charts yet. Enjoy Christmas In July AMD

The Kaby Lake-X processors here are explicitly just the small processors being ported onto the new socket, so I'm not sure why people are astonished that a 2-channel die only has 2 channels. Same thing for people who are complaining about it only having 16 lanes (not here). Putting it on a different socket doesn't magically add hardware that isn't on the die.

It's a "low-end" option so that people who want to buy a 4C8T can buy an X299 motherboard and leave themselves an upgrade path. That's it. Nothing that's not on desktop or server Kaby Lake will be on the Kaby Lake-X chips. It's the exact same silicon.

If the official spec is actually moving up to 2666 instead of 2133... that's actually still upward movement, since officially only 2133 is supported. 3200 is not officially supported and going past 3000 is not super reliable. On the other hand, it's not usually a big deal either since quad-channel has excessive amounts of bandwidth for desktop users anyway.

DrDork
Dec 29, 2003
commanding officer of the Army of Dorkness

Paul MaudDib posted:

All things equal, L2 cache may be more of a "bang for your buck" in terms of performance than L3 is, such that it makes sense to go for an overall smaller amount of cache but more heavily tilted towards L2. After all, the higher-tier caches are usually faster and physically closer.
Agreed; L2 is certainly faster than L3, so presumably Intel figures that the large Haswell-E L3 caches were not being efficiently used compared to what they could get out of larger L2 caches. The real interesting thing, to me, will be to see if that trend continues with their other future lineups, or if the larger L2 is more a response to usage scenarios more commonly featured in HPC applications than your more pedestrian desktop/server stuff.

Paul MaudDib posted:

The Kaby Lake-X processors here are explicitly just the small processors being ported onto the new socket, so I'm not sure why people are astonished that a 2-channel die only has 2 channels.
It's not really surprising in that it's a Kabby derivative, you're right. The surprising thing (to me) is that they'd bother with this option at all since unless the larger L2 cache really improves performance (and doesn't hilariously increase cost), they don't bring anything to the table that you can't already get with normal Kabby Lake, except being on the X299 boards.

Paul MaudDib posted:

It's a "low-end" option so that people who want to buy a 4C8T can buy an X299 motherboard and leave themselves an upgrade path. That's it. Nothing that's not on desktop or server Kaby Lake will be on the Kaby Lake-X chips. It's the exact same silicon.
But who would actually do that? You're (presumably) paying a premium for this "new" chip, plus a premium for X299. In exchange you get...nothing? The details we've gotten for the X299 chipset really doesn't seem to be much of a move from the currently available 200-series, and as you say, it's the same chip on a different socket. So you pay all that extra money in order to save a little hassle down the road? Finance-wise you'd probably be better off getting a Kabby + Z270 and then simply selling both when you're ready to move up to Skylake-X, vs just selling the chip and keeping the motherboard.

And if you're the type of person/business where selling things off isn't worth the effort, you probably aren't going to give much of a crap about that sort of upgrade path to begin with.

Paul MaudDib posted:

If the official spec is actually moving up to 2666 instead of 2133... that's actually still upward movement, since officially only 2133 is supported.
It is, but only for 1DPC. If you want 2DPC, it drops to 2400. I'll be curious to see what the real-world implications of that turns out to be.

DrDork fucked around with this message at 02:17 on May 15, 2017

PC LOAD LETTER
May 23, 2005
WTF?!

Slippery Tilde

DrDork posted:

Overall cache hasn't really gone down by too much, though.
Any decrease at all seems like a big deal to me. Normally the cache is one of those things I'd expect to get bigger not smaller. Especially for something that is targeted at server work loads. edit: AMD's Threadripper or Naples cache size (ie. 32-64MB) is closer to what I'd expected for Intel to have done here but Intel has also always been ahead of pretty much everyone when it comes to cache design too so maybe they've got a huge efficiency advantage of some sort going on too.

DrDork posted:

So it's more of a shift from L3 to L2 cache than a huge drop in cache. Considering that L2 undoubtedly takes up more die space than L3, the drop of 2.25MB/3MB might be the sacrifice needed to keep the total cache die sizes similar.
Yeah that is a good point, didn't think of that.

DrDork posted:

It also looks like The 6/8/10/12c's will be sticking with quad-channel 2666 memory, while the 4c chips are stuck with dual-channel. Considering X99 already defacto supports quad-channel upwards of 3200, it doesn't seem like much of an improvement.
Ah crap, I guess only the super high end stuff with Purley will be using 6 channels? Yeah if they're only sticking with quad channel DDR4 2666 by default (I'm sure with overclocking it'll do 3200+ DDR4 too) then that is a minor bump from DDR4 2400.

edit: \/\/\/\/\/\/\/AFAIK the higher DDR4 speeds possible with Intel's memory controller have more to do with allowing T2 command rates and much looser timings than AMD's currently, the newer or upcoming AMD AGESA BIOS updates should help lots here\/\/\/\/\/\/

PC LOAD LETTER fucked around with this message at 07:27 on May 15, 2017

BurritoJustice
Oct 9, 2012



If it carries over the better DDR4 controller of mainstream Skylake we'll probably see DDR4-3866+. Haswell-E got to 3466 with quad channel on the highest end boards so it wouldn't be too unreasonable of a jump.

Scarecow
May 20, 2008

3200mhz RAM is literally the Devil. Literally.


Lipstick Apathy

Meanwhile at amd (salt etc)
http://wccftech.com/amd-ryzen-9-lineup-threadripper/

Anime Schoolgirl
Nov 28, 2002






the only thing that seems too good to be true is that they're actually using the threadripper trademark

Nomyth
Mar 15, 2013

And if a Nyto get a attitude
Pop it like it's hot
Pop it like it's hot
Pop it like it's hot


I'm the rumored model differentiation numbers that sound like years from the 20th century

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!

Oh dear, 12C/24T Ryzen for same or less than the octocore Skylake-X, that'd be hilarious. The overclocking headroom is a little meh, but it'll be interesting to see what the quad channel stuff will do. Assuming they fix that poo poo with the CCX.

Malcolm XML
Aug 8, 2009

I always knew it would end like this.


EMIB when

I want edram on the cheap!!!

PerrineClostermann
Dec 15, 2012

by FactsAreUseless


Drakhoran
Oct 21, 2012



Fuad Abazovic can now confirm that Kyle Bennet was not bullshitting about Intel licensing graphics tech from AMD. At least he claims he can. He doesn't say much about how he knows so...

Are we sure this isn't part of some Reality TV show?

eames
May 9, 2009





http://cpu.userbenchmark.com/SpeedT...telR-CPU-0000--

http://cpu.userbenchmark.com/SpeedT...telR-CPU-0000--




http://www.tweaktown.com/news/57568...-24t/index.html

eames fucked around with this message at 08:04 on May 16, 2017

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!

Weird dip in performance for the 7900X.

BIG HEADLINE
Jun 13, 2006

Make your move...'cause mine's gonna be ugly.

Combat Pretzel posted:

Weird dip in performance for the 7900X.

Almost assuredly an ES chip on a board with not-quite-finalized drivers and BIOS, but yeah.

eames
May 9, 2009



Yeah, looking at the benchmarks the 7920X ran on a single 8GB stick of DDR4-2113 (9GB/s) and the 7900X benchmark is 3 months old.

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!

Oh this is exciting and annoying at the same time. Either tons of cheap cores for the media/render and compilation stuff, or higher IPC (and less cores) and eventually better overclock for gaming and CAD.

eames
May 9, 2009



It's a slow day at work so I overlaid the two cache latency graphs. This is pretty deep into territory but whatever, maybe one of you smart guys can explain what we are looking at.

orange = 10C 7900X
blue = 12C 7920X

Only registered members can see post attachments!

DrDork
Dec 29, 2003
commanding officer of the Army of Dorkness

Well, both chips have 1MB L2 cache per core, and presumably are close on clocks, so they should have very similar cache performance up to the 1MB mark, as your chart shows.

The 7900x has 13.75MB L3 cache, while the 7920x has 16.5MB, which explains why the 7900x takes the biggest hit at 16MB: right where it's L3 cache is no longer big enough, but the 7920x's still is. Above that, more of the same: the 7920x can leverage that larger L3 cache for a bit better performance.

No idea what causes the 2-8MB performance gap, though.

e; as far as conclusions, it's the same as what you'd expect from the outset: moar cache is moar faster when you're using big data sets.

Potato Salad
Oct 23, 2014

Nobody Cares




This page is kinda providing further indication that we might be better served by a vendor-neutral CPU thread. There's a large amount of overlap between the AMD and Intel threads.

EdEddnEddy
Apr 5, 2012





At this point I tend to agree. Now that they are starting to trade blows yet release little else exciting, combining the two might not be a bad idea.

priznat
Jul 7, 2009

Let's get drunk and kiss each other all night.

Also perhaps talk about server CPUs like the ARM push etc?

Or just Consumer PC stuff, perhaps one of the enterprise threads instead.

GRINDCORE MEGGIDO
Feb 28, 1985


priznat posted:

Also perhaps talk about server CPUs like the ARM push etc?

Or just Consumer PC stuff, perhaps one of the enterprise threads instead.

AMD / Intel / ARM CPU and Platfrom Megathread

But if has to be spelled platfrom or there's no point.

priznat
Jul 7, 2009

Let's get drunk and kiss each other all night.

GRINDCORE MEGGIDO posted:

AMD / Intel / ARM CPU and Platfrom Megathread

But if has to be spelled platfrom or there's no point.



I want to nerd out on new and exciting server architectures etc

Paul MaudDib
May 2, 2006

"Tell me of your home world, Usul"




There is an extreme amount of crossover between the two threads, and GPU stuff goes in its own thread anyways.

GRINDCORE MEGGIDO
Feb 28, 1985


I wonder what kind of an op that would need?

mewse
May 2, 2006




GRINDCORE MEGGIDO posted:

I wonder what kind of an op that would need?

Quick and dirty hierarchy chart a la Tom's Hardware would be cool

GRINDCORE MEGGIDO
Feb 28, 1985


Can mods actually merge threads, or do we just create a new op?

Gyrotica
Nov 26, 2012

Grafted to machines your builders did not understand.

GRINDCORE MEGGIDO posted:

I wonder what kind of an op that would need?

One that updated after August 31st, 2014.

Rastor
Jun 2, 2001



The AMD thread has a fresh OP, just do a pull request with the Intel info.

Boiled Water
Apr 5, 2006

YOU ARE A BRAIN
IN A BUNKER


GRINDCORE MEGGIDO posted:

AMD / Intel / ARM CPU and Platfrom MegaTHREADRIPPER

Watermelon Daiquiri
Jul 10, 2010




Boiled Water posted:

AMD / Intel / ARM CPU and Platfrom MegaEPYCTHREADRIPPER

GRINDCORE MEGGIDO
Feb 28, 1985



Oooo

How do we do this, shall I put an op together?
Is anyone not into this?

SamDabbers
May 26, 2003



I object! Just to be contrary.

Sinestro
Oct 31, 2010

The perfect day needs the perfect set of wheels.


I feel that it would be disrespectful to the time that I, a Valued Member of the esteemed Something Awful community has put into the AMD thread...

Nah, it's a good idea.

movax
Aug 30, 2008



GRINDCORE MEGGIDO posted:

Can mods actually merge threads, or do we just create a new op?

We should probably just do a new OP -- I admit I don't check into the AMD thread that often. I'll swing over there and see how it's going -- a vendor neutral thread that includes some discussion on ARM/SPARC/POWER/etc could be cool and nerdy as gently caress.

Feel free to sling drafts of your OP in this thread for now though -- when it's ready, we can unleash hell nerd.

e: I forgot RISC-V, the future.

GRINDCORE MEGGIDO
Feb 28, 1985


Will do. It'll need some redrafting but I'm happy to. I'll get on that and post it in next couple of days.

BIG HEADLINE
Jun 13, 2006

Make your move...'cause mine's gonna be ugly.

Before that happens, I do have to admit that the first look at the actual CPU has me a touch worried: http://www.guru3d.com/news-story/am...-confirmed.html

Hopefully they've fixed the interconnect speed. Intel might charge a fortune, but at least they make unitary dies. I'm also wondering if there might be a thermal issue since there'll be four hot spots along the periphery of any contact plate with a 'void' in the middle.

Watermelon Daiquiri
Jul 10, 2010




BIG HEADLINE posted:

Before that happens, I do have to admit that the first look at the actual CPU has me a touch worried: http://www.guru3d.com/news-story/am...-confirmed.html

Hopefully they've fixed the interconnect speed. Intel might charge a fortune, but at least they make unitary dies. I'm also wondering if there might be a thermal issue since there'll be four hot spots along the periphery of any contact plate with a 'void' in the middle.

you'd just do a 'ring' rather than a 'plate'

Adbot
ADBOT LOVES YOU

priznat
Jul 7, 2009

Let's get drunk and kiss each other all night.

I do like that no one cares even in the Intel thread to mention the discontinuation of Itanium that was announced recently.

Rest in piss, Itanium. Owned by amd64. (A rare win)

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply