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 $3,400 per month for bandwidth bills alone, and since we don't believe in shoving popup ads to our registered users, we try to make the money back through forum registrations.
«245 »
  • Post
  • Reply
NewFatMike
Jun 11, 2015



Thatís more or less exactly what Iím hoping for. If I can put something together thatíll run Overwatch 1080p, medium 30+FPS and other stuff, Iíll pair it with that lovely Asus FreeSync monitor that has LFC, and call it a day with an SFF system for like $650 thatíll do all them video games things and Netflix/what have you.

I reconfirmed that B350 motherboards will let you over lock, so I just need to find one that has a DP output so I can do the experiments.

Adbot
ADBOT LOVES YOU

Craptacular!
Jul 9, 2001

Fuck the DH


Yes B350 will let you OC, but Iíve read that the VRMs or something component-wise were less than ideal on some/many boards.

NewFatMike
Jun 11, 2015



Ah did not know. I mean it's a pretty weird use case since it needs DP anyway, so it's extra research. I'll keep a lookout. Thanks!

eames
May 9, 2009



So yeah, I bought a Coffee Lake CPU yesterday because It didnít look like the AMD/KVM/NPT bug ever gets fixed and then this happens.

https://lists.linuxfoundation.org/p...ber/024822.html

NewFatMike
Jun 11, 2015



eames posted:

So yeah, I bought a Coffee Lake CPU yesterday because It didnít look like the AMD/KVM/NPT bug ever gets fixed and then this happens.

https://lists.linuxfoundation.org/p...ber/024822.html



Papa bless.

For folks who want a little more info/a full article on AMD increasing R&D spending (like I did):

https://overclock3d.net/news/cpu_ma...the_past_year/1
Not much special in the way of information, but interesting nonetheless.

Combat Pretzel
Jun 23, 2004

No, seriously... what kurds?!

One less blocker for Threadripper.

SwissArmyDruid
Feb 14, 2014



eames posted:

So yeah, I bought a Coffee Lake CPU yesterday because It didnít look like the AMD/KVM/NPT bug ever gets fixed and then this happens.

https://lists.linuxfoundation.org/p...ber/024822.html



rip u

Is it too late to cancel? =P

SoftNum
Mar 31, 2011



eames posted:

So yeah, I bought a Coffee Lake CPU yesterday because It didnít look like the AMD/KVM/NPT bug ever gets fixed and then this happens.

https://lists.linuxfoundation.org/p...ber/024822.html



RIP though I'm glad you posted this I was just about to shop my 1080 around for a Vega so I could use Xen instead. Noice.

NewFatMike
Jun 11, 2015



Combat Pretzel posted:

One less blocker for Threadripper.

Speaking of, 1950X got price dropped to $880:

https://www.techspot.com/news/71582...-1950x-880.html

Craptacular!
Jul 9, 2001

Fuck the DH


eames posted:

So yeah, I bought a Coffee Lake CPU yesterday because It didnít look like the AMD/KVM/NPT bug ever gets fixed and then this happens.

https://lists.linuxfoundation.org/p...ber/024822.html



Explain this to a person who doesnít really know how operating systems work

Xae
Jan 19, 2005



Craptacular! posted:

Explain this to a person who doesnít really know how operating systems work

Hardware can send signals to the main processors called Interrupt Requests or "IRQs". These signals do a poo poo load of work connecting the GPU and CPU.

It looks like the signals were getting ignored because some stuff wasn't properly implemented.

dont be mean to me
May 2, 2007

I'm interplanetary, bitch
Let's go to Mars




Xae posted:

Hardware can send signals to the main processors called Interrupt Requests or "IRQs". These signals do a poo poo load of work connecting the GPU and CPU.

It looks like the signals were getting ignored because some stuff wasn't properly implemented.

Finish the explanation: The problem was with the software, and lying dormant for around a decade - it could not have been AMD's architecture at fault.

Craptacular!
Jul 9, 2001

Fuck the DH


Xae posted:

Hardware can send signals to the main processors called Interrupt Requests or "IRQs". These signals do a poo poo load of work connecting the GPU and CPU.

It looks like the signals were getting ignored because some stuff wasn't properly implemented.

Oh, I remember IRQ conflicts in the olden days.

I guess my bigger question is, given the source, is this a Linux specific thing or relevant in Windows, too.

Paul MaudDib
May 2, 2006

"Tell me of your home world, Usul"


Craptacular! posted:

Explain this to a person who doesnít really know how operating systems work

When you're hosting VMs and they need access to raw hardware, there's a couple ways you can do it. One is called SR-IOV and basically involves guests "sharing" a PCIe device. There are a lot of problems with this, for starters the device drivers need to support it, there's some obvious security problems on certain devices/tasks (GPUs don't wipe memory before reallocating it, etc), but not least that AMD and NVIDIA won't let you do it on consumer GPUs (only compute-oriented GPUs that cost 10x as much at a given performance level). The other way is called "PCIe passthrough" - basically you pass the whole PCIe device through to the guest OS and let it handle it, so the guest OS simply sees an attached GPU just like if it wasn't virtualized. This is preferable in some ways, less preferable in others (after all you can't share a device between multiple guests). But it takes native hardware support, the CPU has to say "ah yes PCIe device 0:1:0, that belongs to VM #3" and pass the IRQs along appropriately (which is part of what VT-D and AMD's equivalent do - "directed" PCIe virtualization for guests).

KVM has had some long-standing bugs when doing this with AMD processors. The fact that this is fixed is a Big loving Deal for cloud hosts like Google and Amazon (and frankly I wouldn't be surprised at all if you told me their coders had a hand in this).

edit:

Craptacular! posted:

Oh, I remember IRQ conflicts in the olden days.

I guess my bigger question is, given the source, is this a Linux specific thing or relevant in Windows, too.

KVM is the Linux "Kernel-based Virtual Machine" but this bug would also have applied to anything else you ported KVM to. IIRC Xen (another hypervisor) never had this issue. It was an implementation problem with KVM - but a lot of cloud infrastructure runs on KVM (Google's iirc).

Paul MaudDib fucked around with this message at Oct 26, 2017 around 01:15

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Paul MaudDib posted:

KVM has had some long-standing bugs when doing this with AMD processors. The fact that this is fixed is a Big loving Deal for cloud hosts like Google and Amazon (and frankly I wouldn't be surprised at all if you told me their coders had a hand in this).

Could be a reason why moving EPYC shipments was overall slower for Q3 than expected?

NewFatMike posted:

Speaking of, 1950X got price dropped to $880:

https://www.techspot.com/news/71582...-1950x-880.html

Someone was telling me that it's possible AMD is doing this in prepr for Pinnacle based Threadripper processors, but besides getting power consumption under further control I don't see the advantage? Could AMD do a 24 core TR by having it run in single channel per die?

Paul MaudDib
May 2, 2006

"Tell me of your home world, Usul"


FaustianQ posted:

Could be a reason why moving EPYC shipments was overall slower for Q3 than expected?

Virtualization is a big deal in the cloud. No one running at scale handles individual machines. My guess would be that everything is run on a bare-metal hypervisor that PXE boots an image and auto-joins a service cloud. Everything is run virtualized and PCIe passthrough bugs are no-poo poo a big deal for that market. Passing something like a GPU device or SAS controller through to a guest is the easiest way to handle it.

I've never handled an Epyc, nor seen one tested. I'd be real, real curious to see if those segfault too. That came out relatively late - it's possible B2 may already have been taped out at that point. That one is another big concern for me both for the consumer and server markets. OEMs aren't going to like the support calls that happen if there's a bad MS or game patch that starts segfaulting on some popular title. poo poo starts mattering when Raven Ridge hits the OEM market, let alone in the server market.

Paul MaudDib
May 2, 2006

"Tell me of your home world, Usul"


Opinion:

Lisa Su via WCCFtech transcription

quote:

ďOur goal is to be very competitive in terms of our long-term roadmap. If you look at the foundryís 7nm roadmap compared to some of the other technologies out there itís actually really competitive. I think the gap between the foundry roadmap and the Intel roadmap has gotten a lot closer. Our goal is to be aggressive with 7nm technology. We will be doing tape outs later this year and as we get closer to production will give more insights there. But the idea is to be more competitive throughout the portfolio.Ē

Are those two sentences related, and if so what are the odds of it not being bullshit? Is GloFo in place to do some early pre-risk runs of engineering masks or something similar that might let them drop Ryzen2 late 2018?

Paul MaudDib fucked around with this message at Oct 26, 2017 around 03:03

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Isn't that setting up for another round of "Ryzen vs Kaby" comparison, except Intel will still be on 14nm IIRC? I keep hearing how Intel keeps pushing back 10nm and I'm getting visions of AMD leapfrogging them for 6 months which can be outright devastating.

Drakhoran
Oct 21, 2012



Paul MaudDib posted:

Opinion:

Lisa Su via WCCFtech transcription


Are those two sentences related, and if so what are the odds of it not being bullshit? Is GloFo in place to do some early pre-risk runs of engineering masks or something similar that might let them drop Ryzen2 late 2018?

According to this article :

quote:

GlobalFoundries has been processing test wafers using 7 nm process technology for clients for several quarters now. The companyís customers are already working on chips that will be made using 7 nm DUV process technology, and the company intends to start risk production of such ICs early in 2018.

So, maybe?

priznat
Jul 7, 2009

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

Oddly the suppliers I deal with at work can find Epyc CPUs but have had no luck getting any motherboards, Supermicro is only selling complete systems right now. With the pricey as hell 7401s in there so around $6k for a dual system (Still, 48 cores )

Anime Schoolgirl
Nov 28, 2002

~perfect~
battlebrother





FaustianQ posted:

Isn't that setting up for another round of "Ryzen vs Kaby" comparison, except Intel will still be on 14nm IIRC? I keep hearing how Intel keeps pushing back 10nm and I'm getting visions of AMD leapfrogging them for 6 months which can be outright devastating.
AMD coming out with a 7-actually-10nm product that performs before Intel can push out their 10nm laptop chips would be quite hilarious but given how tapeouts work we won't see such chips until Q1 2019 +/- 3 months, unless they want to do another crash test dummy run and push out barely-ready-for-production chips for consumers like they had to for Zeppelin.

priznat posted:

Oddly the suppliers I deal with at work can find Epyc CPUs but have had no luck getting any motherboards, Supermicro is only selling complete systems right now. With the pricey as hell 7401s in there so around $6k for a dual system (Still, 48 cores )
A non-rackmount non-proprietary motherboard will have a hell of a time tracing 3700-4094 working pins, imagine how thick the PCBs will be on EATX, to say nothing of anything smaller...

Anime Schoolgirl fucked around with this message at Oct 26, 2017 around 06:27

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Anime Schoolgirl posted:

AMD coming out with a 7-actually-10nm product that performs before Intel can push out their 10nm laptop chips would be quite hilarious but given how tapeouts work we won't see such chips until Q1 2019 +/- 3 months, unless they want to do another crash test dummy run and push out barely-ready-for-production chips for consumers like they had to for Zeppelin.

Look man if I am stuck in this dumpster fire reality made only of the wrong choices on the probability tree I demand it at least entertain me.

PC LOAD LETTER
May 23, 2005
WTF?!

Slippery Tilde

Craptacular! posted:

Yes B350 will let you OC, but Iíve read that the VRMs or something component-wise were less than ideal on some/many boards.

NewFatMike posted:

Ah did not know. I mean it's a pretty weird use case since it needs DP anyway, so it's extra research. I'll keep a lookout. Thanks!
Unless you're trying to OC a 8 core Ryzen or really pushing a 6 core as hard as possible even the cheap mobos are supposed to be fine for OC'ing the 4 core APU's that are coming.

Buildzoid did a very good video on this shortly after launch so its a bit old now but not much has changed with the cheap mobo's or their VRM's since then.

Think this is it:

https://www.youtube.com/watch?v=ZGrxhf_xZWI

No one really knows about OC'ing the built in iGPU. That has its own power plane that isn't really being used much by current Ryzens since none of them have iGPU's. vSoC power delivery will matter lots for that sort of thing though I strongly suspect that just slapping in some DDR4-3000 will be of more benefit than trying to crank iGPU clocks.

Anime Schoolgirl posted:

AMD coming out with a 7-actually-10nm product that performs before Intel can push out their 10nm laptop chips would be quite hilarious

Yeah I don't think AMD has had a chip with a process advantage, or even been on par, on Intel since they beat Intel to copper way back in the early K7 days. It'd be interesting to see if they can really manage to knock it out of the park again.

PC LOAD LETTER fucked around with this message at Oct 26, 2017 around 09:57

Malcolm XML
Aug 8, 2009

I always knew it would end like this.


Glofo 7 has been rumored to be quite good so I'm not surprised that amd would want to jump on it

SwissArmyDruid
Feb 14, 2014



Let's just hope we don't have a rehash of the 20nm debacle.

Four goddamn years on 32nm. Four. Goddamn. Years.

NewFatMike
Jun 11, 2015



PC LOAD LETTER posted:

Unless you're trying to OC a 8 core Ryzen or really pushing a 6 core as hard as possible even the cheap mobos are supposed to be fine for OC'ing the 4 core APU's that are coming.

Buildzoid did a very good video on this shortly after launch so its a bit old now but not much has changed with the cheap mobo's or their VRM's since then.

Think this is it:

https://www.youtube.com/watch?v=ZGrxhf_xZWI

No one really knows about OC'ing the built in iGPU. That has its own power plane that isn't really being used much by current Ryzens since none of them have iGPU's. vSoC power delivery will matter lots for that sort of thing though I strongly suspect that just slapping in some DDR4-3000 will be of more benefit than trying to crank iGPU clocks.


Yeah I don't think AMD has had a chip with a process advantage, or even been on par, on Intel since they beat Intel to copper way back in the early K7 days. It'd be interesting to see if they can really manage to knock it out of the park again.

Thank you very much! That's going to be quite a useful video.

Also, Vega Mobile stuff is announced. R7 2700U and R5 2500U are both 15W parts:

https://www.youtube.com/watch?v=WwWRWC-34xQ

https://www.youtube.com/watch?v=glrw3ibl810

https://www.youtube.com/watch?v=wpWHL34G52o

Probably more videos, this is taking me by surprise!

Risky Bisquick
Jan 18, 2008

PLEASE LET ME WRITE YOUR VICTIM IMPACT STATEMENT SO I CAN FURTHER DEMONSTRATE THE CALAMITY THAT IS OUR JUSTICE SYSTEM.


Buglord

NewFatMike posted:

Thank you very much! That's going to be quite a useful video.

Also, Vega Mobile stuff is announced. R7 2700U and R5 2500U are both 15W parts:

https://www.youtube.com/watch?v=WwWRWC-34xQ

https://www.youtube.com/watch?v=glrw3ibl810

https://www.youtube.com/watch?v=wpWHL34G52o

Probably more videos, this is taking me by surprise!

They were previously thought to be 35w parts, this changes things

SwissArmyDruid
Feb 14, 2014



Holy balls, I take back everything I said about being disappointed in Raven Ridge.

NewFatMike
Jun 11, 2015



Yeah, 15W thermal envelope parts getting decent gaming marks gives me a lot of hope for cranking up a 60+W desktop part:

https://videocardz.com/73702/amd-la...h-vega-graphics

They put some of the slides in there, the gaming stuff is rather encouraging:

LoL, DOTA 2, and CS:GO getting decent marks at 1080p medium, although I'd be willing to bet their 1% lows are pretty low.

Risky Bisquick
Jan 18, 2008

PLEASE LET ME WRITE YOUR VICTIM IMPACT STATEMENT SO I CAN FURTHER DEMONSTRATE THE CALAMITY THAT IS OUR JUSTICE SYSTEM.


Buglord

It's too bad vega doesn't do full hardware decode on vp9, but i'm sure the hybrid approach should work fine @4k given the zeppelin cores. Raven Ridge is looking good for HTPC

Generic Monk
Oct 31, 2011



tbh this is looking more and more what vega was actually designed for. maybe the imac pro won't be a fire hazard after all?

Risky Bisquick posted:

They were previously thought to be 35w parts, this changes things

according to that hardwarecanucks vid they're available in both 15 and 35w configurations

Generic Monk fucked around with this message at Oct 26, 2017 around 14:38

Anime Schoolgirl
Nov 28, 2002

~perfect~
battlebrother





NewFatMike posted:

Thank you very much! That's going to be quite a useful video.

Also, Vega Mobile stuff is announced. R7 2700U and R5 2500U are both 15W parts:

https://www.youtube.com/watch?v=WwWRWC-34xQ

https://www.youtube.com/watch?v=glrw3ibl810

https://www.youtube.com/watch?v=wpWHL34G52o

Probably more videos, this is taking me by surprise!
I was expecting 25w, but 15w is trolling the gently caress out of Intel with sick donuts in the parking lot

priznat
Jul 7, 2009

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

Anime Schoolgirl posted:

A non-rackmount non-proprietary motherboard will have a hell of a time tracing 3700-4094 working pins, imagine how thick the PCBs will be on EATX, to say nothing of anything smaller...

This is true, Iím wondering if the Gigabyte one Iíve seen is a real product or just a paper launch. Ditto Supermicro, they only have their rackmount ones. Routing 128 or so pcie lanes per cpu even if not all of those are connected up is going to be a massive job if you canít just send them to a couple mezz connectors.

Still, there are Skylake-sp atx motherboards and those are 3647. Probably just focusing on getting those out first due to demand.

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


I'm a little concerned about the testing methodology used, note that end of the this video


People are saying that "AMD reference" means testing was done on an open air bench with bios throttling to 15W stable while the 8550U had to deal with a lovely OEM design.

I want real reviews yet before I shout with joy, because some of the scores seem fishy.

NewFatMike
Jun 11, 2015



FaustianQ posted:

I'm a little concerned about the testing methodology used, note that end of the this video


People are saying that "AMD reference" means testing was done on an open air bench with bios throttling to 15W stable while the 8550U had to deal with a lovely OEM design.

I want real reviews yet before I shout with joy, because some of the scores seem fishy.

I'm mobile atm, but I'll see if I can dig up the leaks that included mainboard info and compare them to the reference figures

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Raven Ridge die shot.



Is it me or does AMD's uncore seem enormous compared to Intel? Either way the space that the Vega CU's take up is tiny, and that's 11 of them.

EDIT: Quick and dirty paint job seems to indicate you'd be able to get 44CU in something moderately larger than Polaris 10 or 36CUs and something noticeably smaller than Polaris 10.

EmpyreanFlux fucked around with this message at Oct 26, 2017 around 18:47

repiv
Aug 13, 2009



College Slice

AMD has a bunch of on-die USB and SATA controllers so you'd expect more uncore area than Intel, who put all that stuff on the chipset

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


repiv posted:

AMD has a bunch of on-die USB and SATA controllers so you'd expect more uncore area than Intel, who put all that stuff on the chipset

This should save a lot in the mobile space, right? Cheaper boards, easier thermal solutions.

Seamonster
Apr 30, 2007

IMMER SIEGREICH


Anime Schoolgirl posted:

I was expecting 25w, but 15w is trolling the gently caress out of Intel with sick donuts in the parking lot

Yeah most ultrabooks should have no problem cooling a 25w part especially if thats really the only thing in there generating and significant heat (I guess some NVMe drives get a little toasty). But an extra 10w budgeted to the Vega cores for something that stacks up more convincingly against the MX150 would much more desirable. They're about 25% short on FS gpu score.

Adbot
ADBOT LOVES YOU

repiv
Aug 13, 2009



College Slice

FaustianQ posted:

This should save a lot in the mobile space, right? Cheaper boards, easier thermal solutions.

Yeah I think we'll start to see low-cost designs using the X300 and B300 chipsets. They're supposed to be "null chipsets" that just do the bare minimum needed to bootstrap the CPU, but don't provide any additional USB/SATA/PCIe I/O beyond what the CPU has built in, so their cost and thermals should be negligible.

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply
«245 »