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.
«177 »
  • Post
  • Reply
EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


https://www.youtube.com/watch?v=pCDJERMTL3s

TL;DW is that the Athlon is price/perf better than the Pentium 5400, and that despite a similar number of ALUs it still savages Intel HD, and occasionally jumps ahead of the A12-9800 in gaming.

1) Holy poo poo are the older apus choked to hell on the GPU side.
2) Holy poo poo Intel fix your loving drivers, in Firestrike an HD 630 is it on par with Vega 3 but here it gets mangled.
3) AMD shouldn't have locked it, as unlocked it'd be pissing on the Pentiums grave.

EmpyreanFlux fucked around with this message at Sep 18, 2018 around 19:26

Adbot
ADBOT LOVES YOU

Anime Schoolgirl
Nov 28, 2002

~perfect~
battlebrother





i wouldn't put it past them to put up an overclockable 242GX that's 5-8 dollars more expensive than the 240GE

Anime Schoolgirl fucked around with this message at Sep 18, 2018 around 22:35

Munkeymon
Aug 14, 2003

Motherfucker's got an
armor-piercing crowbar! Rigoddamndicu𝜆ous.



Pillbug

Cygni posted:

Didn't see this mentioned, AMD launched more of the Zen+ stack today. 2700E and 2600E (45w limited parts), 2500X (4/8), and 2300X (4/4)

https://www.anandtech.com/show/1334...2500x-and-2300x

So are the E parts OEM-only as well? The article only calls out the X ones as being OEM limited.

mdxi
Mar 13, 2006

to JERK OFF is to be close to GOD... only with SPURTING

Dinosaur Gum

Munkeymon posted:

So are the E parts OEM-only as well? The article only calls out the X ones as being OEM limited.

I really hope not. Intel sells boxed T series CPUs, so I'm hoping for parity there.

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Anime Schoolgirl posted:

i wouldn't put it past them to put up an overclockable 242GX that's 5-8 dollars more expensive than the 240GE

I could see a 220GE (3.4Ghz/3CU; 60$) a 240GE (3.5Ghz/6CU; 70$) and a 260GX (3.6Ghz/6CU, unlocked; 80$).

I mean on paper that 260GX is terrible value so close to the 2200G, but it'd definitely maul anything sub 100$ Intel still, while a 240GE would be an obvious Price/Perf winner.

apropos man
Sep 5, 2016

You get a hundred and forty one thousand years and you're out in eight!

Sooo, after this upcoming 7nm fabrication release, are AMD/Intel looking to push the shrink even further, or will we have reached the limit of what physics will allow by then?

Or will be an elongated mission over 5 or 10 years to push down to 5 or 3nm because of the increase in difficulty? During that time they could sell a chip on new features or a vast number of cores while they slowly fab down to 5nm or less?

I mean, surely the effort required is on the top end of an exponential difficulty curve at the moment?

mdxi
Mar 13, 2006

to JERK OFF is to be close to GOD... only with SPURTING

Dinosaur Gum

mdxi posted:

I really hope not. Intel sells boxed T series CPUs, so I'm hoping for parity there.

Aw poo poo. 2700E has a page on AMD's site now. The only part info given is: "OPN Tray: YD270EBHM88AF"

Every other retail CPU in the Ryzen line has had a part ending in "BOX", like the 2700X's "OPN PIB: YD270XBGAFBOX"

Seamonster
Apr 30, 2007

IMMER SIEGREICH


apropos man posted:

Sooo, after this upcoming 7nm fabrication release, are AMD/Intel looking to push the shrink even further, or will we have reached the limit of what physics will allow by then?

Or will be an elongated mission over 5 or 10 years to push down to 5 or 3nm because of the increase in difficulty? During that time they could sell a chip on new features or a vast number of cores while they slowly fab down to 5nm or less?

I mean, surely the effort required is on the top end of an exponential difficulty curve at the moment?

7nm is going to be around for a looooongg while. Think 7nm++++

By the time they're done we are probably looking at consumer grade parts (NOT even TR) with 16c/32t @ 5.5 GHz for ~ <150W though so its not all bad.

PC LOAD LETTER
May 23, 2005
WTF?!

Slippery Tilde

apropos man posted:

Sooo, after this upcoming 7nm fabrication release, are AMD/Intel looking to push the shrink even further, or will we have reached the limit of what physics will allow by then?
Read this here: https://semiengineering.com/big-trouble-at-3nm/

tl&dr: 3nm will probably be the practical limit mostly due to economics, and due to economics for many of the cost focused products they won't ever go past 5nm, but there are also huge and real physics problems with trying to mass manufacture chips at less than 3nm that will probably make sub 3nm untenable too.

apropos man posted:

Or will be an elongated mission over 5 or 10 years to push down to 5 or 3nm because of the increase in difficulty?
Yes, 7nm, in various forms (ie. 7nm+ 7nm++ or whatever they call it, expect the marketing in naming these new processes to get more aggressive) will be around for a long long long time because its about as small as you can go without EUV for practical mass manufacturing and still get enough performance gains of some sort to make it worthwhile.

Even Intel for instance has been rumored to be renaming their 7nm improved version processes "5nm" or even "3nm" instead of 7nm+ or 7nm++.

apropos man
Sep 5, 2016

You get a hundred and forty one thousand years and you're out in eight!

Hmm. Yes. Let's give the time taken to beat 7nm an arbitrary length of time. Say 10 years R&D.

That delay could cause a shakeup in how we use out PC's, do ya think?

We could all be using PC's with much many cores or perhaps your average PC lover would end up having a cluster of machines around the house, all working together since the compute power of one CPU has hit bottleneck?

Maybe we'll see gaming being delivered as a service, so that you only need a basic terminal and the game streams from Steam?

Maybe we'd go through a period of PC's becoming extremely optimised for I/O etc, since that is the main area of development left as the CPU designs slow to a snail's pace.

apropos man
Sep 5, 2016

You get a hundred and forty one thousand years and you're out in eight!

PC LOAD LETTER posted:




Even Intel for instance has been rumored to be renaming their 7nm improved version processes "5nm" or even "3nm" instead of 7nm+ or 7nm++.

Seriously? How can they announce blatant lies if they're selling 7nm fabbed chips and passing them off as 5 or 3?

PC LOAD LETTER
May 23, 2005
WTF?!

Slippery Tilde

Well if things continue to follow the current trends you'll see lots more cores at a minimum.

How that gets implemented exactly will be where things get interesting (ie. chiplets, die stacking, etc.).

I think the biggest possible shake up could be if we see real big and powerful FPGA's get integrated into these future CPU's as a standard feature. Along with possibly some sort of "AI" machine learning engine of sorts. You're starting to see some of that being integrated into some of these new phone/mobile SoC's.

I think right now they're not very practical since the software to make proper use of them in interesting ways isn't really there yet but potentially they could become a real big deal.

apropos man posted:

Seriously? How can they announce blatant lies if they're selling 7nm fabbed chips and passing them off as 5 or 3?
They'll call it marketing and no one will care. TSMC and others have been doing it to some degree too for years. The process names have been largely BS for almost everyone (Intel was the least BS'y though) for a long time.

PC LOAD LETTER fucked around with this message at Sep 20, 2018 around 04:41

apropos man
Sep 5, 2016

You get a hundred and forty one thousand years and you're out in eight!

Yep, I do agree that a bottleneck could alter the home computer experience.

Aren't FPGA's slugging, though, because of the circuits inside them being adaptable to whatever program you give it, there's a performance penalty due to the complexity compared to a traditional CPU.

PC LOAD LETTER
May 23, 2005
WTF?!

Slippery Tilde

apropos man posted:

Aren't FPGA's slugging, though, because of the circuits inside them being adaptable to whatever program you give it, there's a performance penalty due to the complexity compared to a traditional CPU.
FPGA's do tend to have a significantly lower peak clock speed than a highly optimized ASIC but the flexibility can allow them to be highly tuned to a specific application and can potentially get massive performance improvements over a more general purpose ASIC.

The gotcha there is that the FPGA has to be properly programmed to achieve that level of performance and doing so is usually not something that is easy and that it won't be good at general purpose tasks. But you'd have a "traditional" x86 CPU there to handle the general purpose/legacy stuff in this case so you wouldn't need it be good at general purpose stuff. Programming the FPGA would be something that would potentially be handled by the software guys so when you'd go to load DOOM2026 or whatever the software would also set up the FPGA to run a given task instead of trying to run it on the CPU or over on the GPU. (edit) Doing that would free up either of those devices to either do more work elsewhere or just to potentially use much less power (ie. the CPU cores would turn themselves off mostly or go to a low power state while most of the heavy lifting would be done more efficiently + faster by the FPGA).

At least that is potentially how it'd be done. Whether or not things pan out that way is something that remains to be seen.

PC LOAD LETTER fucked around with this message at Sep 20, 2018 around 04:56

fishmech
Jul 16, 2006

~death to capitalism~
Chrome OS is shit
Every DSA is a cop


Salad Prong

apropos man posted:

Yep, I do agree that a bottleneck could alter the home computer experience.

Aren't FPGA's slugging, though, because of the circuits inside them being adaptable to whatever program you give it, there's a performance penalty due to the complexity compared to a traditional CPU.

You're getting things a bit twisted around.

An FPGA attempting to emulate a full traditional CPU will be slower, because FPGAs usually can't do all that work at high rate of speed in the same price point. But what you're doing with an FPGA in a situation like this is you're setting it up to only handle a reduced amount of things, and it is then able to perform that restricted set of tasks much faster than a general purpose CPU of similar costs and power budget.

In this situation of integrating a bunch of additional FPGA space with the CPU, you'd have people customizing their setup towards some particular task they do often - say routines to speed up video rendering - while using the typical CPU for other tasks. You might even have it setup to switch between a few particular feature sets as needed, though you can't quite do that on the fly. So there'd need to be some wait time while the FPGA is reprogrammed to switch between a profile for fast video rendering over to a profile for accelerating crunching a large dataset for an "AI" thing.

As to whether the benefits of including the FPGA for such a use are really worth having it in a future CPU design, it's hard to say. But it's not because an FPGA is always slower or anything.

apropos man
Sep 5, 2016

You get a hundred and forty one thousand years and you're out in eight!

Thanks for all the detailed replies.

I wonder if we may end up seeing PC's with several FPGA's, perhaps a bank of then.

One could be programmed to act as a wireless adaptor, another for a USB chipset, another for a RAID controller, etc etc.

That approach could lead to some really tidy little PC's, since there would be a lot of modules on the board of the same dimension for placement.

LRADIKAL
Jun 10, 2001

A Very Useful Person

Fun Shoe

apropos man posted:

One could be programmed to act as a wireless adaptor, another for a USB chipset, another for a RAID controller, etc etc.

These things currently, are the opposite of a general purpose processor. These are examples of dedicated hardware which is built for a relatively simple task and allows it to be small and fast and efficient. They are also very cheap to produce. FPGA's are expensive like a CPU, but can be programmed repeatedly to optimize certain instructions.

edit: perhaps the next step is faster communication between an APU and separate graphics hardware. I think this is possibly with AMD stuff, but not very practical.

LRADIKAL fucked around with this message at Sep 20, 2018 around 08:08

SwissArmyDruid
Feb 14, 2014



I would like to add that none of what has been discussed so far has even scratched the surface of moving to a different material. All of the current projected woes are due to the physical limits of Si or SiGe, there's an entire other bag of problems that may arise if the industry moves to a different substrate, such as InGaAs, or maybe we finally tackle graphene some time within the next decade, or something a little less exotic, like boron-doped diamonds.

EmpyreanFlux
Mar 1, 2013

The AUDACITY! The IMPUDENCE! The unabated NERVE!


Is there a paper available for a deeper dive on the benefits to moving to a new material? Like besides allowing smaller node sizes, do they enable better performance in general? Would it be worth it to migrate such techniques back up the stack to 12nm, 28nm and 40/45nm?

Sidesaddle Cavalry
Mar 15, 2013

65535

dispel please


The general goals of moving to new materials are 1. Making the electrons inside go faster so the transistors switch faster so computers go faster
2. Using fewer electrons to make a transistor switch so that it consumes less power and doesn't get literally hotter than the freaking sun when you pack billions of them together
And certain III-V materials can do one or the other, given their crystal structure.

Even with our current materials there's a ton of research being done on new transistor "mechanisms", like single-atom transistors or spintronic ones that don't have to worry about pushing electrons at all.

ConanTheLibrarian
Aug 13, 2004


dis buch is late

Fallen Rib

Additionally: there's no reason to suppose any of them will work.

SwissArmyDruid
Feb 14, 2014



Yes. That is a BIG problem. Some of them are already in use *now*, but not in the kind of ways that we see silicon currently being used at that kind of resolution and detail. The industry might burn as much money as they have to get to 7nm now, just to get one of those materials working at a node that we would consider enormous by comparison.

Khorne
May 1, 2002

Goonstone Champ x2

apropos man posted:

Sooo, after this upcoming 7nm fabrication release, are AMD/Intel looking to push the shrink even further, or will we have reached the limit of what physics will allow by then?

Or will be an elongated mission over 5 or 10 years to push down to 5 or 3nm because of the increase in difficulty? During that time they could sell a chip on new features or a vast number of cores while they slowly fab down to 5nm or less?

I mean, surely the effort required is on the top end of an exponential difficulty curve at the moment?
5nm will come fairly quickly (4? years) after because the nm ratings aren't true to size and production-ready EUV is/was one of the bottlenecks for 7nm.

Continued shrinks might cause longevity issues we haven't encountered much before. Degradation due to use becomes a concern eventually at smaller sizes. That's one thing people usually fail to mention because it's not as cool as tunneling or density related issues.

Khorne fucked around with this message at Sep 21, 2018 around 21:25

Adbot
ADBOT LOVES YOU

Potato Salad
Oct 23, 2014




Tortured By Flan

apropos man posted:

Maybe we'd go through a period of PC's becoming extremely optimised for I/O etc, since that is the main area of development left as the CPU designs slow to a snail's pace.

This is already where we've gone in enterprise. Because cpus haven't been improving all that much in comparison to the conpute load humanity demands of the internet, in storage alone we are: gutting protocol overhead (SATA/SAS to NVMe), eliminating the need for storage protocols altogether (3dxpoint dimms now sold as "pmem"), and making relatively-secure storage retrieval directly from network interface to block device without even going though the cpu (RDMA). Both dx12 and vulkan are minimizing the involvement of system calls. The cpu is being eliminated as a middleman where possible.

Potato Salad fucked around with this message at Sep 22, 2018 around 01:39

  • 1
  • 2
  • 3
  • 4
  • 5
  • Post
  • Reply
«177 »