Friday, January 17th 2014

AMD Readies 16-core Processors with Full Uncore

AMD released developer documentation for a new processor it's working on, and the way it's worded describes a chip with 8 modules, working out to 16 cores, on a single piece of silicon, referred to as Family 15h Models 30h - 3fh. This is not to be confused with the company's Opteron 6300-series "Abu Dhabi" chips, which are multi-chip modules of two 8-core dies, in the G34 package.

What's more, unlike the current "Abu Dhabi" and "Seoul" chips, the new silicon features a full-fledged uncore, complete with a PCI-Express gen 3.0 root complex that's integrated into the processor die. In what's more proof that it's a single die with 8 modules and not an MCM of two dies with 4 modules each, the document describes the die as featuring four HyperTransport links; letting it pair with four other processors in 4P multi-socket configurations. Such systems would feature a total core count of 64. There's no clarity on which exact micro-architecture the CPU modules are based on. Without doubt, AMD is designing this chip for its Opteron enterprise product stack, but it should also give us a glimmer of hope that AMD could continue to serve up high-performance client CPU, only ones that can't be based on socket AM3+.
Source: Planet3DNow.de
Add your own comment

92 Comments on AMD Readies 16-core Processors with Full Uncore

#1
RCoon
I sincerely doubt this will ever come to any other market besides Opteron. If it does, enjoy the 50000W TDP.
Posted on Reply
#2
FreedomEclipse
~Technological Technocrat~
16 'real' cores, amirite?
Posted on Reply
#3
Fiery
FinalWire / AIDA64 Developer
"Family 15h Models 30h - 3fh"

"There's no clarity on which exact micro-architecture the CPU modules are based on"

Family 15h Models 30-3Fh means Steamroller uarch.
Posted on Reply
#4
buildzoid
If there was a Opteron compatible motherboard that was laid out for desktops not green PCB, high power VRM, BIOS with OC features, less lan connectors and more USB I'd buy this or a 12 core version for my next build.
Posted on Reply
#5
Yorgos
RCoonI sincerely doubt this will ever come to any other market besides Opteron. If it does, enjoy the 50000W TDP.
call it 100 Watt SDP and sell it to customers then ;)

edit: all jokes aside, for those of us that are in deeper waters there is this thing called linux that provides info about cpus and their μ-arch earlier than anyone else.
lkml.org/lkml/2013/6/5/989
On Wed, Jun 05, 2013 at 03:50:03PM
Posted on Reply
#6
west7
it will be great to see 8-10 cores desktop version of this
Posted on Reply
#7
ZetZet
west7it will be great to see 8-10 cores desktop version of this
not happening. just get the apu :roll:
Posted on Reply
#8
Baum


Well why not start to work on better power efficency for the high performance segment, just not to get phased out by cpu power some years ahead....
Posted on Reply
#9
Mathragh
Fiery"Family 15h Models 30h - 3fh"

"There's no clarity on which exact micro-architecture the CPU modules are based on"

Family 15h Models 30-3Fh means Steamroller uarch.
If this indeed is a CPU based on the steamroller core, putting even more cores on die makes sense. With steamroller they repositioned the power-optimum of the cores way downwards, resulting in slightly worse scaling in high frequencies, but much better scaling at lower frequencies(just take a look at the anandtech 45W vs 95W parts, 45W keeps a huge chunk of performance while using way less power). If you want to optimally use this property, more lower clocked cores makes a lot of sense, and SR cores should mitigate atleast some of lost performance as a result of lower clockspeeds.

I'm not to sure about whether they'll be able to effectively compete with intel high core count products though, but I suppose they're certain enough themselves since they're bringing out this product.
Posted on Reply
#10
Fiery
FinalWire / AIDA64 Developer
MathraghIf this indeed is a CPU based on the steamroller core, putting even more cores on die makes sense. With steamroller they repositioned the power-optimum of the cores way downwards, resulting in slightly worse scaling in high frequencies, but much better scaling at lower frequencies(just take a look at the anandtech 45W vs 95W parts, 45W keeps a huge chunk of performance while using way less power). If you want to optimally use this property, more lower clocked cores makes a lot of sense, and SR cores should mitigate atleast some of lost performance as a result of lower clockspeeds.

I'm not to sure about whether they'll be able to effectively compete with intel high core count products though, but I suppose they're certain enough themselves since they're bringing out this product.
What AMD uses since they won the console deal is a LEGO-like architecture. They have a 2-core Bulldozer/Piledriver/Steamroller module or a 4-core Jaguar/Puma module, and they can pack in as many of them on a CPU/APU die as the target market expects them to do. They also have a variety of uncore modules (e.g. 2-channel DDR3-2133, 4-channel GDDR5, etc), and they can pick the one best for the target market. Then of course they have various GCN iGPUs, ranging from 128 units up to 3072 units, and so they can pair it with virtually any CPU. That approach is very useful when you build so many variants of APUs for very different markets, e.g. Temash, Kabini, XBox One, PS4, Kyoto (8-core Jaguar), Kaveri, etc.

The only problem is: AMD doesn't have any Steamroller or Excavator uarch based 2P/4P/8P CPUs or APUs on their roadmap. No matter if they talk about such possibilities in one of their many documents, they do not wish to compete with Intel in those markets anymore. The best they have and will have is Kaveri for 1P servers/workstations, a.k.a. Berlin.
Posted on Reply
#11
TheoneandonlyMrK
btarunrAMD released developer documentation for a new processor it's working on, and the way it's worded describes a chip with 8 modules, working out to 16 cores, on a single piece of silicon, referred to as Family 15h Models 30h - 3fh. This is not to be confused with the company's Opteron 6300-series "Abu Dhabi" chips, which are multi-chip modules of two 8-core dies, in the G34 package.

What's more, unlike the current "Abu Dhabi" and "Seoul" chips, the new silicon features a full-fledged uncore, complete with a PCI-Express gen 3.0 root complex that's integrated into the processor die. In what's more proof that it's a single die with 8 modules and not an MCM of two dies with 4 modules each, the document describes the die as featuring four HyperTransport links; letting it pair with four other processors in 4P multi-socket configurations. Such systems would feature a total core count of 64. There's no clarity on which exact micro-architecture the CPU modules are based on. Without doubt, AMD is designing this chip for its Opteron enterprise product stack, but it should also give us a glimmer of hope that AMD could continue to serve up high-performance client CPU, only ones that can't be based on socket AM3+.



Source: Planet3DNow.de
I like your assertion that you were right about AM3+ Bta

No comment about you being WRONG about AMD big cores though eh FX is dead my ASSSSSS.

and you may have to re address your assertion AM3+ is gone bacause even with that uncore this could all still go in a G34 socket and likely will,, since in server land socket swaps are even rarer then consumer land.

If it has the hyperlink bus still in it and likely will have 2-3 ,,everything stays on the table as you can theoretically slap a NB off of it, and pciex 3 support direct linked to the uncore could very easily bypass any Nbridge which would still have pciex lanes available. The memory controller as happens today is available on or off die via the NB or CPUNB rending this chip supportable by older sockets even with ddr3, though the memory controller on this chip is not listed so is up in the Air.

The Big Cores live nice and steamy too ;)


As for the road map dude, Those we have actually seen have been complete BS or only consumer level parts covering 2013- end of 2014 start 0f 15 and not beyond so wadya know that we dont and any slides??.
Posted on Reply
#13
JDG1980
ZetZetnot happening. just get the apu :roll:
Did it occur to you that AMD wants people to think that it's not happening, because they want people to buy Kaveri now instead of waiting for something better later on?

Roadmaps are marketing documents first and foremost.
Posted on Reply
#14
JDG1980
I wonder if this will be done on the 28nm process (like Kaveri), or back-ported to 32nm.

28nm is more energy-efficient, but 32nm offers higher clock speeds. On one hand, server parts generally don't require the highest possible clock rates. On the other hand, higher clock rates would be helpful if they want to sell the lower-binned parts to enthusiasts (as is the common practice), and AMD may still have obligations to buy a certain number of 32nm wafers from Global Foundries for the next few years.

Either way, it's good news that AMD has not given up on the high-margin big-core server segment.
Posted on Reply
#15
librin.so.1
IF this is not only for the Opteron line and will include models for desktop computers, even if they have a smaller core count (e.g. max 8 cores for desktop models)...
...I am going to cry in happy manly tears.

like this:
Posted on Reply
#16
lilhasselhoffer
...interesting...

I don't understand it, but I see where this is going.

AMD has pushed for more cores. In order to get these extra cores, they first compromised components, and are now compromising integration. From a consumer level CPU standpoint, this is absolutely moronic. From AMD's perspective this is gold. Allow me to explain.

Let's say that AMD is one cohesive unit, not a GPU/CPU divided company. The CPU is one set of cores, designed to work (originally) in very linear tasks. The GPU is a collection of processors, which lack the raw performance of a CPU with regards to precision. AMD is trying to marry the two ideas together, like they have with the APU.

Now, I don't believe this is a great idea for everything. I don't care that a computer can have integrated graphics and play a 4 year old game at low resolution with an acceptable frame rate. I do want my tablet to be something more than a toy. Give me two different processors, and we'll be on the same page. The truth is that software isn't out there, on the consumer side, to make use of huge core count processors. In a few year, it'll be different. Hopefully the APU isn't too far ahead of the curve to survive. They'll have the core counts high, the heat output low (due to lower clocking), and a decent install base. I definitely hope they will beat Intel, without even being seen as a valid competitor.

Edit:
Forgot to mention this. The software doesn't exist on the consumer side, but the server side is aching for more cores. This isn't great news for the consumer, but it is great for servers.
Posted on Reply
#17
fullinfusion
Vanguard Beta Tester
Hell yeah I'll take one! I miss amd as intel is just boring as rcoons posts lol.. ;)
Posted on Reply
#18
harry90
All amd Needed to do was to improve the single core performance of their cpu's. who needs more than 8 cores? Just enhance the IPC, single core performance by 40-60% and they could compete with intel!!!
Posted on Reply
#19
Blue-Knight
anonymous6366and then a quad core i7 is still faster -.- lol
I was thinking exactly that... LOL!
Posted on Reply
#20
newtekie1
Semi-Retired Folder
I will say it again, I still want AMD to move to a unified socket. Use the same LGA socket for servers, for performance desktops, and for mid-range to low end desktops. Of course all their processors will be APUs(Intel's already are), allow us to put anything from dual-cores up to 16-core processors in desktops or even put Opterons in desktops if we wanted to. That to me is a winning idea for AMD. Don't segregate the market by sockets, let someone start with a super cheap processor to get them up and running, and still have the option to upgrade to a high end processor if they want.
harry90All amd Needed to do was to improve the single core performance of their cpu's. who needs more than 8 cores? Just enhance the IPC, single core performance by 40-60% and they could compete with intel!!!
Blue-KnightI was thinking exactly that... LOL!
The thing is, AMD designs their CPUs very differently from Intel. Intel, because they rely on hyper-threading, designs their single core to do the work of two cores. Of course this means when it is only loaded with a single threaded work load, it is extremely fast. This is one of the reasons AMD was reasonably close to Intel in single threaded performance during the early Athlon 64/x2 days when they were competing with Conroe(yes, conroe was still faster, but AMD was a lot closer back then). Intel wasn't designing their processors to use hyper-threading.

Honestly, I don't see a need for AMD to increase single threaded performance to meet Intel. The reason is that AMD's single threaded performance is good enough. Yes, they lag behind in benchmarks, but in real world use there really isn't anything that is single threaded that AMD can't handle. Saddly, games are still heavily dependent on single threaded performance, but most modern games still run perfectly well on AMD processors despite this, because AMD's single threaded performance is good enough. There are a few exceptions, StarCraft II comes to mind, because it is extremely CPU heavy and extremely single threaded.
Posted on Reply
#21
BiggieShady
harry90Just enhance the IPC, single core performance by 40-60% and they could compete with intel!!!
Just? They would have to do ballsy thing Intel has done when they sucked with NetBurst - they need to go back to their older, more efficient architecture and improve on that. Just like Intel ditched NetBurst in favor of Pentium M architecture which improved became Core architecture.
Posted on Reply
#22
Popocatepetl
lilhasselhoffer...interesting...
Wow, a post made of pure 100% stupid. Never, ever utilize your 100% stupid when posting.
I don't care that a computer can have integrated graphics and play a 4 year old game at low resolution with an acceptable frame rate.
According to various online reviews of Kaveri that thing runs Battlefield 4 (without Mantle which shoudl enable even better performance) with reasonable framerates. Most reviews used recent titles and came up with figures of 30+ fps.

From this point on your post devolves into incoherent babbling about tablets (?!) and whatnot. Again, never use 100% stupid in your posts. Thanks.
Posted on Reply
#23
BiggieShady
PopocatepetlWow, a post made of pure 100% stupid. Never, ever utilize your 100% stupid when posting.



According to various online reviews of Kaveri that thing runs Battlefield 4 (without Mantle which shoudl enable even better performance) with reasonable framerates. Most reviews used recent titles and came up with figures of 30+ fps.

From this point on your post devolves into incoherent babbling about tablets (?!) and whatnot. Again, never use 100% stupid in your posts. Thanks.
I just went through all 5 of your posts. Up until this post you seemed like a completely normal person :shadedshu:
Posted on Reply
#24
FX-GMC
PopocatepetlWow, a post made of pure 100% stupid. Never, ever utilize your 100% stupid when posting.



According to various online reviews of Kaveri that thing runs Battlefield 4 (without Mantle which shoudl enable even better performance) with reasonable framerates. Most reviews used recent titles and came up with figures of 30+ fps.

From this point on your post devolves into incoherent babbling about tablets (?!) and whatnot. Again, never use 100% stupid in your posts. Thanks.
Posted on Reply
#25
lilhasselhoffer
PopocatepetlWow, a post made of pure 100% stupid. Never, ever utilize your 100% stupid when posting.



According to various online reviews of Kaveri that thing runs Battlefield 4 (without Mantle which shoudl enable even better performance) with reasonable framerates. Most reviews used recent titles and came up with figures of 30+ fps.

From this point on your post devolves into incoherent babbling about tablets (?!) and whatnot. Again, never use 100% stupid in your posts. Thanks.
Did you read the original post? You claim that the post is 100% stupid, but you don't even bother to read.

I stated that AMD is increasing core count, and that's not an optimal strategy for the consumer CPU market. More cores are bringing us closer to a true fusion of GPU and CPU. I stated that the core count increase is detrimental to consumers, because they are fusing the tablet, desktop, and server market. At no point in time did I reference graphical performance numbers. The point of this article was not graphics, but a CPU with a monstrous core count more akin to what you might see in a GPU.

If you need to stick your foot in your mouth, please do so without calling someone else an idiot. Every person eventually commits a stupid act (I've had my fair share); but missing the point of a post, the point of the thread, and then calling someone else stupid is an act of either brazen ignorance or willful trolling. I have no love or hate for AMD, Nvidea, or Intel. I find it interesting that they are so blatantly leaving the consumer CPU market. Either their vision is so forward thinking that they will have the last laugh, or they have just run full speed into a brick wall. It will be interesting to see what comes of this, and how it hopefully influence desktop computing.
Posted on Reply
Add your own comment
Dec 5th, 2024 09:39 CST change timezone

New Forum Posts

Popular Reviews

Controversial News Posts