Speculative Talk: AMD K10

[MadModMike]

As with Nehalem, I want to get views on AMD's next architecture, K10 and I ask for non-fanboyism. Some sources say it is dead, while others say it is on track for release. What I know about K10 is that it will bring Multi-Core, possibly 8-Cores or more, DDR3/DDR4, use HyperTransport 3.0 or 4.0, probably use Z-RAM or another form of Static RAM for L2 Cache and may include L3 Cache. I also heard AMD is shooting for 10GHz Operation, including additional registers, possibly SSE4, and looking for a 2007-2009 release date. What you guys think?

~~Mad Mod Mike, pimpin' the world 1 rig at a time

 

[Chil]

10 GHz seems to be quite a step up for a 07-09 release. How long did it take to go from 2 GHz to 3? Also, how are software companies going to keep up with a jump like that? I can only assume that development costs are gonna go way up to be able to use all 8 cores.

I'm a gamer by nature, so this kinda stuff doesn't bother my wallet much as I know that a dual core will still be viable in '07 and probably '08.

 

[MadModMike]

Sounds more and more like a Prescott to me...

That was my first impression, but it depends on where AMD goes with this, it could turn out to be a big surprise, or a huge flop.

~~Mad Mod Mike, pimpin' the world 1 rig at a time

 

[MadModMike]

Sounds more and more like a Prescott to me...

We don't need more Prescotts.

hmm... 10GHz? I thought the GHz race was over?

That doesn't mean more GHz won't be better. I'm sure Intel is aiming for alot higher clock speeds as well, especially with a 3.33GHz Conroe. W/ 45nm and 22nm, 5GHz+ should be easily obtainable.

~~Mad Mod Mike, pimpin' the world 1 rig at a time

 

[BGP_Spook]

Isn't there a bit of a snag when smaller than ~45nm?

 

[luminaris]

I think Intel will go for higher clocks as well. Especially now that they're on the right track as far as getting the right technology in place. If their new processors put out as much efficiency as they claim they do, they can certainly do it now vs. Prescott.

 

[corvetteguy]

it all seems very attainable, except the 10ghz optertron, unless theyv are THAT confident that going to 65nm will allow them to clock that high. I am really liking this z ram, it sounds great with the size reduction, and performance boost, just hope its as amazing as it sounds.

 

[MadModMike]

it all seems very attainable, except the 10ghz optertron, unless theyv are THAT confident that going to 65nm will allow them to clock that high. I am really liking this z ram, it sounds great with the size reduction, and performance boost, just hope its as amazing as it sounds.

Yeah, Z-RAM will be the best thing for the new SMP Opteron 64 Servers, it should secure AMD's hold on the Server market for the future.

~~Mad Mod Mike, pimpin' the world 1 rig at a time

 

[Snorkius]

What happened to K9?

 

[MadModMike]

What happened to K9?

K9 was dropped as the codename, but it originally was the Athlon 64 X2's.

~~Mad Mod Mike, pimpin' the world 1 rig at a time

 

[malphadour]

There will need to be some big advancements in silicon technology to achieve 10ghz. The biggest issue is that as the die size gets smaller and heads towards 45nm and less other problems have to be overcome, and the primary one which is also the speed hurdle is (to put it in simple terms) cross talk in the chip. Fixes such as strained silicon have been brought out to help this, and subsequently other new ideas abound on how to reduce this, but until a stable way can be found to reduce data contamination within the chip speeds much over 5 to 6 ghz will be extremely difficult.

This article is slightly old but might interest some of you, it has some basics about strained silicon and also a bit about some other techniques, but it gets across the main issues with increased speed and reduced die size.

http://www.penstarsys.com/editor/tech/cpu/amd/str_sil/

Someone asked how software would keep up. Once software has been developed with any sort of multi core in mind, it doesnt care how many cores your chip has. The compiler libraries that developers will use will basically compile for either single core, or more than one core, its that simple. If a prog runs dual strings on a dual core, it will run 4 strings on 4 core and so on. This is basically the same as software which has been designed to work on multichip server or workstations. For example a lot of the video rendering packages are known to run well and make great use of dual core chips. This is because video rendering was originally compiled for use on workstation class machines which would quite commonly have 2 or more processors, so the software reacts in pretty well the same way to dual cores as it does to 2 processors (not exactly but close enough). As dual core and becomes more prevalent in the market more and more software will be compiled for this, and effectively support any number of multi cores. There will of course be optimisations for more cores and depending on the compiler and the skill (or time/cost restrictions) of the programmers the increase of cores will not necessarily cause uniform scaling of performance with more cores.

That all said, 10ghz 8 core would be rather nice.

As far as will it wont it is concerned, we will have to see what happens after AMD release its labradour and Husky cores (K9).


Sorry, couldn't resist....

 

[BGP_Spook]

I did not articulate it very well since I was in a hurry, but I tried (in Msg #7) to say about what you just said, malphadour.

Hello, everybody. I am the new guy.

Be gentle.

 

[raven_87]

The problem with 45nm much like decreasing any die size is cooling.
Its not that the core puts off extreme heat, its the fact that the surface area
is now becoming so small, that properly setting up a cooling solution could
be difficult.

As for the GHZ. Its over guys...Intel isnt going with the netburst philosophy of the higher the frequencies the better. Now its starting to depend on instructions per clock cycle, energy efficiency, heat dissapation and features. Notice how low the consumption for AM2 and Conroe Dual cores are? ---

Yes, Quad core is 07. So I'm assuming we'll be seeing 4-8 cores per silicon
around this time. I think its going to be about marketing this time around.
Now that Intel has a respectable product to compare, AMD might hurt itself
with the lack of advertisement. Notice any good AMD commercials lately?? (Web doesnt count) Which is my point exactly. Now look at Intel's marketing department.....they just signed Mariah Carrey for christs sake,
these guys know what there doing. So AMD better start understanding,
that they need to somehow reach midstream users.

 

[9-inch]

1) First of all, AMD may not include SSE4 since they're going to implement a new instruction set to take full advantage of the extra registers for multimedia purposes.
3D-Now performed very well for the K6 and early Athlons, the reason why AMD didn't continue with it was because programmers around the world were using Intel's SSE standard (indeed, SSE was never better than AMD's 3D-Now).

2) They'll have double the floating point and integer units for K8L and double K8L's floating point and integer units in K10 (this is 4x K8 floating point and integer units).

3) They'll have a revamped memory controller thanks to Rambus.

4) HTT 4.0.

5) A revamped cache hierarchy with huge L1 and L3 caches thanks to Z-RAM technology.

6) Some sort of DSMT for better multitasking (akin SUNs Niagara chip)

7) Improved RAS features.

These are some of the features of what we can expect from K10.

 

[TechnologyCoordinator]

This is a very engaging discussion. I hope that AMD does not rush K10 in an effort to compete against Intel Core Micro-Architecture (ICM).

I do not know if 10 ghz would be possible from the next generation, but you have to keep in mind the Intel's ICM is going to start off around 2ghz and only go up from there, I could see ICM hitting 5 or 6 ghz as manufacturing processes improve.

I read into Z-RAM for processor cache and it sounds EXTREMELY exciting. I don't know how L3 cache would fit into the picture, but it sounds neat! In terms of release date, I hope AMD gets it out soon, but does not leave things out or make compromises to meet a time-frame.

Thanks for the info MadModMike and others.

 

[9-inch]

This is a very engaging discussion. I hope that AMD does not rush K10 in an effort to compete against Intel Core Micro-Architecture (ICM).

AMD doesn't needs to rush out K10. Conroe is already beated with K8L.
And if I'm not wrong, Intel rushed Conroe to "ease" the beating they were getting from AMD.

 

[clue69less]

There will need to be some big advancements in silicon technology to achieve 10ghz. The biggest issue is that as the die size gets smaller and heads towards 45nm and less other problems have to be overcome, and the primary one which is also the speed hurdle is (to put it in simple terms) cross talk in the chip. Fixes such as strained silicon have been brought out to help this, and subsequently other new ideas abound on how to reduce this, but until a stable way can be found to reduce data contamination within the chip speeds much over 5 to 6 ghz will be extremely difficult.

I wonder if Intel and AMD are currently trialing polymeric dielectric materials? At one time, Intel was testing various polymeric products in processes down to 20nm. A few years ago, IBM was touting polymeric dielectrics as the key to 5GHz and beyond, but I think they now just use them in one process. SiO2 has been stretched to it's limits. What's gonna be next?

 

[FITCamaro]

No ones made the point that maybe they mean 10GHz combined. 4 x 2.5GHz cores. Its a long shot but possible.

Anyway. I don't see anything over a 4GHz processor happening for a while. Maybe when Conroe gets down to 45nm we will but until then probably not. Definitely not over 5GHz for a long time.

I hope AMD doesn't try to pull an Intel though. They proved that clockspeed doesn't matter.

 

[Stimpy]

Sometimes clockspeed does matter.
If you have a process that can't be parallelized the only way to make things faster is clockspeed (going from 4 to 5 to 6 issues is quite difficult and still will not work for all types of load).

Here you can use a nice engine analogy.
Although the saying goes "there any no subsitute for cubes", you'll find the most powerfull engines aren't the ones with the most cubes, but the ones with the most revs (look at F1 engines, small 3.5l but rev upto 15000rpm).
A big 7litre is going be heavy and would not suite a lightweight racing car, great for trucks where you need the low down torque, but you'd just spin the wheels in a lightweight car.

 

[TechnologyCoordinator]

No ones made the point that maybe they mean 10GHz combined. 4 x 2.5GHz cores. Its a long shot but possible.

Gigehertz do not exist. It is a measure of operating frequency and therefore cannot be added together such as something like processor cache amounts.

Adding operating frequencies together would be like saying you can travel at 220 mp/h when in fact you just have two vehicles capable of traveling at 110 mp/h.

 

[enewmen]

I am also interested in 2007/2008 CPUs and Chipsets.
2007 might sound like a long time away, but I only buy new (everything) systems about every 3-4 years.

Knowing this can keep me from buying a 939 CPU if it's not supported much longer. Or knowing PCIe will be supported for years to come will help me decide on what graphics card to get.

 

[stOrmy]

MadModMike

I need your expert advice. I want to upgrade my existing AGP motherboard to a PCIe to run some of the later video cards but not sure how far to go with the upgrade. I have a FX-53 processor currently and would like to continue using it unless you feel dual core is just that much better. I would like to overclock the processor but not sure how well a FX-53 will perform in that area. I know it is fully unlocked but never attempted to take it past the normal clock speed. I know, so please don't jump on me too bad. The thing is I don't really feel like putting a ton of money into this upgrade right now with so many changes coming from both Intel and AMD. I'm a gamer and really just want a really good PCIe video card. I have come to trust you over several weeks of reading your posts and would like you take on it. Thanks. :lol:

 

[TechnologyCoordinator]

I am also interested in 2007/2008 CPUs and Chipsets.
2007 might sound like a long time away, but I only buy new (everything) systems about every 3-4 years.

Knowing this can keep me from buying a 939 CPU if it's not supported much longer. Or knowing PCIe will be supported for years to come will help me decide on what graphics card to get.

Well, first off, Socket 939 is in its final few months. In June Socket AM2 will come out. In your situation I would at the very least wait for Socket AM2 or Conroe to come out.

If you are going to stick with AMD, then you may want to consider waiting for K10, but there are no solid time frames for this, so that would be your call.

 

[Hawkeye22]

I just hope they fix the problem with TSC drift, thus eliminating all these dual/multi core patches. It looks like one is in the works.

excerpt:
Future TSC Directions and Solutions
===================================
Future AMD processors will provide a TSC that is P-state and
C-State invariant and unaffected by STPCLK-throttling. This
will make the TSC immune to drift. Because using the TSC
for fast timer APIs is a desirable feature that helps
performance, AMD has defined a CPUID feature bit that
software can test to determine if the TSC is
invariant. Issuing a CPUID instruction with an %eax register
value of 0x8000_0007, on a processor whose base family is
0xF, returns "Advanced Power Management Information" in the
%eax, %ebx, %ecx, and %edx registers. Bit 8 of the return
%edx is the "TscInvariant" feature flag which is set when
TSC is P-state, C-state, and STPCLK-throttling invariant; it
is clear otherwise.

The rate of the invariant TSC is implementation-dependent
and will likely *not* be the frequency of the processor
core; however, its period should be short enough such that
it is not possible for two back-to-back rdtsc instructions
to return the same value. Software which is trying to
measure actual processor frequency or cycle-performance
should use Performance Event 76h, CPU Clocks not Halted,
rather than the TSC to count CPU cycles.
.
.
.
Until TSC becomes invariant, AMD recommends that operating
system developers avoid TSC as a fast timer source on
affected systems. (AMD recommends that the operating system
should favor these time sources in a prioritized manner:
HPET first, then ACPI PM Timer, then PIT.)

The full report can be found here.

 

[allhell]

All this talk about cpu speed as nice to dream about.... BUT!!! even if it were to come to reality right now it will not offer much benefit to systems using the current popular HD's.
The HD is probably the biggest bottleneck to systems performance today. Personally I would prefer significantly faster HD's right now than a 10GHz CPU & I await the day when the Hybrid ram/disk drives become available. & my system will boot in 1-3 seconds.... my .02c