AMD does 4x4 desktop computing
4x4 is a long term strategic solution to pin down Intel at half the performance
Two dual core CPUs in desktop, 4x4 computing. This is also great for cheap servers, Google should definitely look into it. There will be two banks of memory, two CPUs connect via ccHT gluelessly as in a 2P Opteron server. The two processors may be sitting next to each other, one heatsink is needed... This reduces motherboard cost substantially.
I hope it can mixed CPUs, say one dual core and one quad-core, or a dual-core and an FPGA co-processor. Or, you can have one 4GHZ single core CPU for single threaded games, and one dual core for multithreaded ones. I also wonder about the rumored inverse threading.
4x4 is a long term strategic ccHT solution to pin down Intel permanently at 50% of AMD's performance. As Intel moves to quadcore, so does AMD, 4x4 will have 8 cores. AMD can keep this 2x advantage until Intel develops something similar to ccHT and IMC.
AMD is so smart.
I can hear Intel folks crying about AMD changing the rules, "unfair" they complain.
Intel's thinking is outdated. While the whole computing industry is looking to increase throughput, Intel is trying to increase complexity and squeeze last bit of single thread performance. Intel already ran into a wall with GHZ, but they never learnt the lesson. The AMD 4x4 technology will give AMD 80% higher performance on 80% of the benchmarks.
Do you want to buy an Intel Extreme processor for $999, or two 35 watt Athlon 64 X2 3800+ AM2s for $600? Do you want 100% performance improvment or alleged 20%? I think the choice is clear.
AMD should introduce a "4x4 capable" label for its boxed AM2 CPUs, much like the "HDTV capable" sticker on TVs. When people see the sticker, they know the 4x4 potential of the processor. The rest are all low end.
More on AMD's plans is here.
posted by Sharikou, Ph. D. @ 6/01/2006 12:56:00 AM
36 Comments:
Of course, using two processors means increased power consumption. Even more so on the old outdated 90nm technology AMD is using.
But, naturally, if Intel does anything that would increase the TDP of a processor by even 1W; this would brazenly be displayed on this blog!
I really don't get it - what is so smart about 2xDual Core?
This is same stupid as SLI which gives you 30% increase doubling the cost.
I like AMD technology and I understand that this is pure marketing move to spoil Conroe fest. But what is so smart about that??? And what is a difference between 2-socket AM2/940 etc solution with two dual-cores?
Usability of such machine is kind of wired because desktop with four-cores can act as a video/renderinng workstation. Multithreading gaming is a myth.
We are waiting for dual-core usability, another two idle cores is a pure hype like Quad-SLI.
I dont know if you didn't expect someone to read the same article, but I noticed a couple of things...
"The new 4x4 socket will help it compete for the speed crown with Intel, which is poised to launch its new Conroe desktop microprocessors in July."
That doesn't sound good for AMD.
Also I am going out on a limb here and am gonna say that I guess there having trouble with the quad core K8L, otherwise why bother with a new socket?
"The company expects to have one computer maker showing off the performance of the 4x4 socket today at its analyst briefing."
Wow one setup, and all the while Intel is set to launch Conroe in July. Looks like AMD is in real trouble.
Also...
"AMD is so smart. I can hear Intel folks crying about AMD changing the rules, "unfair" they complain."...
I dont think AMD is changing the rules, no, I think they are trying to do what ever they can to stay competitive.
I also think that a dual dual core board and two processors and one "GOD LIKE" heatsink is gonna be a fairly expensive venture.
They are slowly faidng, but they are kicking and screaming... poor AMD.
Dual Opterons are truly TWICE the CPU performance (i.e for 3D rendering) than one Opteron. Hopefully this 4x4 solution is close.
BTW Intel TDP numbers are averages. AMD TDP numbers are maxmiums. With the EE CPUs Conroe and X2 are about the same TDP (ignoring the FX-62/60 which doesn't have a EE version)
AMD represents the new schood of thinking. Intel's original plan was to pump Pentium 4 to 10.2GHZ, but AMD was doing dual core, so Intel abandoned the chase for frequency and went dual die. Now Intel has to follow AMD again. Intel will have to port its Woodcrest with blackford chipset to desktop to compete. For AMD, it's almost free. HyperTransport does all the trick.
Quadcore you ask? You will hear it more. With quadcore and 4x4 socket, you get 8 cores of computing power.
Toms Hardware
Seems Intel is set to release Kentsfield in Q1 2007.
Whens the release date for Quadcore AMD?
Quadcore you ask? You will hear it more. With quadcore and 4x4 socket, you get 8 cores of computing power.
This is great, but who needs that?
I'm having trouble to feed my Athlon 3200+ Single Core...
Unless that reverse hyperthreading is a real technology - but I don't belive it.
I don't see what a big deal this is. I know of a few people who have server boards with two dual core chips overclocked to FX speeds. I don't see why Intel can't follow suit.
personaly, i don't buy this story.. AMD is a server minded company, they have a very good 2way system on the opteron 940. i work on that btw. they wont release a 2 way system on desktop for marketing wise.... they better put 2 x 2 3800 (meaning 4cores) energy efficient together... huge performance and still nice TDP.
if they realy want the performance gaming crown again, just do like intel and a large fast cache.
on server level i am sure that the woodcrest doesn't outperform the opteron, but the mid-level conroe is price/performance(at least when they have mass produktion finaly) much better then current AMD procs. we know that they are overprised at the time, and will be lowered at certain point, but still it will be a hard performance match... maybe until 65nm or i hope not k8l
I don't see what a big deal this is. I know of a few people who have server boards with two dual core chips overclocked to FX speeds. I don't see why Intel can't follow suit.
The issue is cost. A 2P Bensley motherboard costs near $800. A 2P Opteron board about $500.
The beauty of AMD64 is that you can connect the CPUs directly to form SMP machines that scales linearly with core count. It's called glueless SMP on ccHT. I expect the 4x4 board costs around $200 or below--all you need to do is add another socket.
What AMD does is to exploit the flexibility of its CPU architecture. That brings the perfect migration path.
Basically, it is an old concept. Big-iron server does SMP with CPU boards, but AMD now does it with socket. Think of it! They are the same thing.
There are a few benefits:
Additional processing power in heand : (50%-100%)
Cost : since the price vs processing power is not linear, SMP tends to be cheap for the same processing power. It is very true in server or gaming environment, (Desktop ?)
Migration: Simple
What is the big deal ? This kind of SMP architecture is fleasible to CPU with fast I/O infra-structure like AMD Operton, but not Intel's CPU. This also explain why INTEL's CPU can not complete with AMD's CPU in 4 P server.
This time, AMD does with simple socket upgrad, and INTEL simply has no chance to copy before it fixs it I/O infra-structure.
When will the official announcment be made? I heard that it might be today, but I've yet to see some official info on it.
ASUS M2N32-SLI Deluxe Wireless Edition Socket AM2 NVIDIA nForce 590 SLI MCP ATX AMD Motherboard - Retail
$239 + $50+/- (for new socket) + $289
AMD Athlon 64 3800+ Orleans 2000MHz HT 512KB L2 Cache Socket AM2 Processor - Retail
$290 x 2 = $580
Crucial Technology Ballistix 2GB (2 x 1GB) 240-Pin DDR2 SDRAM Unbuffered DDR2 800 (PC2 6400) Dual Channel Kit System Memory - Retail
$319
Total = $1188 just for a motherboard processors and memory.
Seems pretty expensive for twin 2.4Ghz, this isn't even FX or 5000+.
Couldnt Intel modify the North bridge, split the memory and have up to 2 memory slots per socket(if they were to attempt this)?
"Basically, it is an old concept. Big-iron server does SMP with CPU boards, but AMD now does it with socket."
Except for AMD dual-processor is really ccNUMA, not SMP per se (the performance is reduced if only SMP is used).
A number of cheap 2 socket Opteron boards already do this trick. Tyan has one, MSI has one and I am sure there are others.
The upside is a bit lower cost vs. implementing a second memory bus and second set of DIMMs for the second physical processor.
The downside is that you drop down to 50% memory bandwidth compared to a full 2 physical processor implementation with one memory bus per processor.
Of course with the terribly poor or non-existent NUMA implementations of all modern PC operating systems, maybe this downside doesn't amount to much.
If AMD were a tiny bit smarter they would add a few new modes to their memory controllers... like 4X interleaving. That way you could populate all 4 DIMMs and get better bandwidth.
This type of a system is perfect for developers like me who like to do some development at home.
Right now i am running Opteron 165 with 2GB of ram. But with 4x4, i can easily see myself getting 2 x AM2 5000+ procs and dedicating 1 dual core proc to my Vmware VM running Oracle, BEA and other server software and dedicate other dual core proc to my native OS running the development and build tools.
AMD is lowering the cost of quad cores.. and i like it!
Just hear what kind of technologies AMD is going to announce during next two years. Poor Intel. I'm going to buy Pentium D or Core - as an ancient, money and marketing driven idea.
Intel is going to loose another 20-30% of the market during next 18 months.
Conroe is not going to help them even with real 20% of performance advantage.
Can you imagine: AMD has 75% shipments of HP blade!
this has KNEE-JERK REACTION written ALL OVER IT. just like p4 dual cores, perhaps more so than the p4 dual cores.
just like p4 dual cores
No. The 4x4 is like a 2P dual core server adapted for desktop needs. Low cost, small form factor. As we have seen from Opteron benchmarks, AMD's HT can allow system performance to almost double with double procs. This moves desktop performance up by almost 100% in one shot.
about multi threading usage.
its not far away. if u will look into task manager, all windows processes are multi threaded. win itself, office, etc.
non windows OSes like macosx or linux are multi threaded from the beginning.
same goes for application like photoshop or 3ds max.
software for x2 and like it is not far away. its here already.
Quake 4, Call of Duty 2, Oblivion, UT2007, Far Cry 2, 3dmark06, multimedia and archicing... are multi-threaded. Expect 4x4 to have huge advantage.
"just like p4 dual cores, perhaps more so than the p4 dual cores."
Completely wrong. They are two totally diffrent animals.
The two cores in a p4 dual-core are so far away from each other that data on one core has to travel a main-memory round-trip-time to get to the other core (roughly 90 cycles). Here the two cores are only one hop of HyperTransport link (~10 clock cycles) away.
AMD 4x4 also employs a ccNUMA-like architecture that could offer even better performance optimization.
"The downside is that you drop down to 50% memory bandwidth compared to a full 2 physical processor implementation with one memory bus per processor."
AMD 4x4 each processor has its own memory bank. See here on page 36.
Also, memory vicinity system calls already exist in Windows and Linux. The problem is software must be re-written to take advantage of them. That doesn't mean bad NUMA implementation from the hardware perspective (which consists just the MOESI protocol and HyperTransport spec - MOESI is granted a hard thing to do correctly, though).
I don't think I quite understand this product! As a blogger has stated, it's almost useless for non-servers.
Unless AMD has some magic trick up their sleave I really don't see how this product can be all that successfull to the exception of low-end server use.
about multi threading usage.
its not far away. if u will look into task manager, all windows processes are multi threaded. win itself, office, etc.
To all those who hype multi-threaded applications... I don't mean to burst your bubble but this is one of the hardest programming aspects. The results are many, many bugs OR worst dead-locks.
Its too expensive for most companies to spend time developing correct and optimized multi-threaded code. Instead they just lock entire blocks of code or entire functions which results in bad performance.
Do you think MS is competant enought to code this correctly? Only game and daemon developers will have great multi-threaded code.
I've seen some older linux code which locks entire functions of frequently used low-level libraries. This means that only one core can use these functions and the other cores sit idle.
I don't think I quite understand this product!
For those who don't need performance, a Sempron or even a Geode is good enough. For anyone wants performance, this is a dream product. The whole thing about dual core, multi-core is adding more compute engines. 4x4 is a major innovation. Currently, there is no such product for desktop. If you go the 2P Opteron route, it is expensive and doesn't support dual and quad graphics.
A lot of the demanding tasks, such as media encoding and newer games are multithreaded and multi-core capable. You can see from those Quake 4 benchmarks, dual core wins hands down. Now you got 4 cores and double memory bandwidth. Now you can finish encoding your video in 1/4 time.
AMD 4x4 will represent almost double performance for desktop. That's very tangible difference. It will take an Intel Woodcrest server to come close to 4x4.
With this announcement I would hope that for the sake of everyone, Intel would meet with Amd to talk about conceeding that Amd have the best x86-64 bit design and agree that Intel would sell Opteron/Amd64 under license.
Amd would share design details that would allow "say Itanium" to work with the Hyper transport and be a high end optional accelerator for big Iron.
Otherwise its going to be a blood-bath for Intel with only one outcome.
Thats what I would do NOW if I was the boss of Intel, there is room enough for both companies to make good money but only room enough for one to be the best.!
SERIOUSLY
4x4 is a con, I can buy 4 Dcore pentiums for the price of 2 X2 AMD64s and I will put them close together and they will co-operate to-gether on the the 1333Mhz FSB and they will kick asse... I think... gezz ok ... stupid things went up in smoke.? Duh ?? wonder how they do that.!
OK I see what you mean.. glueless CPU's is the ONLY way you could do this, man those AMD guy's are just..just so superior.!!!
"Also I am going out on a limb here and am gonna say that I guess there having trouble with the quad core K8L, otherwise why bother with a new socket?"
maybe a K8L + K8l = 4x4??
didn't think of that did u.
You might understand glueless design one day,
not Intel clueless design!!. Haha
"Its too expensive for most companies to spend time developing correct and optimized multi-threaded code."
It's time for these companies to go away. This is the evolution for higher performance.
"maybe a K8L + K8l = 4x4??
didn't think of that did u."
No I hadn't... because what I have read points to the K8L being the basis of their Quadcore design. The link below talks about what may be present on or with the new design.
Wikipedia - K8L
I understand glueless design, but just having a new socket does not mean you dont have to modify the board, right?
Most AMD people love to say that Intel doesn't have a true dual core chip, and that it was just taped together to beat AMD to market.
Well if AMD does in fact make this dual socket design are you people gonna say the same thing about AMD?
They are clearly behind on a quad core design.
If they weren't why this gimmick to try and beat Kentsfield to market?
HKEPC
Quadcore AMD release date.
Well if AMD does in fact make this dual socket design are you people gonna say the same thing about AMD?
No. You don't understand. AMD's DCA doubles the bandwidth when you add a processor, unlike Intel's shared FSB approach. When Intel tapes cores to the same FSB, the cache coherence traffic takes away more FSB bandwidth, strangle the cores even more, a primitive and inefficient solution. AMD's two socket system doesn't have a FSB. Its performance almost doubles with the doubling the socket. 4x4 is just like a 2P Opteron server, but made available to desktop environment with desktop CPUs and components.
By the time Intel releases quadcore, so will AMD. The 4x4 will have 8 cores, putting Intel at 50% of AMD yet again.
As I have tried to say plenty of times before to countless fanboys (especially TomsHardWare), AMD doesn't (And I have spoken with reps to validate this) bring out products instantly to win a crown right away, AMD works on platforms and systems to bring the best performance and performance per watt to customers based on what the customers suggest and what they say.
AMD listens to their customers to see what they want. Why did AMD release a 254 Opteron and 256? Because the customers said "we want more speed" over the 252...so AMD gave it to them. Why is AMD releasing 4x4? Because sure, AMD just won the race to "Quad-Core", but it also solidifies them in the future to have "Octal-Core" when K8L releases, once again putting Intel back of the line in the race.
AMD does not take shortcuts, they strive for excellence and bring out products consistently and on time to meet the demands on customers. Intel said themselves they rushed the Pentium D's, and most people don't realise the first P-D's were literally 2 processors stacked on top of each other. Alienware was the only vendor to sell such a disgraceful product, and it quickly vanished after Intel used their "Intel Inside" stickers to hold the cores together instead of Paul Otellini's Wrigley's Gum.
"Well if AMD does in fact make this dual socket design are you people gonna say the same thing about AMD?"
No, because AMD's Athlon64 were designed with multicore in mind. It's easy to add coprocessors and scale up their performance that right now there are companies doing just that, including AMD (4x4).
Try adding a coprocessor to any Pentium; it'll be an ugly hack that clogs the FSB. So connecting two Athlon64 with HT is completely different from gluing two Pentiums with the FSB.
4x4 is gonna be hella expensive to conroe with lackluster performance improvement...
4x4 is FX or FX2 ONLY
FX and FX2 ARE EXPENSIVE
thefore a 4x4 system will be REALLY EXPENSIVE
Post a Comment
<< Home