AMD's Barcelona to "stink" in ORACLE - Page 4

Last message on previous page:
"Surely, for applications that have lots of inter-process communication, no 8S would perform well unless say it has a 4GHz broadcast bus with"

...not without some NUMA-awareness built in to the code along with NUMA-optimized mutual exclusion primitives...but who am I to know?

But the CPU doesn't know if the memory is shared or not unless it asks every socket first.

Will this snooping function be like a SSE type instruction or is it something entirely different?

Ah thanks.


That post was just a way of changing the subject header into something less annoying.

And the only way to find out is to ask other CPUs, which consumes HT bandwidth, which is why latencies of Opteron systems increase as socket counts are increased.
Yes and no. The local processor do not "ask" other processors about current cache line state; all it has to do is to "push" (broadcast) state changes of shared cache lines to other processors.

The local processor needs to "ask" other CPUs when there's L2 cache miss. But that is not unique to Opteron. Every multi-processor system using a snooping protocol must do so, by definition. With HT, this type of messages are switched through all processors in a manner much more efficient than with FSB.

Also, "asking other CPUs" itself won't increase main memory access latency unless it's a remote cache miss, since in cHT cache coherence and memory access can go in parallel. The increase latency you see is the result of remote memory access, not the result of snooping.

With proper NUMA support, the OS will attempt to use local memory as much as possible per CPU. NUMA support is for the OS and software, not for the CPU.
Under NUMA, the OS can reserve an address space exclusively for a process. Then it is straightforward for the CPU to know that any memory access to this address space is local only.

Snooping is a protocol used by multiprocessors to track and snoop global cache states. Its operation is like caching - usually not controlled by (and transparent to) software.

LordAMD,

In this case I would have to reply that this article is useless.

Again speculation on performance Positive or Negative, AMD or Intel does not matter.

It is a matter of testing engineering samples to establish an early baseline (still not the total performer that should be released but close) and then expanding on the trended information from reputable sites.

Even then throwing out the outliers.

AMD may in fact perform well in this arena.

Also LordAMD,

Posting for the sake of posting is really not called for. A half hearted attempt at appeasing the seemingly Intel crowd is not what is needed either.

What "is" needed is factual information both FOR and AGAINST reviewed CPUs.

Early speculation is bound to happen... But it is by no means a substitute for accurate and legitimate info.

But the CPU doesn't know if the memory is shared or not unless it asks every socket first.


The MOESI protocol marks cache lines. It only has to check how the cache line is marked, not query the other sockets. If that were the case, Opteron wouldn't rule the 4S transaction world.

The cores would spend most of their time snooping HT.

That is one reason why scaling to 8 sockets is poor.

That is one reason why scaling to 8 sockets is poor.

Beyond 8 sockest, AMD uses an HT switch to keep the number of hops down to a reasonable level.

Yeah, or the Horus chipset, which had some attention a couple of years ago but suprisingly didn't make it to market. Possibly the appearance of dual-cores reducing the need of really big x86 servers.

http://www.realworldtech.com/page. [...] 202353&p=1

It's the cause of the problem. As the Quad FX is targetted for enthusiasts and gamers, its inferior performance in games versus the single-socket FX-62 is a problem.

Having NUMA has its advantages, but it also has its disadvantages in the desktop market. Clearly it was a poorly planned knee-jerk reaction on the part of AMD in trying to bring a 2-socket platform for the desktop market, giving up memory latency and thus reducing performance in desktop applications and games in order to increase throughput, though still not a match for the mega-tasking ability of the QX6700.

And I doubt if it's true of ccNUMA as implemented in Opterons.

It's about the difference between DDR2-800 and DDR2-533.

Actually, it's a lack of benefits. Desktop apps benefit more from lower latency than higher bandwidth.

I wouldn't use a Quad FX system since a) it has only 4 DIMMs and b) it doesn't support registered ECC memory. So the Quad FX is a product in search of a market. Too slow in desktop applications and games (not to mention too much power consumption and require too large of a case), while not enough RAS and memory and too much extraneous stuff for workstations.

A simple tweaks such as a complete reworking of the Opteron platform.

I hadn't noticed this trend. It's actually more the opposite, everything coming to a single socket.

This is the case in reality.

Desktop apps = games and single-threaded apps that the Quad FX is slower than the FX-62 in.

And yet no such example exists.

And you can't even fit 2 memory controllers in a normal size board now.

That is one reason why scaling to 8 sockets is poor.

That is why Barcelona will have L3. This allows better sharing of data amongst sockets. You either need L3 or a helluva fast bus as someone said.

No it doesn't.

I always thought cache was used to bring needed data/instructions closer to the core to avoid main memory latency and reduce overall demand on the bus. Needless to say, adding another level of cache simply means another cache pool that will need to be checked when sharing --- hence the drive to a new HT revision with more BW and lower latency.


That's what the extra bandwidth is for. HT latency is almost nil. Barcelona also has 4 HT links which connects 8 CPUs with 1 hop NUMA. If you ubgang you can get out to a cache coherent 32S.

It's the cause of the problem. As the Quad FX is targetted for enthusiasts and gamers, its inferior performance in games versus the single-socket FX-62 is a problem.

Having NUMA has its advantages, but it also has its disadvantages in the desktop market. Clearly it was a poorly planned knee-jerk reaction on the part of AMD in trying to bring a 2-socket platform for the desktop market, giving up memory latency and thus reducing performance in desktop applications and games in order to increase throughput, though still not a match for the mega-tasking ability of the QX6700.

And I doubt if it's true of ccNUMA as implemented in Opterons.


So what you're saying is that SSE performance doesn't count because it does have proper SW support?

If you don't have an awareness of dual+ sockets, then your code will be fragmented and it's "kernel" will not look for local RAM vs. just calling malloc.

I have long said that using SLI-like "Profiles" would allow for smart swapping so that the game will take all of the RAM in the banks local to the executing processor (QFX banks have a max of 2GB ). This would only mean slightly longer loading times as data is swapped.

Vista shows nice improvements with everything and people who buy this could sacrifice a few fps for the ability to run a virus and spyware scan while playing at 1920 (that can happen).

You have long said it, but that does not make it true.... 'what is smart swapping'? Does it start here and end there? Exactly what is it about smart swapping that makes it localize better?


Are you being funny or are you just dumb? It seems pretty obvious it means swapping one bank up to the max RAM in the profile.

For example, I have viewed RAM on Q4 (Task Mgr) and it peaked at about 1.5GB. If the bank has 2GB, that leaves .5GB for other apps. This would have to also be careful of memory fragmentation within the bank but the algorithm would not be difficult and could even be generic.

Which are almost every application.
Stop playing a dumb enthusiast, will you? Singled-threaded and game apps are not all apps. There are probably more apps optimized for NUMA than for SSE (as Baron made the good point).

That's what it was targetted for, enthusiasts. I guess until AMD realized what it a dud it was for enthusiasts.And for mega-tasking, it's slower than the QX6700, and usually the Q6600.
QX6700 can't even perform the mega-tasking workload that QFX demonstrated.

Now what are you? Mind reader? IMO you're just either playing dumb or being blind.

What does an academic research project, that has their own OS and compiler, have to do with production machines today. And it doesn't really look much different than say IBM's X3.
Search the paper for "multicast," which show an example of what I said, multicast. Again, you're playing your extremely short memory here.
BTW, no, this architecture is totally different from IBM's X3.

What about the pin count to get memory to the controllers. The Quad FX motherboard is a clear example, larger than the standard MB. Good luck fitting it into a typical Dell case.
Thousands of pins per chip is technically nothing. 10 years ago it wasn't possible to fit 1207 pins to one single socket. You're just too backward looking. Besides, hard to connect everything to one place is yet another reason why NUMA is preferable.

I doubt that considering SSE is by default the mode for floating point operations in Windows x64 mode, while many compilers are primarily compiling floating point operations to SSE ops now.

And if so many apps are "optimized" for NUMA, why does the Quad FX with a greater than 100% advantage in bandwidth lose so badly to the QX6700?
there are lots of sweet fruits but they pick the sour ones. does that make all fruits sour?

Of course it can, and as all the reviews prove it can do it faster.
what review shows mega-tasking? don't give me a list of reviews each a copy of one another. point out which is precisely measuring mega-tasking performance.

It's not hard to see with the disastrous results of the Quad FX. It's pretty easy to see the lack of markets for it, considering its excessive power consumption, excessive costs, excessive size and lack of memory slots.
you are confused of QFX's technical merit with its business/marketing ones.

It has nothing to do with what you're talking about. It's basically just like IBM's X3, its stations are equivalent to IBM's quads, it utilizes ideas similar to IBM's snoop filters.

And its from 1995.
I am talking about multicast. are you saying the ability to do multicast has nothing to do with the multicast I talk about? It may use a different mechanism but the possibility of multicast is due to its NUMA nature.
Not just IBM's X3 uses quads, so does Newisys Horus. Not just X3 uses snoop filters, so does Horus. In your superficial view, everything becomes the same.

But thousands of pins across a cheap motherboard is not. The width of memory has grown much smaller than the clockspeed speed of memory over the years.
you're again confused of the memory clock freq with access time, and the board level with chip level. The clock freq increases not because of shorter access time but more pipelined/parallel access ports.

Frankly, you were talking but not really speaking of anything material.
- An Opteron do not need to snoop another chip when it's accessing memory exclusive to itself.
- As I have said, in all cases that an Opteron has to snoop another Opteron, a Woodcrest/Clovertown with FSB must snoop another Woodcrest/Clovertown, too. The latter will actually put more pressure on the shared bandwidth and thus experience higher delay. I.e., snooping or not makes little difference between K8 and Core 2 (K8's MOESI is actually more efficient than Core 2's MESI). Snooping over direct connect arch with cHT is in fact more efficient than that over shared FSB.
- X3 and Horus have vastly different implementations. If you're talking about similar ideas, then yes, from K8 to Conroe to Power are all "the same" in this regard; but in fact they all have different mechanisms and pros and cons.
- It takes about 10 years for memory speed to double (measured by reduced access time). OTOH, from Socket 7 (1995) to Socket F (2006) the pin counts almost quadrupled.
- Your talk about business/marketing strategy affecting technical merits is just nonsense. You wouldn't have thought like this if you knew the loathe some Intel engineers feel about marketing BS (even from their own company).

It's clear to me now that you live in a totally different world from mine in terms of technical understanding. I'll only say that if you don't like QFX, then know that it was never meant to please you; if you think FSB/UMA is sufficient/efficient, then know that you'll soon become part of the forgotten past.

BS. I said "NUMA," not "Opteron." Here is my original claim: "Snooping efficiency has little to do with being NUMA or UMA. It's a factor of system architecture (FSB or direct connect) and algorithm (MOESI, etc). Further, if a process is bound to a node, and an address space is reserved exclusively for the process, then a cache miss in that address space do not need to consult remote cache at all. This is true to NUMA, but not UMA." Where did I say Opteron?

I studied the architectures of NUMA (and COMA, MPI, among others). I was debunking your false claims of NUMA not suited for mega-tasking; I was not trying to make technical inductions based on a few on-line enthusiast reviews.

QX6700 wins over QFX the same or less than Conroe over K8. The former's win is clearly on single core optimization (both software and hardware), and has little to do with system architecture. In terms of memory architecture, direct connect (plus NUMA) is huge when it comes to mega-tasking and virtualization.

BS. I was speaking of the NUMA architecture, of which both NUMAchine and Horus are examples. X3 OTOH is what you brought on to the table, and indeed has little to do with what I spoke of.

Another BS. You simply don't understand the working of SDRAM. The "10x" increased clockrate is not due to memory access speed (which increases only ~2x in 10 years), but intensive use of banking inside/among the DRAM chips. The "double-ness" of DDR and higher frequency over the SDRAM are not due to memory access speed increase, either, but increased parallel access of memory cells; such a parallelism went from 64-bit of single rate to 128-bit of twice the double rate (DDR2), or about 8x increase over 10 years.

AMD never talks about QFX the way enthusiasts like you did.

From this discussion I clearly see that you are a FUDer, nothing more. I never said QFX able to handle 4 dual-slot video cards; I said you don't need 4 dual-slots to run 4 video cards. I never said QX6700 can't handle two intensive tasks; I clearly said mega-tasking is more than running two intensive tasks.

OTOH, it was you who claimed that QFX has higher memory access delay (than QX6700) due to snooping(!). You went on to claim that it's all because NUMA that is not going to be supported, or NUMA performs poorly even if supported, or you simply don't know how could NUMA be properly supported (please at least pick your favorite...). If you only criticized QFX for its lack of business tactics, I would have no objection. But you went further to make false technical implications based on wrong (single-threaded & games) benchmarks and (lack of) business appeal. You were just plain wrong.

Besides, Tom Yager's article has nothing to do with our discussion here. Does it talk about NUMA? Does it talk about snooping? Are you sure your mind is clear, or were you on some drugs?

This is what you said.

2) I see no reason that memory latency on Quad FX would be higher if NUMA is properly supported. Maybe you can tell me how and why?

Even if NUMA is properly supported, a Quad FX's memory latency will simply be unable to match a comparable clocked single socket FX.
First, you didn't tell me why.
Second, your claims are wrong, anyway. QFX's (essentially dual-Opteron) memory latency will be smaller due to its integrated memory controller and direct connect architecture. Under high bandwidth, it's memory latency will be smaller due to its NUMA, with benchmarks that are NUMA-optimized.

The QX6700 wins on almost everything, from single applications to mega-tasking and basically shows the direct connection/NUMA advantage has little effect on the overwhelming majority of applications.
You're simply docking the real point. QX6700's win has little to do with memory architecture, but the (single) core optimization.

And neither of which have anything to do with your little flip a bit and an Opteron won't have to snoop another socket claim.
Really, how much a FUDer you're going to be? First, my claim is NUMA, not Opteron. Second, just because they use another mechanism doesn't mean 1) the mechanism I described can't be done, 2) NUMA must snoop all other sockets.

As I said, the bandwidth has increased from the increased clock rates of the memory devices, not by increasing the width of the memory connection or by adding more memory controllers.
As I said, you're confused/mixing up the intra-chip and inter-chip levels. Inside the chips, width of memory access has been increased to support the higher clock rates.

Let's read what AMD originally said about the 4X4:

New Enthusiast Platform
Building on AMD’s already recognized leadership among PC enthusiasts, AMD also announced plans for a new enthusiast platform codenamed “4x4” that will extend AMD’s long-standing commitment to those consumers who demand the highest-performing PCs. The 4x4 platform features a four-core, multi-socket processor configuration uniquely possible via AMD’s Direct Connect Architecture. The 4X4 platform will be designed to be upgraded to eight total processor cores when AMD launches quad-core processors in 2007. Project 4x4 represents system-level enthusiast enhancements and is designed for ultimate multi-tasking performance across gaming, digital video, processor-intensive and heavily-threaded applications.

I must've missed it! Where did AMD say it'll run faster than QX6700 on games and single-threaded applications?

Let me refresh your memory:

I guess those reviews were performed with 4 graphics cards performing 2+ graphically intensive tasks? Ouch, aren't they?

Or do they simply think like most people on this forum did, that hacking (C2D and C2Q) around for SLI doesn't worth one's efforts? Last time I checked quad-SLI isn't even possible on any Core 2 platform (please correct me if I'm wrong).
You FUDer, where did I say QFX able to handle 4 dual-slot video cards? OTOH, QFX is able to run 4 video cards. It's demoed by AMD. And I was right that quad-SLI isn't available on Core 2 platform. The Dell system you happily suggested was a Netburst system, not Core 2!

Let me refresh your memory:

OTOH, one would react to such a phrase so much probably because it was right on the spot for him. 4x4 and C2Q are different animals, for different purposes. AMD showcased 4x4 by running 4 heavy weight tasks on 4 screens smootly at the same time; can C2Q do so? Can C2Q even do two?
In case you're reading deficient (not to suggest you are, but you might since you really can't read what I said), I was asking, Can C2Q even do two.

I claimed that the purpose of SLI is to run games, and that games are heavily dependent on memory latency which is why the Quad FX-74 loses to a FX-62 in games. Which is true as proven out by every review. I never said the Quad FX has higher memory latency than the QX6700.
How many times do I need to tell you that those games are just not optimized for NUMA. What's wrong with your mind, something blocking its working? If you run a game on QFX with a single critical path sometimes on the first socket and sometimes moved to the second, of course its memory latency will be higher. It's not the problem of QFX, its your problem of not making the game NUMA-aware.

Your argument is stupid. First you claim QFX is (only) for games just because AMD mentioned "gamers" in its ads. Then you claim the games mean only currently available, non-NUMA-aware ones. Then you start to FUD on NUMA of QFX based on the two erroneous implications above. Yes, games performance depend on memory latency, I never said otherwise, but that doesn't make your arguments a bit more correct.

I went on to say that NUMA will not fix the memory latency issue, and the FX-74 will simply never be able to match the memory latency of a single socket AMD platform and that NUMA will not be a magic fix that suddenly improves performance in desktop applications.
You were again FUDing. NUMA will make memory latency worse if the program is not NUMA-aware. NUMA should not make memory latency worse if the program properly support NUMA. Your QFX to FX comparison proves nothing, because the games used are not NUMA aware or optimized.

It just goes to show your lack of technical knowledge when you post an article that has show a staggering number of mistakes and inaccuracies.
I quote Tom Yager's article for its description on the shared L3 cache on Barcelona. So that shows my lack of knowledge on QFX? You're just BS-ting.

Where are these benchmarks?
it's not that difficult to write one and test the memory latency yourself.

You're missing the point. The direct connect architecture barely matters in comparison to the core.
direct connect architecture matters when it comes to memory latency.

And the fact is, in an Opteron system today, each socket must snoop every other socket before it can do a local memory transaction.
the fact is, this is also true to the dual-socket Intel machines. maybe truer. so now you're not talking against NUMA, but multiple sockets on desktop? If so, don't mention NUMA or QFX, because these are not the only multiple socket desktop systems.

And you're confusing the fact that this doesn't matter, what matters is that PC designers are not willing to increase the width of the memory connections on the motherboard significantly, which would be needed if multiple memory controoler
I'll leave that decision to the market, not you.

The target is enthusiasts and gamers. The fact that its slower than a FX-62 in games makes it dead on arrival, let alone the across the board loss to the QX6700.
Can't you just answer my question straight? Where did AMD say it'll run faster than QX6700 on games and single-threaded applications? Because that's the way you described it.

It hasn't been demoed by AMD at all. The physical locations of the x16 slots make it impossible to fit 4 top-end video cards so it's a useless feature anyways. And I didn't recommend any system, that was Gojdo who only pointed out that its readily available on the Intel platform. But that's to be expected, given your lack of memory.
gOJDO or you, doesn't matter. quad-SLI is not available on Core 2, and I didn't say it's available on K8. Plus, I said 4 video cards, not 4 top end or dual-slot ones. You really can't read, and keep being a illogical FUDer.

And the way you phrase clearly suggested you didn't think it can. Despite the evidence that the QX6700 handles 4 threads better than your beloved Quad FX.

And after that you went on little incorrect rant about running 4 video cards.
The way I phrase? it's your way of reading that has problem. And just admit it, there are 4 video cards (built-in or dual-slot) in the 4x4 system AMD demoed.

Well, the Quad FX's problem is software engineering then. And of course it's the Quad FX's problem since games will probably never be NUMA optimizable.
If we run a game inside a HW VM, the game itself doesn't need to be NUMA optimized to benefit from QFX, as long as the virtual machine that runs it is NUMA-aware.
But this is not the point. The point is running a game as-is on QFX doesn't show the system's strength nor benefit. You keep talking as if all gamers want is just an increase from 100fps to 110fps - which is bull shit. Your eyes can't even tell the difference. What's far more interesting is to run one game on one (top-end) card, a few other intensive tasks on another card, and yet a few other programs probably in a different virtual machine. AMD demoed a smooth operation in a similar scenario; I've yet to see an Intel box that demos such.

I never claimed it was just for gamers. I claimed it was targeted for gamers and enthusiasts because that's what AMD announced the platform for, and what the FX series of processors were meant for over the years. If you want AMD to switch their advertising so that the Quad FX is aiming at the heater market, you should tell them that because that would be a good idea.
You claim the purpose of QFX is to run games. so... that's not just for gamers, but for enthusaists (whose purpose is still to run games) as well? please... where's your logic at all?

Where are these NUMA aware ones? There won't even be many multi-threaded games in the near future.
Games is not the only type of program that enthusiasts run. Nor is running one game with all the power of a system the only plausible scenario.

If memory latency is not higher than why does the 3GHz FX-74 consistently lose to the 2.8GHz FX-62 in games?
BECAUSE THE PROGRAMS ARE NOT NUMA AWARE/OPTIMIZED.
Gosh.... what's the problem with you?

Yawn give it up. A NUMA aware program can't stop the Opteron from snooping each other, which is what causes the memory latency increase.
Bullshit. All Intel-based multiprocessors must snoop each other; NUMA doesn't solve it, but it doesn't make it worse. I've said this to you before, but it seems your head has problem acquiring knowledge. THE CAUSE OF NUMA LATENCY INCREASE IS NOT SNOOPING, BUT LACK OF SOFTWARE SUPPORT. (Don't even mention snoop filter here, because it helps Opteron, too.)

What description? He doesn't reveal anything new is and even in that paragraph alone, has at least two mistakes.
whether he made mistake or not doesn't change the fact that the article has nothing to do with what i'm saying here, and your mentioning his article only shows a logic as messed-up as a plate of spaghetti.

What's the point, when I can see that the applications are slower.

But it doesn't particularly matter for the desktop.

Yes in general, which is why AMD's Quad FX is a knee-jerk reaction.

The market's already made the decision. Systems are getting smaller, notebooks are increasing in popularity.

The point is that it's supposed to the ultimate platform for the enthusiast, your youtube video spokesman even said it.

It's not available for any platform, unless you count the 7950 GX2s. And 4 sli-ed less than top-end video cards will get destroyed by two top-end video cards in the end.

AMD has never made such a demonstration.

Well, they sure don't want a decrease to 95fps after spending a couple of grand.

They haven't demoed anything like that. They had a WOW title screen, a couple of movies and simple encoder. Not particularly tough for a dual-core.

And for gamers its slower than the FX-62 and for enthusiasts, its slower than the QX6700.

Well, the Quad FX will be waiting a long time for the particular software to arrive.

And almost all applications will never will be. That's AMD's problem.

Intel's snoop doesn't have a significant impact on memory latency, so there is no noticeable loss in performance in games and desktop applications unlike the Quad FX.

It does have to do with your lack of knowledge.