Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
Register for free to join our community of investors and share your ideas. You will also get access to streaming quotes, interactive charts, trades, portfolio, live options flow and more tools.
mmoy,
I was thinking more in terms of L1 and L2 caches than system memory and swapping to disk. I remember some comparisons showing that compactness of code, even though less efficient comes close to bloated, but optimized code, once usage of L1 and L2 is taken into effect (and multitasking, where other tasks may want to claim some of the L1 and L2 caches).
With typical memory sizes on current PCs, it is not the size of the code that is very significant. Data is more likely to be the bigger consumer of memory, when you get to 1GB or so usage.
Joe
mmoy,
I've noticed on Win32 is that the size of the executable is 60% larger with PGO. PGO adds about a 7% performance improvement on my application so I think that it's well worthwhile
But it must increase usage of L1 and L2, which apparently is fine (since you have a gain) but when the user has other apps that running at the same time, increased usage of caches will have adverse effect on other apps running...
Joe
wbmw,
Wrong. If you look up my old posts, I said I wouldn't buy at $24, but I told the AMD investors here to hold until they could find a top.
I have looked at your posts. You told muzo:
"Find a good place to sell, you won't want to be part of the group holding the bag next year when valuations drop like a rock."
http://www.investorshub.com/boards/read_msg.asp?message_id=7905326
The stock closed at $23.77 on that day.
As far as "finding a top", I am sure you will find it, retroactively, and with no money on the line.
But as of your 3 months worth of posts, since you did not put up any money to back up your ominous warnings, how can anyone take anything you post seriously? This is an investment forum. It is called InvestorsHub.
I said I wasn't interested in playing the speculation game, which is the only thing keeping AMD out of the doldrums. It's nothing more than a hope and a prayer that Intel won't win back any competitiveness in the server market, and it's based on some rather pessimistic analysts estimates on the part of Intel (about them not having something competitive until late 2007). I think it's bull, and Intel will kick ass in the latter half of this year. Now go ahead and bookmark this so that you don't mess it up next time you want to throw it in my face.
Ok, so in your infinite wisdom, you know all of this, buy you have not tried to "capitalize" on this infinite wisdom of yours. I wonder why. Is it, maybe that you yourself did not believe the stuff you posted in last 3 months?
I told you 3 months ago to put up or shut up. You have done neither.
Joe
Smallpops,
Mike, you're wasting your breath. wbmw's not too keen on 64-bit laptops because Intel doesn't have it. If Intel did have it, his argument would have changed completely to support it. Logic doesn't quite have a place in that kind of mentality.
I think you hit the nail on the head. It is not about logic. It is about supporting Intel position of the day.
If Intel position changes tomorrow, so will wbmw's.
Joe
wbmw,
Yeah, I made the mistake of believing that the stock price would lead actual results by about a year. Of course, they didn't because investors are far more exuberant than I ever thought possible. It'll make AMD a very interesting stock to play on the way down. I'm looking to buy some puts as soon as we near a top.
I see. And you will define that top retroactively, if it happens.
I thought you recommended to others to sell at $24 (while you said you have not done so yourself). Put up or shut up.
Joe
Has IPF been delayed again / cancelled yet?
No, chipguy is happy. IPF survived the first 2 business days of 2006.
Joe
wbmw,
I don't see Intel stumbling; rather, they have made a conscious decision to reach performance points through micro-architecture in the upcoming generation, while leaving the integrated memory controller for when FBD ramps up to reasonable price points and is able to take the place of direct connect DDR-n.
Those prices may prove reasonable for the server market, but I doubt there will be an option for desktop and mobile anytime soon.
If I were an AMD investor, I'd be worried about Intel beating AMD in performance prior to integrating the memory controller.
Lack of integrated memory controller makes it more difficult, which is one of the reason AMD investors have been happy campers lately.
Joe
SmallPops,
I recall reading something along those lines also. It was about the possibility of one in the future though. What current chipset could support such a thing right now? I can see it possible in the distant future with Opteron being glueless to 32-way and then throwing in the next generation Horus. But in the works or imminent?
Glueless means no chipset. As far as imminent, my opinion is that if AMD adds extra HT link (something that should not be too difficult to implement), it would take AMD to the next level pretty much right away.
Joe
wbmw,
You most certainly did doubt Centrino's success, along with a number of others, right up to the point where the data most decidedly contradicted you.
I did? Can you show me? I repeatedly pointed to Pentium M as a good architecture and step in the right direction for Intel. My criticisms had to do with Netburst and Itanium. It is good (for my net worth) that it is taking so long for Intel to realize this.
My major worry was, and continues to be, that Intel will match AMD with integrated memory controllers, and deny "free" 10 - 20% performance advantage AMD get out of them, but again, good thing (from point of view of AMD investor) is that Intel continues to stumble here.
Based on pricing, single core Turion already competes with Celeron M, so it will be interesting what happens to AMD's ASPs as Yonah ramps.
So you are expecting ASP erosion on top of loss of 1/2 of the market share?
I think nothing happens on ASP front, since Celeron M is not Yonah, and Celeron M doesn't really get much out of the new single Yonah core.
Personally, I'm expecting a Turion dual core paper launch later this quarter with availability mid-year, enough to make the first Turion launch seem on time.
On that, I have similar expectations.
Joe
wbmw,
I don't think you should go that far, but we are comparing a dual core chip on 65nm designed from the ground up for low power, vs. a single core 90nm chip from AMD that achieves low power through mostly backend engineering.
Looks like you are back to knocking Turion. Ok, so tell me by what percentage Turion shipments will drop from Q4 to Q1.
It's not hard to predict a winner here.
Well, you tell me.
By the way, do you still believe that Centrino isn't pulling its weight?
What do you mean by "still"? Do you know if I:
a) ever believed it
b) believe it now
That would make Turion a Celeron competitor.
Well, you have to start somewhere, but yeah, Celeron or slightly above Celeron. BTW, I think you are starting to see the picture.
Joe
wbmw,
Turion started from nearly 0% several quarters ago, but has risen recently to roughly 10% of the market, or about double the share that AMD had prior to the launch of Turion. I'd say that gives Intel roughly a 5% MSS winback opportunity if Turion DC is sufficiently late.
So Turion is a failure, paper launch, just a desktop chip relabeled, that will never go anywhere, never touch Centrino, and at the same time it will take Intel's latest and greatest to stop the bleeding of the market share, or at best, recapture some of the market share lost to this non entity of a chip.
You seem to be at the same time disparaging Turion as a non-entity, and building it up so that Yonah you can celebrate Yonah as some kind of dragon slayer.
My advice is, pick one line of argument and stick with it.
Now let me tell you how it really is:
The small market share that Turion took (after a very slow ramp up of the design wins, and design work around it) may be 5 to 10%, but it is not the top 5 to 10%. It is in the bottom 1/3rd of the market. All the changes that Intel does in top 1/3, 2nd 1/3 are completely irrelevant to Turion.
If Intel wants to introduce Yonah Duo to the bottom 1/3rd of the market, it may have some effect on Turion, but I doubt Intel will do that. Who would, then, pay the prices for top 2/3rds?
Joe
chipguy,
LOL. Yeah the fox really didn't want those crappy
grapes hanging out of reach in the first place.
The funny thing is that the grapes want to get to the fox as much as fox wants to get to them...
Regarding ASPs, I guess you are scraping the bottom of the barrel. Is that the last Itanium's last claim to fame?
Joe
mmoy,
I received that catalog as well.
Joe
Keith, what do you think is AMD's exposure in the mobile space with Turion DC's schedule still unknown.
If you are starting from nearly zero market share in mobile - the downside is not much. Just lost opportunity to take market share from Intel.
Joe
chipguy,
You guys still don't get it do you? In another few
quarters IPF will overtake SPARC and PA-RISC
Overtake in riding into the sunset?
BTW, thanks for clearing the deck of those (once) high end architectures. Alpha too. It is not clear Opteron could do it alone.
Not suprising AMD
partisans hate IPF so much given the ASP of IPF servers
was about 20x higher than Opteron servers in Q3.
These statistics sound like North Korea claiming that they are producing the highest tonnage of cement per capita...
Joe
LOL, like I said before this looks like it is shaping up to be a doozy of a year. :-P
No wonder chipguy is excited. 2 full days of the year passed without Itanium cancellation...
Well, those 2 days were holidays, but still, it is a success.
Joe
Keith,
I just looked at JJ's data, and there is a big jump on 12/26 for Opterons, that continues into this week (may be the same orders not yet delivered, hard to tell).
The quantities were much lower prior to 12/26.
http://epscontest.com/prices/rm_2005_12_26.htm
Also no inventory for the lower clocked A64 parts.
That's not so good. Hopefully, the situation will improve in the next couple of weeks.
Joe
Mike,
Is it just me, or did Opteron orders jump a lot? Time to check JJ's site...
Joe
Darbes,
i called for $500/share a few years back
So looks like we have an upgrade. ANALyst Darbes raises target on AMD from $500 to $600 American.
Joe
Darbes,
i see AMD at $600/share
American?
Alan,
Thanks for the link.
It looks like the prices are going to remain the same, except 940 is cheaper than 840. These should be a little more capable, due to the 2x2MB L2.
BTW, what happened to the Pressler EE, the one that was previewed last week? It does not seem to be on the list, only 2 other EE parts.
Joe
Keith,
what you´re not seeing yet at Compusa (and BestBuy) are the new AMD-based models on both the desktop and notebook side. I wonder how the picture will look in 2-4 weeks time.
Sounds promising.
BTW, did you notice that the retail version of A64 3700+ can be had for only $13 more than the 3500+ (lowest PW pricing including shipping)? You only get a 3.2Ghz P4 (640) for around $210 as the 3700+.
The difference on Newegg is now only $23. ($201 and $224). I bought 2 of the 3700 in December for $233.
BTW, on Newegg, the top selling Intel processor is 630 (3 GHz?) at $172. Top 3 are:
A64 3200 $171
A64 3700 $224
AX2 3800 $322
Not bad, if you asked me.
It looks like there will be an inevitable price cut in January, but it may be just a shift by a speed bin. Speed binning should not be an issue, since the top mainstream part (a64 4000) is still 2 speed grades below max speed bin. The top selling mainstream bin may actually be the sweet spot of the speed distribution.
The bigger challenge, especially in Q1 will be transition to of bigger percentage of output to dual core. It should be handled without much disruption if Fab 36 initial output is satisfactory. It seems (from your post about Gateway, and from other availability - HP) AMD is going to push the 2x512 version of X2 for mainstream, which obviously makes sense.
Depending on where Intel pricing of their Pressler comes out, AMD may need cut prices on X2 somewhat in Q1. Especially 3800 and 4200.
Joe
Keith,
I stopped by CompUSA today. The desktop units were back to 2:1 Intel (was a little better earlier in the quarter) but there were a few more Turion based notebooks. On thing that's new is that they now offer Acer, and there were 2 new Turion based Acer models. There is also an Avaratec, Turion based, one of those small form factor notebooks. i think I counted 8 Turion models plus 2 or 3 Sempron.
Joe
Alan,
100% more cache
33% faster FSB
13% less power
8% higher clock
While it only beats the top of the line AMD 4800+ in a few benchmarks, it is MUCH better than what INTC had at 90nm.
Kind of tells you how pathetic Smithfield is.
Joe
rlweitz,
It looks like it didn't happen today. Trading stops at noon today, right?
Joe
64-Bit Power Struggle Heats Up
http://www.eweek.com/article2/0,1895,1900721,00.asp
Itanium not even mentioned...
Joe
drjohn,
No it means when QingHuaTongFang comes crying to Intel for mermons and conroe's for next years back to school season Intel will tell them your very welcome please get in line, in the mean time please fell free to sell as much of that low margin crap as you can from that other supplier.
Oh, so you are saying that threats of reprisals were used, but had no effect... Is that a speculation on your part, or do you have an inside info on this?
Joe
Chris,
Intel said it will respect QingHuaTongFang's choice.
I wonder what that means. Does that mean that the threats, bribes and intimidation did not work, or that Intel is not going to try these tactics this time?
Joe
BTW, in that test, they call Opteron based HP workstation the "fastest PC system in the universe"...
Joe
Bobs,
One thing you need to add into the equation is packaging and testing (and its capacity), which we hear little about, but it can be a bottleneck the same way wafer processing can be.
I think it is quite possible that AMD has been running full steam for the better part of this year in Dresden, saving K7 wafers, some 130nm K8 wafers as well, and tons of Rev E wafers. Maybe it is the packaging and testing that is operating on the Just-In-Time basis...
Joe
Paul,
LOL, something for our "Dell Gamers" to read:
The Bottom Line – 4/10
On Dell’s Web site, the XPS series is marketed directly to gamers and high-performance power users. The XPS series is under the gaming tag, and they claim right on the front page of the Dell.com site that, “We take your gaming experience as seriously as you do.”
The next line: “XPS show just how serious we are” is more telling. If the Dimension XPS400 is any indication, Dell considers computer gamers a joke.
Harsh, yes. But we think it’s accurate. The system itself is a decent gaming platform and the hardware was well built. It was put together decently with parts that can pull the weight required to play today’s graphically intensive games.
But we couldn’t even install one of the most popular games on the market, Sims 2, and trying to play other popular games would lock up the system and gaming sessions, when they would run, would get interrupted.
The pre-installed programs that Dell chose to include on its computer were almost certainly the cause of all these problems, and unloading these programs from the boot-up routine fixed the problems.
Because Dell is a large company that sells many identical systems, and because their restore solution uses a hidden partition with a disk image on it anyway, this leads us to believe that Dell uses a disk image, loaded via network or CD-ROM, to set up the software on its computers.
Now, if you were the largest PC retailer in North America, you might make sure that the disc image you’re loading on the many, many computers you are selling has a configuration that works properly with the computers you sell. This, to me, would include at least a moderate amount of in-house game testing.
The fact that our very popular, mainstream games wouldn’t work on what Dell markets as its gaming line leads us to believe that they simply didn’t test this configuration extensively, and certainly not by trying to run Splinter Cell: Chaos Theory, Quake 4, or The Sims 2.
Furthermore, the fact that the video card driver they used was more than two months out of date also implies to me that it has been at least that long since they last looked at the configuration.
It comes down to this: Most people who buy a computer from Dell instead of a lesser-known builder do so specifically because they just want something that works out of the box, and the Dell XPS 400 doesn’t.
Our experience with the Dell, overall, was unpleasant and unsatisfactory. We cannot with good conscience recommend this computer to anyone.
Our goal is this evaluation, and all our other product evaluations, is to give consumers the ability to make informed choices. As a side effect of that, we hope that computer companies will also take this information to make their products better. We think that Dell can, if it chooses, rise to the challenge of improving their product line. We look forward to evaluating other Dell products in the future.
http://www.hardocp.com/article.html?art=OTI0LDg=
Joe
chipguy,
Intel's collosal and consistent profits and enviable
gross margins are the first clues anyone whose head
wasn't firmly planted up their fundament would notice.
These are functions of ASPs, volume primarily, process technology secondarily.
BTW, do you have a braille keyboard?
No. But, I always wondered if you have one of thos old style typewriters connected to your PC as an input device, one with the big carriage return lever.
Joe
chipguy,
Duh! Lower yields, higher defectivity, longer product
development and verification time, lower circuit density,
reduced circuit design options, extended and often
problematic production ramp.
Are you sure you are not talking about Intel?
Joe
alan,
Feel free to calculate the growth in network traffic with your method as you scale to 16 and 32 sockets, but I think my approximation is close enough... especially the 64X what is required for a 4 socket system...
On one hand, if you keep the number of links going from CPU constant, and than if you keep doubling the number of sockets, yes, it does converge to n^2. But on this side of infinity, it is 1x 2 in socket system with current 3 links, 1x in 3 socket system , 1.4x in 4 socket with 3 links (and cross connect. 5x in 8 socket system.
With 4 links, you stay at 1x in 4 socket system. So the challenge is to either find a way to increase the number of links and connection so that you stay close to 1, as opposed to going down the n^2 path, of find an entirely different method.
I don't believe that 16 and 32 socket systems are the plan with the current implementation. I think the IDs set a hard limit at 8 for current implementation, which says that AMD does not believe so either.
The 16 and 32 socket reference linked to Inquirer must be referring to something on AMD roadmap labeled as DC 2.0, which incidently AMD has for 2008, not 2007 as the Inq reported.
This is just mental masturbation until we see some actual 8 socket benchmarks for opteron. We know both POWER and Itanium scale VERY well going to large SMP systems.
AMD's current ambition ends at 8 glueless. The first Tier 1 to implement that may be Sun. Whether they do it with current Socket 940 or the next Socket 1207 is a question. If Socket 1207 adds an HT link, it would make a lot of sense for Sun to wait until the new socket arrives.
One class of apps where 8 way systems based on the current system are likely to challenge the interconnects to the max are databases (the most important app for servers). Coherency traffic is an issue, but a small one. The bigger issue likely to be huge memory transfers, because the data is predominantly global, rather than local and it may need to be accessed at distance of up to 3 hops. This is the other extreme of Spec_Fp_Rate.
BTW, because of this issue, I believe that AMD will increase the number of links to 4, along with slight increase of the speed of the HT links in the new socket. If my assumption is correct, the change will make Opteron undisputed champ in 4 to 8 socket space even after Woodcrest ships. 1 to 2 socket will be more up for grabs, since they will depend more on CPU performance than on memory bandwidth and interconnect (where Opteron will retain the advantage in larger 4 and 8 way systems)
Joe
Alan,
I think you have shown a good case where 8 socket can be done, but may require some improvement.
That's not correct. 8 Socket Opteron systems can be done and are being done without any improvements. They certainly need no improvements to beat 8 way Xeon based systems silly in number of areas. See Spec_fp_rate for an example.
4 Socket 2400 MHz Opteron 153
8 Socket 3000 MHz Pentium 4 Xeon MP 41.1
8 Socket Xeon systems certainly don't look viable here. It is (or was) being not because it passed the test of viability, but because it is Intel.
That's obviously one extreme, where you have localized data, and large bandwidth requrement. On the other extreme, where you have global data, Xeon could even reach competitiveness with Opteron (without any improvements).
So your original statement:
The broadcast coherency method AMD uses runs out of gas once they get up near 4 sockets, which is why even though it currently allows 8 sockets you don't see many of those systems out there.
and your modified statement:
8 socket can be done, but may require some improvement
are not much more than FUD.
You don't see very viable Opteron system, at the same time you do see less viable Xeon system is not because of the cache coherency, but because Intel has a head start in the market, allowing inferior Xeon to be in this market, while superior Opteron is not.
Remember, this started with a claim regarding 16 and 32 socket servers.
True, but I think that pertains to either 32 core system, or something that AMD calls Direct Connect 2.0, coming apparently in 2008, and none of us knows what it is about.
I certainly don't believe that 32 socket is a viable option with current strategy (it is not even supported). 32 socket and 16 socket systems are not very relevant, IMO. You need to think in terms of 128 socket systems today (non-clusters). How many are sold? 10s? 100?
That will be the size of the market of 32 socket servers once quad core mainstream CPUs are on the market. That market is going to implode.
So to address that market, AMD would most likely need to change the cache coherency but AMD will do that only if changes to cache coherency would bring measurable improvements to 8 socket systems (with possible 4 links). Otherwise, why bother? Certainly the market for 32 socket system (= todays market for 128 socket system) is not really worth bothering with. Well, not unless the roles and server market share of Intel x86 + POWER + Sparc + Itanium vs. Opteron are moving toward reversal of current situation <g>.
If we carry out the table with the other configurations we get:
2 socket = 4
4 socket = 16
8 socket = 64
...
Obviously you ignored my point that the number of connections in the system reduces the rate of growth of the bandwidth necessary for coherency, and that it is the number of these necessary connections would make the system impractical to build. This is true even more now, when we can go the other way, doubling number of cores, which is easier, cheaper, and has very broad market appeal (beyond tiniest niches of the server market).
Joe
bobs,
I don't know if it was Q2 CC or Q3 CC (AMD increased inventories in both quarters), but in one of them it was stated that the increase was mainly in CPUs.
Joe
highlandpk,
that could change significantly when 32 socket glueless SMP arrives.
I would rather think of 32 core glueless, which I can readily envision.
Joe
Alan,
2 sockets ( 2 requests (one per socket) 2 answers (one per socket ) total 4 messages
4 sockets (4 requests, 12 answers (3 per request X 4 requests)) total 16 messages
8 sockets (8 requests, 56 answers (7 per request X 8 requests) total 64 messages.
You have to also consider the #of connections (you can consider it a unit of bandwidth) as well. There is one additional complication, which is number of hops, that works to saturate these connections more.
Also, IMO, it is better to think in terms of "traffic" rather than requests and answers, because what you call request, meaning a requesting processor sends a message to every receiving processor, the amount of traffic for requests is the same as for the answers, and since we are dealing in orders of magnitude, you can just think of one way traffic.
For a 2 way system, you have 2 units of traffic, and 1 connection. (2 / 1 = 2)
In a 4 way system (with a cross connect), you have 14 units of trafic, 5 connections (units of bandwidth) (14 / 5 = 2.8)
In an 8 way system, there are 100 units of traffic and 10 units of bandwidth (100 / 10 = 10)
So, it does jump. But the increasing number of connections reduces the rate of growth. For example, to go from 2 way to 8 way system you go from ratio of 2 to 10, or 5x increase, rather than your example where you went from 4 to 64, or 16x increase.
Further, you need to consider how saturated the connections become with these coherency requests. If it is 1% in 2 way system, going to 5%, we don't really have a problem. If it goes from 10% to 50%, it may start to become a bottleneck, since the HT bus needs to respond with data from distant processor request as well.
If you consider the possibility of adding 4th link than the data ratios become:
2 socket system: 2 / 1 = 2
4 socket system: 12 / 6 = 2
8 socket system: 84 / 14 = 6
Meaning 4 socket system would have equal coherence traffic as 2 way system (no penalty), 8 socket system would have only 3x coherence traffic vs. 2 way system.
Anyway, it would be helpful if we had some 8 socket data which would speak to the ability of the current HT implementation to handle this much data. It also gets to be more important for the OS to be more NUMA aware so data can be stored in the local memory rather than in the memory of a remote CPU, which would generate even more HT traffic.
True, and it also depends greatly on types of app. So the answer is not really simple. Not as simple that HT coherency strategy is inadequate and runs out of steam.
Also, note that just because Elmer said something does not make it wrong
Not at all, but you have to consider that he perfected the art of FUD, and clearly states it (in his handle), so you can't really get angry, or suspect any kind of deception.
When it comes to AMD, when AMD does not do something, Elmer says AMD can't do it - until proven otherwise, by which time he moves on to the next subject of what AMD can't do.
Well, the answer to the question why AMD does not do something does not automatically mean it can. It can be that AMD has not been presented with (or has not won) the opportunity to prove that it can.
Joe
Bobs,
Yeah, that's just one of the many things that causes a gut feeling that while AMD isn't in desperate straits in the supply area, fab36 production could have been put to good use awhile ago. From my point of view it does look like AMD has done a creditable job in managing the capacity issues. However, one has to wonder if the spot outages at places like Best Buy that are being reported would have occurred if fab36 had come online 6 months ago?
I think it is hard to call it a shortage, if CPUs are available everywhere 10 days before Christmas. Mike's data shows that inventories are depleting fast, and some spot stretching of deliveries. But again, I think it is good to see that 10 days before Christmas.
It is hard to reasonably argue that after 26% QoQ growth and another strong growth quarter in Q4, that something should not have been left on the table.
BTW, I think AMD prepared for Q4 by building up inventories throughout the year. Even in Q3, when AMD experienced 26% growth, inventories grew. Seeing those inventories depleted would not be such a bad thing.
BTW, a side benefit of selling out your inventories (in addition to cash) is that in the next quarter, AMD will be selling largely the freshest dies, with the latest process optimizations build in. Also, when the new core revision hits, there will be less of an overhang of the older rev chips.
Another benefit is that it is a good situation to be in (low inventory) in the quarter when additional capacity is turned on.
Joe
Alan.
Why has AMD increased HT speed?
Actually, I got the broadcast cache coherency scaling from an AMD fan on another site. He is generally correct when it comes to server information, so I trust it to be correct even though I have not seen nor done the math myself.
Sorry, but I have to LOL. When pressed, Elmer used exactly (nearly word for word) reply. It must just be a complete coincidence.
In terms of competition in this market, I believe most move to some sort of directory cache coherency scheme. For instance, the IBM hurricane chipset has a "bunch" of DRAM in the northbridge to keep track of cache coherency data to reduce snoop traffic.
I believe all the large systems use some sort of directory, filtering, or snoop reduction. AMD could implement such a thing, but I believe they will leave that up to the big system vendors like SUN.
Overwhelming bandwidth works too. It is already implemented, in place. It works. It is proven. It is a little hard to argue with success here.
The overwhelming bandwidth can be made even more overwhelming reasonably easily.
The bottleneck of going beyond that is not really in the scheme, but in the challenge of wiring all these connections. I think that with 4 HT link CPU, 8 socket system would have the same performance characteristics as 3 HT link 4 CPU socket. But you will have to have up to 14 links in that 8 CPU 4 HT link system. It could still be modular (one board with 4 CPU, second board with 4 CPUs, 2 boards connected with 4 HT links. The base board would have 4 internal HT connections (piece of cake), the extension board would have 6 (workable).
Joe