• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

Opteron's big weakness

X

xonik

[H]F Junkie
Joined
Jun 20, 2001
Messages
10,318
The way I see it, AMD has no designated competitor for Intel's Xeon MP series. Instead, they take a vanilla Opteron with 1 MB L2 cache and add 8-way scaling to it through a simple microcode change and by not disabling a few HT links. While that's working pretty well right now, how will it compete with the Xeon MP in the future? And how will AMD rise to the challenge of Itanium 2? These are questions that server buyers are probably asking...what is AMD's answer?

Here's my suggestion to AMD: introduce 2 and 4 MB L2 or L3 cache Opterons to take on Xeon MP at the 4+ CPU market. It'll widen the performance gap in some areas and close up some gaps in other areas. If AMD really wants to do well in the next few quarters, it's going to need to win some more purchases in the enterprise world. That calls for substantially higher performance not only now but into the foreseeable future. If they can't take on the big iron, at least AMD should increase its presence in the 4-8 CPU scene.

So do you think that AMD needs to take further action? Do you have a better idea? Discuss.
 
I totally agree. This is one business where you can't sit still for long, if at all. I like your ideas about the extra cache. I'd also add faster system buses. I know AMD is moving the Hypertransport bus to 1 ghz with the release of socket 939. They should continue to increase the HT bus as needed.

The most important thing AMD can do as far as I'm concerned though, is to create their own compilier with all the necessary options for full 64 bit optimizations.
 
the problem with making a 4MB L2 or L3 cache, is that it would make the die size much, much bigger. It would theoretically have about 250 million more transistors!:)

I dont think that would be a cost-effective strategy for them trying to sell them at a competitive price to the xeons
 
That's to be expected. Intel pays the price for the additional transistors, too. The thing is that AMD can mark up the price significantly and still be cost- and performance-competitive with the Xeon MP, which is like a year behind compared to the Pentium 4 in which it is mostly based on. If buyers ate these large-cache Opterons up, AMD could rake in some serious coin and further develop other weakpoints in the company like marketing, software support, and mobile and embedded PC solutions, not to mention the obvious production capacity shortcomings.
 
yeah, you're right about the cost-effective part, I just realized how wrong that was.
 
I mean, when it all boils down to cost, how much do you think a brand new Xeon MP costs in the factory? $40? $50? Even if it was $100 per CPU, the chips are selling for $3000 or more. That's more than enough to recoup losses from bad dies in the wafer, operate the fabrication facility, validate each CPU, pay the workers, etc. etc. -- and that's just one CPU.
 
I don't think AMD could get high yields from making 2mb or 4mb L2 cache Opterons. Plus, it would be more profitable to focus on the desktop consumer market. The high-end server market is only a fraction of the total market.
 
I'm not in a position to speak on the nuances of processor yields, so I won't comment on your first statement. However, I wholeheartedly disagree with your second statement. AMD might not sell as many Opterons as it does Athlon 64s, but which ones make more money, do you think? And most server operators purchase two or more CPU per computer, so it adds up. Now introduce a line of heavy hitters that can be marked up to Xeon MP levels, and imagine the profits that could be gained from that. And what about those supercomputer projects that employ dozens of CPUs? Opterons have found their way inside several major supercomputer projects, but what if they were "Opteron MPs" instead?
 
Originally posted by DaveX
I don't think AMD could get high yields from making 2mb or 4mb L2 cache Opterons. Plus, it would be more profitable to focus on the desktop consumer market. The high-end server market is only a fraction of the total market.
Click to expand...
If AMD can start selling the Opteron's at the same price premium Intel has, I don't see the problem.

AMD needs more industry recognition to sell more mainstream parts. Bob(names changed to protect the fact that I don't know the guy), the tech guy for GMC, buys Intel based servers and recommends Dell Intel based clients. If they went with AMD for the servers, the dealerships would be more likely to go with AMD for the clients.

High and low end in one fell swoop, if it all worked out that is...

As is said regularly, noone has ever been fired for buying Intel. AMD needs to get the same rep in servers. That mindset is more likely to trickle down than up.

AMD has plenty of end users pushing for it, in order to get numbers behind them they need enterprise solutions. They need their name to be associated with big money servers.

Tell me which sounds better to you, a kid down the street is running an Athlon FX chip, so buy an Athlon based server, or a major business depends on AMD for their server needs, so buy an AMD based desktop. The difference in reputation is what AMD needs to get straightened out.

There are still tons of people that are scared of AMD for workstations, asking if they can run the same programs, etc... I've sold a few and they were very happy with the end result.
When the price difference and name recognition left the discussion table, they had a damn nice computer, and they'd never heard of AMD before.

In the eyes of big businesses a lower price means unproven reputation. It is as if the higher priced part is worth the money, but the cheaper part is just cheaper. If it is faster than an equivilent clock speed Intel system and cheaper, there must be something wrong with it.

This is not the case, but that is how most businesses see it.

It is to the degree that rather than buy an Athlon XP 2400+ system they buy a Dell with a 2.2GHz Celeron. The XP is signifigantly faster, in office apps no less, but the Celeron is chosen on name recognition alone.
 
Hmm Well I agree and don't agree First.

1, Intel BUS plainly sucks ass because they still share the same pipe for memory. So The more Xeons you stick on a mobo the slower memory transfer gets. Hence you only get 5% increase for adding an extra CPU.

2. HyperTransport gives scaling a much needed performance boost. The more opteron's you got will not decrease your memory through put, it will increase it 4 Ways opterons gives you close to 10 gig/s of memory bandwidth.

3. Cache is not the only answer, infact I would rather see Opterons licence Hyperthreading. Then you got 3 FPU Units and 3 integer units being 100% utilized. I also rather see an extra FPU/ALU unit rather then cash. Give the motherboard L3 cache thats DDR2 or SRAM of 2 Megs and (just like the k6-2 days) thats capable of running core speed of cpu.

4. L2 / L3 Cache is nice but L1 is the best. Better prefetching more code is ready for direct processing etc.

5. An opteron can deliver 80% itanium performance at fraction of the cost a top of the line Itanium server/workstation will cost (Hp build) about 28k american I can get 4 Opteron machines of the same spec and better the itanium for the same price. Cost of ownership is also much lower.

6. Itanium Bus is not scalable either so its very poorly suited for scaling and supercomputers.

7. The new Xeons don't look like true 64bit proccessors either, they only have memory extenshions. No where does it say it can do 64 bit math.
 
Yeah, Wrench00 has got it, the XeonMPs need that big L3 cache because their memory bandwidth is so poor. Also bear in mind that P4s/Xeons use an inclusive cache system, so the effective total cache of the XeonMP is 2MB, whereas the Opteron has 1MB L2 + 128K L1 for 1.152MB, which isn't such a huge difference, especially when the Opteron's performance is affected far less by cache misses than the Xeon's.

Seeing as an 846 is literally half the price of a 2.8GHz XeonMP, the price/performance you get from a 4way Opty makes the Xeon look like a joke right now IMO.

Rich :¬)
 
Most of the important reasons for Optys NOT having large caches have been touched here, but to recap:

1.) Opteron scales MUCH MUCH better than Xeons. The more you add, the more memory bandwidth you get. Opterons are also glueless up to 4 way. This drastically decreases the system cost and complexity. HyperTransport is a great interconnect. It will only get faster.

2.) AMD chips do not like large caches. The way an exclusive cache works (which is what all the current AMD have), a large cache can actually hurt performace. This has to do with L1 latency, L1-L2 write back, the Victim Buffer, etc. 2 MB for cache would be abou the max for a usable Opteron and there isn't really a need for it.

Really, the Opti's kick the Xeons all over the map on performance. I would like to see SSE/SSE2 performance increase (like increasing the Wallace tree to 128 bits), but overall it is just about perfect.

Partial or full multicore will come along soon enough. HyperThreading has some serious problems in the server arena (one of the two threads hogging the cache lines and slowing the whole thing down, etc.).
 
I don't know too much about this, or what a corporate market looks for, but the Opteron has a huge lead in this cost-segment of servers. AMD needs to build reputation, not in the desktop markets but rather the high end markets. The Opteron has had a few good sized wins. And when cray can finish redstorm, that's a pretty major win for AMD. When a company sees the world's highest performance super computer being powered by an Opteron, that has to mean something.

As for hyperthreading, it would require an enourmous amount of work to implement and a very small payoff. Instead, AMD needs to work on dual-core chips (which they are). Getting these to the market this year would be absolutely awesome (but as of now they're saying late 2005 i think).

Now that intel has officially announced AMD64 support (IA32e? how confusing) and consequently the rest of the market :rolleyes: AMD's chips could potentially shine if they best Intel's new Xeon. The question of architecture support isn't really a question anymore, as some of the big software producers will start porting and optimizing AMD64 programs.

As for very long run, i'm not qualified to comment on this at all, but wouldn't switching to MIPS or Sparc be extremely beneficial? I know AMD owns rights to use atleast MIPS.
 
There's absolutely no need for more cache for the Opteron as it has been already pointed out in this thread. Opteron's architecture with local memory for each CPU combined with HT links takes care of that.

Xeons need cache since they are bottlenecked.
 
I agree with oldman.
If a coporate exec uses amd for his servers the computer illiterate guy down the street will be like, "ok i guess i'll use them to."

If i say, "use AMD it's cheaper and better then a celeron."

I'll either get, "AMD???" or "are you sure?" or "No i think I'll stick with intel."
 
Originally posted by xonik
I mean, when it all boils down to cost, how much do you think a brand new Xeon MP costs in the factory? $40? $50? Even if it was $100 per CPU, the chips are selling for $3000 or more. That's more than enough to recoup losses from bad dies in the wafer, operate the fabrication facility, validate each CPU, pay the workers, etc. etc. -- and that's just one CPU.
Click to expand...

I would imagine it would highly depend on yields.

Low yields could really hurt them badly.

Also, they are trying to get away from multiple cores... they just don't have the resources. They already are manufacturing K7 and K8 cores, which is probably stretching them pretty thin as it is.

At this point, I'll take a wild guess and say that all of the Athlon 64's they produce go like this:

Opteron Test(s) = PASS (Bin as Opteron)
FAIL Opteron Test(s) = Run Athlon FX Test(s)
Athlon FX Test(s) = PASS (Bin as Athlon FX)
FAIL Athlon FX Test(s) = Run Athlon 64 Mobile / DTR Test(s)
Athlon 64 Mobile Test(s) = PASS (Bin as Athlon 64 DTR / Mobile)
FAIL Athlon 64 Mobile Test(s) = Run Athlon 64 Test(s)
Athlon 64 (1MB L2) Test(s) = PASS (Bin as Athlon 64)
FAIL Athlon 64 Test(s) = Run Athlon 64 (512K L2) Test(s)
Athlon 64 (512K L2) Test(S) = PASS (Bin as 512K L2 Athlon 64)
FAIL - Recycling or disposal of core materials.
 
Originally posted by DryFire
I don't think the tests would go from socket to socket or am i mistaken?
Click to expand...
chances are you're correct, but the Opteron and Athlon 64 cores really are very close to each other, it's probably quite possible for a core built for 940 pin Opterons to work perfectly on a 754pin setup (the core, not the packaging).

wait a second. i'm going to contradict everything i just said - after they physically inspect the cores for damage don't they have to package them before they can test them any further, therefore meaning testing has to split between 940 and 754.
 
It's not actually possible for AMD to say a 940 FX didn't pass muster but does as an Athlon64. They are two different cores, different numbers of transistors, and different core sizes. If they tried to toss an FX on the 754 packaging, it would be easily noticeable.
 
Originally posted by SmokeRngs
It's not actually possible for AMD to say a 940 FX didn't pass muster but does as an Athlon64. They are two different cores, different numbers of transistors, and different core sizes. If they tried to toss an FX on the 754 packaging, it would be easily noticeable.
Click to expand...
What's the difference between an Athlon64 FX and Athlon64? 128-bit vs 64-bit memory channel?
 
Opteron's big weakness......nothing any of you guys mentioned.

Its big weakness is it doesn't say INTEL on it.

Peeps like Dell and Gateway are so far up Intel ass it isn't funny. Seems really stupid to overlook a superior chip now doesn't it?
 
I think some people are missing the obvious fact that this is amd we are talking about. They always seem to snatch defeat from the jaws of victory. I love amd, all my systems are amd based (although i had a couple intel rigs back in teh late 90's) so please don't think i am slamming them for brand loyaltys sake. I'm just saying they have a way of self defeating just when it seems that they are pulling ahead. Why no sub 200$ a64? If i could get a 754 cpu for around 150-175 with aheatsink i would probably buy one, but 200$+ is to much for something with a obsolete socket.
 
Most of the important reasons for Optys NOT having large caches have been touched here, but to recap:
Click to expand...

Actually the biggest reason is that AMD simply doesn't have the fab capabilities to do it, one fab can't support 3+ different lines of cpu's, most of which get an order of magnitude less die's per wafer than intel
 
Originally posted by SnakEyez187
Actually the biggest reason is that AMD simply doesn't have the fab capabilities to do it, one fab can't support 3+ different lines of cpu's, most of which get an order of magnitude less die's per wafer than intel
Click to expand...

We'll see how that changes since IBM can out-fab Intel anyday.
 
Originally posted by Morley
We'll see how that changes since IBM can out-fab Intel anyday.
Click to expand...

If IBM uses all of their capabilites to fab AMD chips, and stop making their own, and the deozens of other companies that depend on them, which they won't, and they aren't going to ever
 
Originally posted by SnakEyez187
If IBM uses all of their capabilites to fab AMD chips, and stop making their own, and the deozens of other companies that depend on them, which they won't, and they aren't going to ever
Click to expand...

In terms of technical abilities, IBM is further ahead of Intel by several years. That's why NVIDIA and AMD have gone to them.
 
just a reminder, but, Xeons render Maya animations, 3dsMax animations, etc much much MUCH faster than Opterons, due to advanced SSE2, and even because of HyperThreading. I've heard in some cases, it's been as much as an hour faster PER FRAME, in a scene dealing with tons of textures, polygons, NURBs, and lighting. But, Opterons CREAM Xeons in Photoshop and Content production. It really just depends on how you look at it. Opteron has as many weaknesses as the Xeon, and either's weakness is the others strength. So, to say that Opterons absolutely destroy Xeons, and vice versa, is like saying my goat is better than your '87 Tercel, their is just no ground for arguement.

So please, if you are going to say Opterons > Xeons, or Xeons > Opterons, please define on what basis you are talking about, cause if you say Opterons are better than Xeons (and just that statement alone), I can easily say Xeons creamate Opterons in SSE2.
 
Originally posted by rayman2k2
just a reminder, but, Xeons render Maya animations, 3dsMax animations, etc much much MUCH faster than Opterons, due to advanced SSE2, and even because of HyperThreading. I've heard in some cases, it's been as much as an hour faster PER FRAME, in a scene dealing with tons of textures, polygons, NURBs, and lighting. But, Opterons CREAM Xeons in Photoshop and Content production. It really just depends on how you look at it. Opteron has as many weaknesses as the Xeon, and either's weakness is the others strength. So, to say that Opterons absolutely destroy Xeons, and vice versa, is like saying my goat is better than your '87 Tercel, their is just no ground for arguement.

So please, if you are going to say Opterons > Xeons, or Xeons > Opterons, please define on what basis you are talking about, cause if you say Opterons are better than Xeons (and just that statement alone), I can easily say Xeons creamate Opterons in SSE2.
Click to expand...

Opteron is faster in Maya, SoftImage, and is equal or better in Lightwave. When someone wants a 3D Studio Max machine, we sell them Intel. In most other cases, AMD. We do our own in-house testing of these systems, and have a large customer base that gives us feedback.

www.boxxtech.com

(Our machines are used throughout the digital content creation world, including major special effects studios, gaming studios, and post production houses.)
 
Well, we don't want AMD to become IBM's bitch. By the way, I thought Adobe was very heavily biased toward Intel. How the hell does the Opteron beat the Xeon in Photoshop?
 
Originally posted by Morley
In terms of technical abilities, IBM is further ahead of Intel by several years. That's why NVIDIA and AMD have gone to them.
Click to expand...

What? IBM still hasn't gotten low-k on any of their chips yet, and has horrible yields for many customers. Nvidia is actually running back to TSMC for it's products because IBM can't meet demand and deadlines, and requirements for future products. iIalso believe there are many technical reports and awards for Intel having the most advanced and knowledgable fab workers on the planet. if you're going to say "several years", explain exactly what IBM has that Intel doesn't (i'm guessing your first and probably only answer is going to start with an s)

By the way, I thought Adobe was very heavily biased toward Intel. How the hell does the Opteron beat the Xeon in Photoshop?
Click to expand...

I thought they were biased for Apple
 
Originally posted by SnakEyez187
What? IBM still hasn't gotten low-k on any of their chips yet, and has horrible yields for many customers. Nvidia is actually running back to TSMC for it's products because IBM can't meet demand and deadlines, and requirements for future products. iIalso believe there are many technical reports and awards for Intel having the most advanced and knowledgable fab workers on the planet. if you're going to say "several years", explain exactly what IBM has that Intel doesn't (i'm guessing your first and probably only answer is going to start with an s)
Click to expand...

They do have Low-K, and yes, SOI's another good thing. They were first to copper as well.
 
Originally posted by DaveX
Well, we don't want AMD to become IBM's bitch. By the way, I thought Adobe was very heavily biased toward Intel. How the hell does the Opteron beat the Xeon in Photoshop?
Click to expand...


the new Photoshop CS (7.0)


Morley, are you sure? Cause in my experience, Xeons beat Opterons clean out of the water.



oh wait, then again, i am using a SSE2-based render proggy, but regardless, SSE2 is much more stronger on the Xeon platform than Opterons.


EDIT: And is Boxxtech your company/do you work for them, cause if so, you might see a couple of orders heading my way;)
 
Originally posted by DaveX
Well, we don't want AMD to become IBM's bitch. By the way, I thought Adobe was very heavily biased toward Intel. How the hell does the Opteron beat the Xeon in Photoshop?
Click to expand...

IBM's bitch....lol. Remember the old days when you had an IBM or an IBM/compatible and not a PC.....lol. I think Intel and AMD were both IBM bitches then!
 
Maybe IBM is still sore about losing the PC industry in the '80s. Maybe by controlling AMD production, they are secretly trying to get back into the PC industry. OK, that's my conspiracy theory. AMD is lucky the German government is investing so much money into their fab at Dresden.
 
i just want to point one thing out about itanium scalability, every 4 processors has their own snc and their own bank of memory, so that while not as efficient with the memory bandwidth as the opteron, the bandwidth situation isn't as horrible as you make it out to be in massive smp systems.
 
IBM appears to have good yields on the G5 and have apparently already segued to a .09micron process.
I wouldn't call them ahead of Intel but their fabs are some of the best in the world.

regardless, SSE2 is much more stronger on the Xeon platform than Opterons.
Click to expand...

Vectorized SSE2 is clearly better on the PIV architecture as a function of clockspeed.
The A64/Opteron platform on the other hand, blows the Xeon/PIV's out of the water when it comes to Scalar SSE2.


The Itanium and the Opteron are positioned in quite different segments. While it's true that the Opteron scales superlatively, the quad-capable Opterons lose a bit of their advantage due to AMDs pricing policies.
An Opteron 848 retails for about $1500 for one processor, which isn't a bad deal compared to quad Xeon or Itanium processors, but still somewhat less compelling than if it were $1,000.
I think AMD should drop their prices to gain additional market penetration since they need more Opterons out in the wild to encourage AMD64 development.
 
Originally posted by rayman2k2
the new Photoshop CS (7.0)


Morley, are you sure? Cause in my experience, Xeons beat Opterons clean out of the water.



oh wait, then again, i am using a SSE2-based render proggy, but regardless, SSE2 is much more stronger on the Xeon platform than Opterons.


EDIT: And is Boxxtech your company/do you work for them, cause if so, you might see a couple of orders heading my way;)
Click to expand...

I'm an Account Executive there, lol...
 
Originally posted by Morley
They do have Low-K, and yes, SOI's another good thing. They were first to copper as well.
Click to expand...
...and first to practical implementations of the 90 nanometer process. Some of Xilinx's FPGAs were first produced at the East Fishkill facility in mid-2003.
 
You must log in or register to reply here.
Back
Top