desktop Clawhammer to feature 1MB of L2 cache

[andreasl]

Page 10 of this document

Seems clawhammer desktop will come in 2 versions. One with 256KB of L2 and one with 1MB of L2! This is interesting indeed and may indicate a direct positioning against Prescott and with the 256KB L2 version as the value processor (whether this one will be branded under the Duron marketname is not known however, probably not since AMD wants to get rid of this one)

 

[RaynorWolfcastle]

Interesting, but I wonder if this implies that it will have a huge die at .13 micron. Maybe Wingznut or pm can comment on the effect of a huge L2 cache at .13 micron?

 

[Wingznut]

1mb of cache would pretty much double the die size.

 

[CrazySaint]

Hmmm...if 1MB of L2 cache would double the size, then between the 1MB of cache, the much smaller yield rate (due to larger size) I guess we can probably expect the 1MB version to be 2-3x the cost of the 256k version. I'm kinda surprised that they are going with 256k and 1MB versions instead of 512k and 1MB versions since Barton will have 512k.

 

[glugglug]

0.13 micron is a little over 1/4 the size of 0.25 micron (have to take into account both directions).

So there is no reason a 0.13 micron 1MB cache would be much bigger than a 0.25 micron 256K cache (half what the first PIIIs had....)

1MB cache really is NOT "huge" Its long overdue that desktop processors started using it.

 

[Macro2]

Except AMD has never been able to exceed 256k L2 on a chip. Show me and I will believe.
Mac

 

[Tanked]

Glugglug - The first P3's had 512K of external cache, so it did not consume die space.

The K6-III, however, had 256K of integrated L2 cache, and was manufactured (initially) on the .25 micron process.

Wingznut, was the Pentium Pro's cache integrated onto the die, or just onto the same package? Wasn't that also on .35 micron technology?

 

[FishTankX]

The pentium pro's cache wasn't on the same package, it was in the same package. When they were manfuactured, the cache and the CPU were integrated into the same packaging. Thus, if the cache, or the CPU didn't work, they both were thrown out. Thus resulting in astronomical manufacturing costs.

I don't beliebe Wingznut was working at Intel during that era, but if you can dig any info to prove that i'm wrong wingznut, be my guest.

Essentially, there are only 2 techniques for cache on socketed processors.

Same package, and same die.

Intel tried same package. Because at that point their 2MB cache Pentium Pro's (Probably costing thousands of dollars each) I believe the cache was several times the size of the actual CPU itself. So trying to integrate the two together would spell disaster.

Anyways, the good thing about same package packaging is that you can select how much cache you want a CPU to have after it's manufactured. The bad thing is that with traces running everywhere the packaging is a pure nightmare.

 

[stonecold3169]

Do we have ANY rough estimating of cost of these guys? Will these be likely to be priced consumer level, or are we talking much larger then this?

 

[Utterman]

Originally posted by: stonecold3169
Do we have ANY rough estimating of cost of these guys? Will these be likely to be priced consumer level, or are we talking much larger then this?

AMD should have a one that will be comparitive to the price of the what the Athlon is now. I would imagine the 1 meg clawhammers would be for the profressional sector so they would be be maybe 2x-3x more than a regular clawhammer. As always, these are guesses and there is no idea as to what AMD will charge when these will be coming out.

 

[stonecold3169]

Originally posted by: Utterman

Originally posted by: stonecold3169
Do we have ANY rough estimating of cost of these guys? Will these be likely to be priced consumer level, or are we talking much larger then this?

AMD should have a one that will be comparitive to the price of the what the Athlon is now. I would imagine the 1 meg clawhammers would be for the profressional sector so they would be be maybe 2x-3x more than a regular clawhammer. As always, these are guesses and there is no idea as to what AMD will charge when these will be coming out.

Right. Basicly, what I was getting at is I can quickly see these processors getting ridiculously expensive really quickly. Originally we were going to see 2 hammer proc come out, one home and one server class. Now we have 1 home, 1 server, 1 inbetween. We all know that the 1mb cache one will beat the hell out of the 256kb one in synthetic benchmarking. Seems like maybe this would be a way to gain on Intel for their mhz myth. Consumer Joe blow sees good reviews of the clawhammer, doesn't know the difference in cache size,a nd never knows the difference. Seems like a good market plan anyways

 

[glugglug]

My understanding, before hearing the cache sizes, was that there will be 3 versions of Hammer at release:

1) entry-level Clawhammer(64-bit data bus), no MP support, marketed as Athlon

2) low-middle tier server, 64-bit bus Clawhammer, supported in up to 2 proc config, marketed as Opteron

3) high-end server, 128-bit bus Sledgehammer, supporting up to 8-proc config, marketed as Opteron



Sounds like the 1MB cache will be in #2 of the 3 above. Because of the marketing as a server proc with the Opteron name, it will be overpriced, much like a Xeon.

It is disappointing to hear there will be a 256K version...

Edit: looking at the DOC, there will actually be 5 versions:

1) 256K cache Clawhammer w/ single 16-bit HT link
2) 1M cache Clawhammer w/single 16-bit HT link (not dual supported)
3) 512K cache Clawhammer w/2 8-bit HT links (supports 2-proc)
4) 1M cache Clawhammer w/2 8-bit HT links (supports 2-proc)
5) 1M cache Sledgehammer w/3 16-bit HT links (8-proc support, 128-bit memory bus)

Note: I wouldn't be surprised to see #3 and #4 in dual config have lower AGP bandwidth than #1 & #2 because it sounds like the HT bus is split in half for communication between the 2 procs in that scenario.

 

[KF]

> When they were manfuactured, the cache and the CPU were integrated into the
> same packaging. Thus, if the cache, or the CPU didn't work, they both were thrown out.
>Thus resulting in astronomical manufacturing costs.

I'm reasonably sure they test the chips individually before they mount them. They don't have to throw them both out. It was the same separate cache arrangement with slot PIIs in the beginning. PPro = PII with a different carrier.

The separate memory chips also worked at half CPU speed or less IIRC, depending on the CPU speed. One of the ideas of a separate cache chip was that manufacturers who specialized in memory chips could make them better and cheaper. Intel bought the cache from contractors.

They also can design the cache memory in sections, with an extra section, so that a single defect can be eliminated by disabling a section. Anyway, they talk about this. I don't know if they actually do it. I am fairly sure that the big memory chip manufacturers do it. This means they can make bigger chips without drastically reducing the yield.

This chip business has a lot of unexpected and unmentioned aspects. A lot of companies design chips but own no manufacturing capabilty. ATI, NVIDIA, Matrox, VIA, SIS. The actual chip makers in turn buy a lot of their equipment from other specialty companies. AMD seems to be planning to contract out the manufacturing of Hammers. They will be paying for part of the equipment I believe. Intel does not make the silicon, or slice and finish the big 300mm wafers that chips are fabricated on; they have contractors that bid on them and have them delivered. It seems strange. I do not know if they contract the CPU packaging, but I think so.

 

[dullard]

Originally posted by: stonecold3169
Do we have ANY rough estimating of cost of these guys? Will these be likely to be priced consumer level, or are we talking much larger then this?

Here are my thoughts (not cold hard facts, but you can see my arguments and tell me if you agree):

Originally the Hammer (I'm focussing on the consumer version not the server Sledgehammer) was going to have dual processor support, and a large cache. I said it would be in the $700s range. Then we were told there will be a single processor and dual processor Hammer - so I adjusted my price range to $500s to $600s for the single processor version and $700s for the dual processor version. Now we are being told that the single processor version will have less cache - I might be willing to narrow my range down to just the $500s for the consumer single processor, low cache Hammer.

Where do I get my $500s estimate from?
1) AMD needs a high priced product since they have been losing money ever since they stopped pricing their chips at $1299. Yes, less than 3 years ago, AMD sold chips for $1299. Their last high priced consumer chip was released Oct 17, 2000 at $612. Then in 2001 they started losing money - lots of money - with top chips priced at $350. If AMD sold chips for $600+ in the past, why can't they now?
2) I lost the link, but AMD said that Clawhammer will be priced at a significant premium over the top Athlon chip. The top Athlon currently sells for $397. I don't know exactly what a "significant premium" means to you, but to me it means at least 25% more. 125%*$397 gives you about $500. Heck just a half year ago, the 2100+ was released at $420. 125%*$421 = $525.
3) Limited production. Very, very few people need a 64 bit processor at the moment. Thus the first Clawhammer will have NO advantage over the current Barton of the time. People have been under the impression that a 3400+ Hammer will be a speed king. Well that sounded great two years ago, but times have changed. Barton will be roughly 3400+ speed at the time Hammer is released. Couple that with the higher price, and there will be little demand for the original Hammer. AMD itself confirmed this recently by saying that they expect Barton to outsell Hammer throughout 2003. Without large quantities sold, the start-up costs must be spread over very few Hammer chips, increasing the cost per chip to manufacture.
4) Added expense. Hammer isn't a small chip. Larger chips cost more to produce. Plus with AMD now using some outside fabrication plants (which will need to be paid, an additional expense), AMD's chip production will be more expensive.
5) AMD said at the end of 2003 Clawhammer will barely fit into a sub $1000 computer. The other parts, motherboard/video card/Hard drive/OS/etc, costs in the $400 range leaving $600 left for a processor. This means that at the mid-2003 the processor may be even higher than $600. Thus I'm still a bit reluctant to reduce the Clawhammer price range from $500s/$600s down to just the $500s.
6) Economy picking up. I feel strongly that the bad economy lead to the low AMD/Intel prices far more than any other reason (including competition). This is since when AMD was the best competition of all, prices soared to $1299 per chip. Only when the economy slump hit did prices drop (in fact as prices dropped, AMD lost their competitive speed edge). When the computer slump ends, thus I think we will again see the days of $750+ processors.

So there are my estimates and my reasoning behind them. Feel free to disagree, but if you post your thoughts please give reasons to back them up.

 

[JSSheridan]

Interesting article. So, 256KB L2 for the desktop clawhammer. I suspect that, when they move to a 9nm process, they will increase it to 512KB. With the onboard memory controller, I wonder what the penalty for a cache miss will be. I expect it would be less than the penalty on a 133MHz FSB, but I'd like to know how many cycles it takes to recover and retrieve the data from main memory. Can anyone give me a number? If that number is small enough, then 256K would be adequate. Of course, with HT, we have seen that there is very little penalty on Intel's HT CPU's. If there is a miss, the CPU just works on the other thread instead of sitting idle. Thanks, peace

 

[MadRat]

I like the idea of how Apple uses small integrated L1 and L2 caches then uses a large high speed, external L3 cache.

 

[KF]

>Where do I get my $500s estimate from?
Sounds like a reasonable estimate, but your premises are strange. I say $500 because unless they can put out a Hammer that will be near what a P4 can do, there is no point in putting them on sale.

>6) Economy picking up. ... When the computer slump ends, thus I think we will again see the days of $750+ processors.
I agee. However I never paid $750 for a processor in my life and never will. People that buy top-of- the-line are complete idiots or have money to burn. The economy and chip sales were doing just peachy when I picked up my 800MHz Tbird for $130, and it was not far below the flagship. AMD was bragging about profits and market share.

>1) AMD needs a high priced product since they have been losing money ever since they stopped pricing their chips at $1299....
>If AMD sold chips for $600+ in the past, why can't they now?
They never sold a lot of $1299 or $600 chips. $1299 is for a prestige item which pulls up the price of the mass market chips. Where they make money is where they sell lots of chips. They are a mass- market manufacturer.

>2) I lost the link, but AMD said that Clawhammer will be priced at a significant premium over the top Athlon chip...
Of course it will. The top Athlon might only be $150 at that point.

>3) Limited production.
Yes, they will produce them in low quantity until they can ramp up to full-scale production. The longer they take, the worse position they will be in to get a better price for their chips and the harder it will be for them to get $500. The competition is in full swing with advanced P4 dies.

> Very, very few people need a 64 bit processor at the moment. Thus the first Clawhammer will
> have NO advantage over the current Barton of the time.
No advantage = low price. If it has no advantage, they won't get people to pay even a low price, let alone a premium. The only way AMD is going to produce ANY hammers for consumption is if it does have an advantage. Again, Hammer is their prestige item whose purpose it to pull up the price of other chips, until they switch production to Hammers.

>... Couple that with the higher price, and there will be little demand for the original Hammer. ... Without large quantities sold, the
> start-up costs must be spread over very few Hammer chips, increasing the cost per chip to manufacture.

No, the start up cost will be spread over the whole production run, and even the life of the company. When you spend billions, it pretty much has to be. Big corporations invest for the long term. That is the only to get big and stay there. One reason corporations lose so much money while there is a slump is because they cannot afford to quit investing in the future. The alternative is to go out of busness.

It doesn't matter what amount AMD NEEDS to cover costs. They have to sell at what people will pay even if they are losing $10000 per chip. You can't charge people what they won't pay. In other word, Hammers will cost a little less than an equivalent Intel chips. A Hammer that outclasses Intel might get a premium over Intel's top performer.

>4) Added expense. Hammer isn't a small chip. Larger chips cost more to produce.
Yes it costs more. Lets get this straight. There is no mass-market chip Intel every produced that cost them $1000 to make. $40 is more like it. (My estimate of Intel's cost comes from the guy that headed the Winchip team, and now runs what they call Cyrix since VIA bought the remnants.) In other words a big chip is not that big a liability. It doesn't mean you have to charge $500 instead of $150. It might mean you have to charge perhaps $170 instead of $150 (to stay profitable.)

> Plus with AMD now using some
> outside fabrication plants (which will need to be paid, an additional expense), AMD's chip production will be more expensive.
Nyet. Outside fabrication does not ordinarily mean higher costs. It could mean lower costs than the alternative and that's why AMD is doing it. Intel does not buy wafers from outsiders because it costs more money; they do it because it is a cost advantage.

5) AMD said at the end of 2003 Clawhammer will barely fit into a sub $1000 computer. ...

No. Some reporter fabricated a story something like.

 

[billyjak]

My old 700mz athlon on my wifes computer has 512k of L2 cache, so they did do it , and a long time ago.

 

[Wingznut]

Originally posted by: billyjak
My old 700mz athlon on my wifes computer has 512k of L2 cache, so they did do it , and a long time ago.

Yeah, but that Athlon's L2 cache was not on die. (Hence, the big Slot A packaging.)

 

[ai42]

You guys realize that a 1MB L2 cashe would most likely not be as good as an overclocker as a 256kb one right? I think it is obvious but more die space creates more heat, and unless you use rather extreme methods to disapate the heat your not going to be able to overclock worth a D&*M.

AMD really must charge more for their processors, and I think AMD is seeing this now. AthlonXP 2800+ cost $397 in 1000 quantities. Far cry from just a few months ago where the top 2200+ would only cost around $200. Quite honestly AMD needs to make some cash for them to survive, and while they were seen as the price/performance king I think that margin will decrease dramatically.

As far as AMD making more money by making chips cheaper. Well its all relative, either you sell fewer at a higher profit per item, or sell cheaper and make less profit. The trick is to find a good middleground between the two. Just AMD is rather bad at this and sell so cheap that they don't make money on the chip. Honestly sub-$75 chips do not make AMD money.

 

[Zugzwang152]

not to mention the $30 Ghz+ Durons I've been tempted to buy for so very long...just holding off on the off-chance theres a deal for less

 

[KF]

Originally posted by: ai42
You guys realize that a 1MB L2 cache would most likely not be as good as an overclocker as a 256kb one right? I think it is obvious but more die space creates more heat, and unless you use rather extreme methods to disapate the heat your not going to be able to overclock worth a D&*M.

Like things in general that seem obvious, when you think about it, it is not so obvious.

A larger die means a larger area to dissipate heat. A larger area to dissipate heat means a lower energy density (heat is energy) and therefore a lower temperature. IIRC P4 dies are said to be rather large in proportion to the number of transistors, and the reason for that is presumed to be to get the temperature lower. OTOH the first Tbred dies were expected to be exceptional OCers because people thought they would run cooler due to their lower power. But they were hotter. If you divide the power by the area, it turns out this TBred should be hotter.

We could assume the heat generated is proportional to the area (for a given technology), in which case the size of the die would be neutral with respect to temperature, but you would still need a bigger heatsink to keep the temperature the same. For on-chip memory, just as with other CMOS technology, I would expect the power dissipation to be about proportional to the number of transitions, which would be memory reads and writes. I would expect heat dissipation to be about proportional to the number of cache accesses. But memory chips are accessed one location at a time. Therefore a larger number of memory locations does not imply a proportionally larger amount of power dissipation. The area occupied by the cache would dissipate a far lower quantity of heat than the rest of the CPU chip. It should be almost the same as a smaller cache. Therefore heat generated by a larger cache is not a problem for OCing.

AMD really must charge more for their processors, and I think AMD is seeing this now. AthlonXP 2800+ cost $397 in 1000 quantities. Far cry from just a few months ago where the top 2200+ would only cost around $200. Quite honestly AMD needs to make some cash for them to survive, and while they were seen as the price/performance king I think that margin will decrease dramatically.

It is not going to drop much if they plan on staying business.

Both Athlon prices are similar to the prices of roughly corresponding P4's, which is a reasonable guess at a feasible introductory price. Therefore there is no indication that AMDs pricing policy has varied. Since the quantity of these chips sold is appoximately zero considered in the context of mass marketing, the contribution to AMDs bottom line is neglible. As a marketing tool, having chips as near to Intels in preformance is a key strategy.

As far as AMD making more money by making chips cheaper. Well its all relative, either you sell fewer at a higher profit per item, or sell cheaper and make less profit. The trick is to find a good middleground between the two. Just AMD is rather bad at this and sell so cheap that they don't make money on the chip. Honestly sub-$75 chips do not make AMD money.

I don't know where AMDs price point needs to be to make a hefty profit. But regardless of where it needs to be, that does not govern the price. It is true they have to make plans based on predictions of what the price points will be, but all that planning means nothing placed next to reality.

They have already spent the money, in advance, for the all the facilities and technical capability required to produce some very large number of chips. It took years to get it to this stage. The bulk of current spending is to service their debt, most of which is to put in place the facilities to manufacture their next generation Hammer and the more advanced versions of the present generation. Given that AMD has already spent the money, it does not help AMD to produce fewer chips than they are able, or to produce them with lower quality than they can attain. All that could do is cause revenue to drop further. It is up to the marketers now to get as much they can for AMDs inventory. It does not matter what they get for each chip, it is what they get for the whole lot. If not selling some portion at all will help, than that is what they will do, but in the real world that seldom accomplishes anything. The idea that they are selling these chip for less than they can get is unbelievable. It the price were too low, indicated by selling more than planned, they could raise the price next week. If the price were too high, they could as easily drop the price. Accumulating inventory is conceivable, since chips don't cost much to store, but hi-tech is perishable. As technology advances and the price drops rapidly, unsold inventory becomes practically worthless.