SiS645 chipset, are you tempted to turn to Intel again?

[CougTek]

Hi,

Like most of you, I read a whole bunch of reviews about the latest DDR chipsets for the Pentium 4. If you put aside the temporary DDR SDRAM price jump, the new socket 478 DDR chipsets make it possible to build a P4-based system at a fairly competitive price, even compared to Athlon XP-based systems.

Among the new DDR chipsets - Intel i845D, VIA P4X266A and SiS 645 - the most interesting seems to be the SiS one. Cheaper than the Intel offering and almost as fast as the VIA at 266MHz (faster at 333MHz) and absolutely legal, it looks like THE obvious choice.

According to a comparative review at X-bit Labs, the Microstar 645 Ultra is arguably the best motherboard using this chipset. Due to a recent price drop at a local store, it's possible to buy this motherboard (MSI 645 Ultra) for a mere 125$CDN (~80U$). I must confess that even if I don't really like the P4 for reasons I've already stated more than once, I'm very tempted by a P4 + MSI 645 Ultra system. And I'll be a lot more when the Northwood 1.8GHz (significantly cheaper than the overpriced 2.2GHz model) will come out.

Am I alone among the usually Pro-AMD to be eyeing the Intel side since the SiS645 came out?

BTW, I didn't read any review compairing the performances of the original SiS645 release against the new "A2" release. It's supposed to be faster, but how much more? My guess is that it's nothing very significant but I might be wrong.

 

[Will Rickards]

The price difference is becoming very small.
I'd chose AMD for my sytem... maybe P4 for a family system.

 

[Tannin]

Time I checked my pricing again, I can see. But last time I looked, the P-4 was not even in the ballpark. I don't mean a little down on the price-performance scale, I mean way down. Even a crappy i845 SDRAM equipped P-4 1500 costs way more than some serious AMD iron.

If the SiS DDR chipset is up to scratch and as cheap as you say, then that's a step in the right direction. I'll be checking what's happening with Intel prices as soon as I'm back at the office. But I won't be seriously considering the P-4 option as something I could recommend to my customers unless they start offering some real, fair-dinkum value for money. Once I can offer a P-4 2000 for about the same price as an equivalent Athlon XP (a 1600 or 1700), then it will be a salable product. Not before.

When you look at what you get for what you pay, in Intel chips, only the Celeron comes close.

 

[Mercutio]

The benchmarks I've seen for newest P4 are pretty impressive. But intel's dizzying array of chipset choices, coupled with the rogue nature of the current Via chipset and my complete distrust of all things ALi and SiS, means that despite good performance (finally!) there's no way I'd consider one.
That, and like Tannin said: They cost WAY too much. Particularly the Northwoods, which are the first decent performers in the P4 lineup.

 

[Pradeep]

Definitely have to agree with Tony, the prices of decent performing P4's aren't even in the same universe as an equivelently performing AthlonXP, certainly in Australia.

 

[CougTek]

...certainly in Australia.

The price drop is very recent here so maybe you didn't have it yet. I agree that a P4 + SiS645 mobo still costs more on average than an Athlon XP + VIA KT266A combo. The gap isn't as huge as it was though. I like the Intel option more an more as the Northwood CPUs are droping slowly but surely in price.

Here I can have an Athlon XP 1700+ OEM for 220$CDN (270$CDN for the retail version) and a Soltek SL-75DRV2 for 155$CDN. But then I must add a 20$CDN for a decent heatsink because almost every AMD CPU is sold OEM instead of retail (which is a good thing IMO because the retail heatsink isn't as good as the one I put). Total for an AMD platform : 395$CDN. The Pentium 4 1.7GHz cost 340$CDN and it's alway retail here. The MSI 645 Ultra cost 125$CDN to bring the total for an Intel platform to 465$CDN. 70$ difference, not that bad. If Intel manage to sell the Northwood at more or less the same price as the older 0.18µm core, I'll consider it a very serious option for sure.

I must resist...don't...chose...the dark side.

 

[time]

Your problem, Cougtek, is that you're falling into the megahurts trap.

That P4/1700 surely isn't a Northwood, so you're comparing Apples to Oranges. Try comparing an Athlon XP 1700+ with a P4 1900 instead to make a fair comparison (although by the time you add in the effects of the chipset, a P4 2000 would be more appropriate).

Unless of course Quake is your only app ...

 

[Mercutio]

Is anyone else really bothered with the AthlonXP nomenclature? Cyrix got seriously slammed for using PR ratings. Now we have AMD doing the same thing - and making essentially meaningless comparisons - and nobody's bothered by it?

P4 does have something else going for it. Parity, for one. Most of the big vendors have dropped AMD (I think Compaq is still shipping Athlons). Cougtek advertises AthlonXPs, and the people who bring in their electronics chain-store flyers are going to be asking where they can find the Pentiums. So Cougtek either has to launch into a long, complicated explanation of AMD, Athlon XP, megahertz and and how crappy P4 is... or he can just be like everyone else and sell P4s.

 

[Pradeep]

I believe Tony is sellling a steady stream of almost exclusively AMD machines, with Intel available on special order?

 

[Pradeep]

Is anyone else really bothered with the AthlonXP nomenclature? Cyrix got seriously slammed for using PR ratings. Now we have AMD doing the same thing - and making essentially meaningless comparisons - and nobody's bothered by it?

But this time, the AMD performance ratings actually understate the performance of the chips compared to the p4. The XP 2000+ is equal to or faster than a p4 2GHz. It doesn't really bother me, I know what the true chip speed is.

 

[Tannin]

Quite so, Pradeep. We offer everything but we hardly ever get to sell anything except AMD. Last year, we sold:

Duron: 300 or 400 maybe.
Athlon: 100 or so.
Pentium III: six.
Celeron: two.
Cyrix III: one (our sample).
Pentium 4: nil.

This year I expect to sell a higher proportion of Athlons - so far it's up to two Athlons to every three Durons, though it's early days yet - and, who knows, we might finally see a P-4 go out the door.

(Actually, it might be interesting to talk about the stuff that people are actually buying, set out our product list and ponder why people choose this system or that one. But I'll do that in another thread, I think.)

 

[CougTek]

Time,

I know the P4 1.7GHz isn't a Northwood. I was just comparing the similar ratings together. Intel is rumored to launch a 1.8GHz and perhaps even a 1.6GHz version of Northwood in the upcoming weeks. Northwood is made to operate at ~2.4GHz and more, so a 1.8GHz would be very overclocking friendly IMO (the 1.6GHz even more).

You are right that the 0.18µm Pentium 4, even at equal price, aren't all that much tempting. Nonetheless, the price difference is melting like snow under the sun. If a 1.8GHz Northwood appears for more or less the same price as an Athlon XP 1800+ or even as an Athlon XP 2000+, it will be quite a good deal considering that it should be possible to make it work on a FSB of 533MHz (133MHz quad-pumped). With a FSB of 533MHz, the CPU would run at 2.4GHz (1.8 * 133/100), still within the limits of the 0.13µm core.

I won't jump the bridge as long as Northwood won't be as cheap as Athlon XP configs, but if/when it happens, I'm not sure at all.

So far, I've only sold AMD-based systems with the exception of a few Celeron (Tualatin core) machine in the last year. You all know that the wind can change direction rather quickly in the computer market though.

I hope Barton will feature huge improvements over the regular Palomino core or else, it will be Intel's year.

 

[flagreen]

Intel Processors with non-Intel chipsets always leads to BSODs.

 

[CougTek]

Intel Processors with non-Intel chipsets always leads to BSODs.

Shameful bias. That's not what most of the 'Net reviewers say about the SiS645.

 

[flagreen]

Intel Processors with non-Intel chipsets always leads to BSODs.

Shameful bias. That's not what most of the 'Net reviewers say about the SiS645.

Owwww that stings Mon Ami. If the Net reviewers said it. It must be true.
Now I must retreat to lick my wounds and work on that shameful bias of mine.

 

[CougTek]

Bill,

It's not necessarily because some 'Net reviewers say it that it must be true, but guys like the ones at Ace's Hardware usually know their stuff and I trust them for the most part.

Besides, I've built a few Pentium III and Celeron II systems using VIA chipsets and they were fine for their intended purpose. I wouldn't have run servers on them, but for average users, they were stable enough.

SiS latest chipsets (from last summer to nowadays) have a good reputation. So far, I've only used the SiS735 for socket-A platforms and I have no complain about it (except perhaps the compatibility problem with the <1GHz Thunderbirds). I don't see why their chipsets for socket 370 or socket 478 wouldn't be as reliable as their socket-A offerings.

 

[CougTek]

But despite your shameful bias, you are still Mon Ami too so no worries ;-)

 

[flagreen]

Coug,

If it's good enough for Ace's it's good enough for me. My statement was way too general in nature anyway. Sometimes I post before I think which is not wise. 8)

 

[time]

Cougtek,

Northwood is indeed a different kettle of fish, by any measure.

I hope Barton will feature huge improvements over the regular Palomino core or else, it will be Intel's year.

But I'm not sure I understand what aspects of the Athlon you want AMD to improve? Surely not IPC?

 

[CougTek]

If they aren't able to follow Intel in the Megahurtz race, they'll have to improve the IPC indeed.

Things that could be done to improve the IPC on the Athlon :

• Widen the bus to access the L2 Cache : It's currently 64bit wide, one fourth of the Pentium III and probably also ¼ of the Pentium 4's. The larger L1 cache of the Athlon is nice, but if you combine both (L1 + L2), Intel still has the advantage on this field.
  
  SSE2 support : AMD's managers already admitted that SSE2 won over 3D Now! SIMD instructions. Unfortunately, Barton won't feature SSE2 support, it won't happen before ClawHammer arrives.
  
  More pipelines for the FPU unit : Or any other significant improvement to increase the FPU power of the Athlon. It can be done in many ways, I just suggest one.
  
  Larger L2 Cache : It isn't impossible to do. I heard the processor inside the GameCube has several MegaBytes of embedded memory with a 2048bit wide bus (I don't know if it's the access bus, the data bus or both combined).
  
  Faster memory bus : DDR II support. Intel prepares a dual channel DDR SDRAM chipset, AMD has to answer.

There are certainly many other things that can be done to improve the current Athlon core, I suggested a few I had in mind but I'm not a CPU engineer. I know Barton will be the first AMD processor to be made using IBM's SOI process. The result will be less heat dissipation and consequently higher possible clock speeds. Will it be enough without any other significant improvement to compete against Intel's products at this time? I think it won't, but time will tell (not you, it's just an expression).

 

[HellDiver]

Widen the bus to access the L2 Cache

Y. At least doubling it would be nice, but there are certain difficulties associated with that (Umm... Aren't there some associated with everything in this world? ), at least if AMD want to keep it cheap - AMD's only way of increasing its marketshare in the near future.

SSE2 support

N. At least not IMO. At the moment, there are no readily available (to programmers) ways to put SSE2 to good use. Compilers support is scarce and results are still unsatisfactory, manually coding in Assembly is not something the "general public" enjoys doing very much, especially since it ties you to the specific platform. Moreover - unless you're very good at Assembly coding, and quite savvy in SSE2 usage - your Assembly SSE2-based implementation may easily end up being somewhat slower than non-SSE2 code. Finally, SSE2 optimizations are only applicable to a relatively small number of applications, something that slows down its acceptance even more.
Properly implementing (meaning both correctly, and making it fast) SSE2 in silicon is not a trivial exercise, certainly costly in both R&D and manufacturing terms, while the expected near-term benefits are slim at best. So, no, I don't think AMD should be in any hurry with SSE2 - let Intel spill its guts trying to achieve industry's acceptance, then come to pick up the fruits of others labor!

More pipelines for the FPU unit

?. Doubtful. Athlons FPU is more than on par as it is, while FP calculations (again!) are not the predominant type (excluding gaming, science, etc). IMHO FPU is not what's stopping Athlon right now...

Larger L2 Cache

<drooling> Y!!! Certainly 512kb would be very nice, but at the moment - prohibitively expensive. The performance benefits wouldn't justify the cost increase (well, on a 0.13 process they would - for Intel, but remember that AMD must be competetive price/performace-wise, if they are to remain in business).

It isn't impossible to do. I heard the processor inside the GameCube has several MegaBytes of embedded memory with a 2048bit wide bus

Bad (and wrong) example. Theoretically speaking - anything's possible. Have a look at some vector processor supercomputers on the market - they have some multi-kb registers too. So?! Comparing a general-purpose desktop CPU to a highly specialized piece of hardware is plain wrong.
Besides, it's not GameCube you're referring to. And not the CPU, but rather GPU. And not 2048-bit bus. And it's not SRAM you're talking about, although it's on-die.
GameCube sports a rather conservative PPC 750CXe spin-off for the CPU, while ArtX-designed (today : ATI) Flipper GPU sports 3MB of "1T-SRAM", not real SRAM!!! 2MB of that cache are accessed through 384-bit 166MHz bus, and another 1MB through 512-bit 166MHz bus. BitBoys with their elusive XBA only theorize about 1024-bit 150MHz+ eDRAMs... Sony's PS2 GPU (again!) has 4MB of eDRAM accessed through 2560-bit connect running at 150MHz , but there are some questions about whether that connect can really be called a general purpose access bus... Now, "1T-SRAM" is actually not SRAM at all - it's DRAM with tricky refresh timings scheme. About the same goes for embedded DRAMs ("eDRAM"s). They're much better than your regular SDRAM, but have a major limitation - clock speed. Athlon's caches are full-speed, while "1T-SRAM" has literally years to go before it'll be able to touch such operational frequencies.
Large SRAM caches (1-4MB) are indeed possible - high-end server CPUs are coming with those on-board for a long while. But the price premium is... well, a killer.

Faster memory bus : DDR II support.

?. I don't think DDRII is here yet. I'm not even sure it's been standartized yet! (haven't been following news lately - got bored). IMHO It'll be at least another year before DDRII will start showing up, and somewhat longer before it'll become mainstream. By now DDR is just becoming mainstream memory type (no, don't tell me you bought a stick a year ago - what counts is how much DDR vs SDR is being manufactured worldwide) - in part thanks to Intel, BTW - and will remain so for a while. DDR333 support seems more realistic, and I'm talking here about a 166MHz DDR FSBs, people, not just the memory. That might do some good, though at first I'd like to see some benchmarks of an OCed AthlonXP sitting on 166MHz. Links anyone?

Intel prepares a dual channel DDR SDRAM chipset, AMD has to answer.

IIRC, nVidia's "wunderkind" Crush is just that - a dual channeld DDR chipset. Which only goes to show, how you can totally twist and make completely useless a basically good idea. So, while yes, generally speaking more memory access B/W is a good thing, it doesn't always make any good, especially unless your FSB can put it to good use!!!. Don't forget that P4 is sitting in a slightly different bus, with a slightly different throughput capacity. A dual channel DDR chipset for P4 makes sense - it simply makes up for the pair of fatal mistakes : first, the decision to design a CPU that would rely on high memory access B/W, second, the decision to abandon the memory that could deliver that high access B/W.
IMHO, a dual-channel DDR chipset for Athlon would follow the fate of RDRAM and DDR chipsets for P3 - a waste of good money for negligible increase in performance. Go for 166MHz bus and DDR333, AMD!

Will it be enough without any other significant improvement to compete against Intel's products at this time? I think it won't, but time will tell.

I personally think Intel will give AMD a good run for its money - now that they're down to 0.13u process they'll be ramping up clockspeed like there's no tomorrow. Just going SOI is certainly not enough, although its a welcome change.
Generally speaking (i.e. leaving effective IPC aside) - a reliable thermal protection for the CPUs would be an excellent idea. A better physical layout ("heat spreader"?) might be on order too. Those two are persistent problems, especially for novice "enthusiasts". I've had entirely too many burnt-out or chipped off Athlons to take care of for people, all due to wrong HSF installations. IMHO, making the internal thermal diode support a mandatory thing for mobo makers ASAP would be a good move.

 

[Tannin]

If there is one thing that really pisses me off about the AMD chips I sell, it's that there is no 100% reliable thermal overload protection. To hell with putting it on the motherboard! The correct place to mount any safety device is as an integral part of the thing that needs protection. The thermal cutout belongs on the part that it is designed to protect: the CPU itself. Anything else is a cop-out.

Hey, you've got 37 million transistors already. You telling me it's too hard to dedicate a couple dozen to thermal overload protection?

BTW, great post, HD. I'm going to have to read it again to pick up on some of the stuff that didn't soak in the first time.

 

[time]

Awesome post, HellDiver! I dips my lid ...

 

[Santilli]

Guys: The AMD pci bus sucks

There have been a couple different testers that have found at high pci bus rates, the VIA chipsets are about 20% or more slower then the INTEL chipsets.

I think, for that reason alone, that I would look at the new intel chipsets, if, that's a factor in the system setup for the person you are configuring a machine for.

Read, if the person has a desire for a fast raid array.

gs
PS:
Otherwise, and even considering this, I suggest supporting AMD, unless you like paying a grand for a piece of sand from Intel.

gs

 

[HellDiver]

(Thanks to Pradeep for heads up on the fact that Greg has a new habit of using message subj field!)

The AMD pci bus sucks

Uh... NO! Unless you have some evidence to back that statement up, Greg!

VIA has certain timings problems with PCI bus implementations in both its AMD CPU based and Intel CPU based chipsets, that much I know. But AFAIK VIA and AMD haven't announced a merger yet, so lets stick to the facts - VIA PCI bus sucks.

 

[Vlad The Impaler]

This thread reminds me of when I first used to read Storagereview. An intelligent, well reasoned and interesting argument with no jumping up and down and waving of hands. Lovely......

On the subject of the discussion, I too have finally succumbed to AMD. It seems that since the XP came out, all of the stupid little compatibilty issues that we noticed in the past have gone. I am very proud to report that we have yet to have one failure with any AMD kit we have sold thus far. Amazing!

Intel's failure to produce any CPUs for us around November last year forced us into AMD's arms, and I am glad they did. Oh dear, Intel!

Now, does anyone know were I can get 14 Xeon 2Ghz from? So far Intel have produced one (1) prototype Prestonia for us to lend to a client (the British Govt) when we threatened (not seriously!) to go to AMD. I wish we could persuade our customers onto AthlonMP. Having to go direct to your IPI representative to get one CPU is a joke and Intel should be ashamed.

 

[Santilli]

Helldiver: Hard to find a mobo with out the VIA chipsets...

and I try.

I have had both the 750 and 760 series boards, and the
VIA chipsets may well be the cause of the spikes we have noticed for so long, on so many raids.

Just a thought.

I would love to see an ALL AMD mobo, without the via chipsets.

Still, this isn't that big a deal, at least to me.

110 MB/sec is still pretty damn fast.

gs

 

[Pradeep]

I would love to see an ALL AMD mobo, without the via chipsets.

Still, this isn't that big a deal, at least to me.

110 MB/sec is still pretty damn fast.

gs

TYAN Tiger MP, TYAN Tiger MPx, TYAN Thunder, pure AMD, no Via to be seen

 

[Santilli]

check component compatibility with the Tyan boards

Last time I checked, ATTO cards were incompatible with any Tyan board.

GS

 

[HellDiver]

TYAN Tiger MP, TYAN Tiger MPx, TYAN Thunder, pure AMD

Actually, Pradeep, none of the boards you listed has 32/33 PCI bus, so it's not really a direct comparison (though Tiger MP and Thunder K7 can probably be considered as such once you stick in a 32 bit card for testing).

And just remember folks - insofar dualie AMD chipsets (760PM/PMX) have had one poor design decision sticking out : S/B attached to the main PCI bus. Some four years ago this really wasn't a consideration for desktop chipsets, but times change - and so do B/Ws of the devices attached to the PCI. To resort to such a clumsy solution in their very first server/workstation chipset was a bad decision IMHO, although I can understand AMD's reasoning behind it. It's just a shame that they made their first shot in the game with the big boys so... "average", that would be the word! May hurt their reputation in the future.

BTW, in case of MPX, it's even worse, 'cause the 32/33 PCI is tunneled through the S/B onto the 64/66 PCI!

Either way, 760MP/MPX chipsets will hardly demonstrate whether AMD's PCI implementation is any better than VIA's, because what you'll be looking at won't be the PCI implementation, but rather a PCI bus under load of several devices you didn't even think were there (or in case of MPX : a PCI bus limited by an extra logic hop - S/B - and an extra bus tunneling hop - 64/66 PCI). Unless you'll be tinkering with a logic analyzer, that is!

 

[Prof.Wizard]

I think you should have named your thread: "Are you tempted to turn to SiS again?" :mrgrn:

No, seriously. SiS has done a hell of comeback. I'm pleased that more chipset manufacturers compete...

 

[NRG = mc²]

For whoever asked:

8KHA+ 183FSB CAS2 max timings
XP @ 1.4Ghz (Low Multiplier )

http://server24.hypermart.net/dynashome/cpu/mem183.jpg

Should give an idea what the XP will do on 166 FSB :mrgrn:

 

[NRG = mc²]

Whoops... screwed that up. Check the quoted link.

 

[Santilli]

HD: Could you dum down your reply so I can understand what

you are saying?

I'm intrested in the topic, but, you are over my head on this one, and, it's an area that is VERY important to me.


gs

 

[Santilli]

HD:What are the abrevs...yausing

gs

 

[CougTek]

It isn't impossible to do. I heard the processor inside the GameCube has several MegaBytes of embedded memory with a 2048bit wide bus

Bad (and wrong) example. Theoretically speaking - anything's possible. Have a look at some vector processor supercomputers on the market - they have some multi-kb registers too. So?! Comparing a general-purpose desktop CPU to a highly specialized piece of hardware is plain wrong.
Besides, it's not GameCube you're referring to. And not the CPU, but rather GPU. And not 2048-bit bus. And it's not SRAM you're talking about, although it's on-die.
GameCube sports a rather conservative PPC 750CXe spin-off for the CPU, while ArtX-designed (today : ATI) Flipper GPU sports 3MB of "1T-SRAM", not real SRAM!!! 2MB of that cache are accessed through 384-bit 166MHz bus, and another 1MB through 512-bit 166MHz bus. BitBoys with their elusive XBA only theorize about 1024-bit 150MHz+ eDRAMs... Sony's PS2 GPU (again!) has 4MB of eDRAM accessed through 2560-bit connect running at 150MHz , but there are some questions about whether that connect can really be called a general purpose access bus... Now, "1T-SRAM" is actually not SRAM at all - it's DRAM with tricky refresh timings scheme. About the same goes for embedded DRAMs ("eDRAM"s). They're much better than your regular SDRAM, but have a major limitation - clock speed. Athlon's caches are full-speed, while "1T-SRAM" has literally years to go before it'll be able to touch such operational frequencies.

Nice copy and paste exercice. I read that article too, I just didn't think I needed to quote the exact numbers to state my idea. The real width of the bus isn't important. My point was that it could be a lot wider than what it currently is. I confused the CPU with the GPU, but then again, it's of little importance in this case.

And I didn't wrote SRAM, I wrote embedded memory : eDRAM if it's too hard to translate.

Comparing a general-purpose desktop CPU to a highly specialized piece of hardware is plain wrong.

Not when I'm only refering to the cache/eDRAM/whatever. Integrated RAM has two main purposes, no matter if it's a general purpose CPU or a specialized CPU or a GPU : store the data needed by the processor while it's doing it's calculations and feed it as quickly as possible when it needs to store/retrieve data. The two most important parameters to check for the integrated memory is bandwidth and latency. The former as to be as high as possible as the later as to be as short as possible. Sure, you can't just plug (way to speak) any type of RAM on any CPU without some (sometimes heavy) modifications on the CPU's bus interface unit. But it's not an impossible task to accomplish. I don't say it has to be +2Kb wide, I only wrote it's possible. Whether or not the performance boost would worth it, I don't know.

I wasn't awared of the severe clock speed limitations of eDRAM (I never wrote anything about 1T-SRAM) when I wrote my previous post. I still believe that, although it might not make a decent L2 cache replacement like I wrote in my earlier post, it would be a very interesting alternative for a L3-like cache replacement. I'm sure you have more loose time than I have to search on Google about the cost comparison between eDRAM and SRAM, but I think the former would be cheaper. eDRAM should also take less space on the core than SRAM IIRC. Since it's DRAM, it would be logical.

eDRAM and other on-die memory types interest me because of the bandwidth they allow. The theoretical implementation of BitBoys' design had a >10GB/s bandwidth to the embedded memory. Not bad. Feasible or not for a desktop CPU? That's the job of AMD's engineers to figure out. I propose, that's all. Same goes for 1T-SRAM. I never mentioned it, but since you bring it to the table, it might also be an interesting alternative. If the performances/cost/space on-die are superior to existing technologies, that it.

Large SRAM caches (1-4MB) are indeed possible - high-end server CPUs are coming with those on-board for a long while. But the price premium is... well, a killer.

The two main reasons why server CPU are so expensive are :

• a) They are targeted to the industrial market. Who says industrial says $$$.
  
  b) They are made in very limited number compared to the desktop CPU, so the R&D and production costs show much more on the final selling price.

Intel won't be selling its Northwood processors a lot higher than their Willamette despite a x2 Cache increased. The 0.13µm process transition doesn't hurt, but the volume nivels the expense much more IMO. For AMD, making the jump from 256K L2 to maybe 1MB wouldn't make their Athlon twice as expensive. Since the rest of the core is still a lot smaller than the Northwood, they would probably still be able to underprice Intel with such modification.

Anyway, let's agree that the Athlon wouldn't suffer from a wider bus to access its cache and that more bandwidth would be a good thing (as long as the latency would be kept low).

It's probably a bit too late for exotic memory implementations for Barton though, since this processor is schedule for late summer/fall. Oh well, dreaming doesn't hurt.

Intel prepares a dual channel DDR SDRAM chipset, AMD has to answer.

IIRC, nVidia's "wunderkind" Crush is just that - a dual channeld DDR chipset. Which only goes to show, how you can totally twist and make completely useless a basically good idea.

I also wrote that the FSB should be increased too. I never wrote in my previous post that AMD should make a dual-channel DDR chipset without increasing the FSB speed. I proposed DDR II instead of a dual channel DDR bus, but since, according to you, this standard isn't near broad industry adoption, a dual DDR channel configuration paired to a significant increase of the FSB should give good results. NRG's graph seems to proove this (higher FSB = big boost). Current Athlon XP seem to choke in memory intensive benchmarks (and demanding memory applications). How the Athlon architecture would benefit from a big boost from the DRAM memory bandwidth? We'll have to see, but I'm optimistic.

Increasing the FSB and pusing a faster memory configurations is a lot more realistic to hope for Barton than heavy cache modifications IMHO.

Generally speaking (i.e. leaving effective IPC aside) - a reliable thermal protection for the CPUs would be an excellent idea. A better physical layout ("heat spreader"?) might be on order too. Those two are persistent problems, especially for novice "enthusiasts". I've had entirely too many burnt-out or chipped off Athlons to take care of for people, all due to wrong HSF installations. IMHO, making the internal thermal diode support a mandatory thing for mobo makers ASAP would be a good move.

100% agreed. These are two points where AMD needs to improve vs Intel.

 

[Vlad The Impaler]

We just got some Northwood 1.6A and 1.8A P4s in today. They were exactly the same price as the old style Williamette P4s. Same goes for the 0.13 Celeron 1100s......

 

[HellDiver]

Nice copy and paste exercice. I read that article too...

Thanks! Apparently I did better than you...

And I didn't wrote SRAM, I wrote embedded memory : eDRAM if it's too hard to translate.

It is hard to translate, because it ain't so. L1 nad L2 caches on Athlon are nothing more than "embedded memory". Except they are embedded SRAM memory, while here we're talking about embedded DRAM memory. And there is a world of difference!

Comparing a general-purpose desktop CPU to a highly specialized piece of hardware is plain wrong.

Not when I'm only refering to the cache/eDRAM/whatever. Integrated RAM has two main purposes

Purposes - yes. But right about there similarities end. Manufacturing costs, yields, scalability, future growth - all those and others have different weights in those two sectors. Flipper was born and will die at 166MHz. P4 went from 1.4 GHz (or was there a 1.3GHz too?) to 2.2GHz in about a year, with ungodly millions being manufactured in the process.

I wasn't awared of the severe clock speed limitations of eDRAM (I never wrote anything about 1T-SRAM) when I wrote my previous post.

Well, now you are. We all learn something from SR community, don't we?

I still believe that, although it might not make a decent L2 cache replacement like I wrote in my earlier post, it would be a very interesting alternative for a L3-like cache replacement.

Already is. HP PA-8800. 2 x 750kb of L1 (SRAM, naturally) per core (8800 has two on-die) + 32MB of 1T-SRAM L2 (off-die naturally, in 4 x 72Mbit chips). And will be used a lot more in the near future, but not as L1/L2 (except for designs akin to PA-8800 where L1 is huge and fast, allowing for a huge and relatively slow L2).

I'm sure you have more loose time than I have to search on Google about the cost comparison between eDRAM and SRAM

LOL - actually, this week I do - I'm sitting home, sick as a puppy! But being the asshole that I am - I'm not going as much as type the word 'google' in the address bar. I'll tell you off the top of my head that 1T+1C is cheaper than 6T !!!

eDRAM and other on-die memory types interest me because of the bandwidth they allow.

I don't think so. There's absolutely no correlation between B/W and memory type here. Both P3 and Athlon have SRAM for L2, but at identical clock speed P3 has several times more access B/W. There's absolutely nothing in eDRAM as in itself that would facilitate higher access B/W than that of SRAM. NADA.

Intel won't be selling its Northwood processors a lot higher than their Willamette despite a x2 Cache increased.

Well, first of all - we're yet to see about that! Besides, x2 cache increase from 256kb to 512kb is not quite the same as for instance from 1mb to 2mb. Lets not forget, those babies come at 6T per bit, plus all the interconnects. 1MB of L2 can easily cost you >50 millions of tansistors. P4 + 256kb L2 is only 42 millions. AthlonXP + 256kb L2 is only 37.5 millions. You double the die - your yields plumet.

For AMD, making the jump from 256K L2 to maybe 1MB wouldn't make their Athlon twice as expensive.

Could you please explain to silly me why do you believe so?

Besides, it's not just about the amount of cache, it's also about its implementation. PIIIs CuMines with 256kb of L2 were faster than PIIIs with 512kb of L2 - latency and B/W as you mentioned are of paramount importance. Having 4 times the cache amount yet keeping the identical properties is very difficult.

Finally, the amount of cache to pay an arm and a leg for is to be very carefully considered against the typical workload of your CPU. If your typical data locality sucks arse (i.e. you pull 4 bytes from here, then another 2 bytes from there, never the same data etc), the benefits of your large and expensive cache will be severely limited...

Anyway, let's agree that the Athlon wouldn't suffer from a wider bus to access its cache and that more bandwidth would be a good thing (as long as the latency would be kept low).

Definitely!

I also wrote that the FSB should be increased too. I never wrote in my previous post that AMD should make a dual-channel DDR chipset without increasing the FSB speed.

Frankly, I just went back to re-read your post I were replying to and found nothing about the FSB, but then again - maybe I'm missing something here... Besides, my reply wasn't supposed to somehow personally attack your opinion - I simply used your bullet points to express my own opinion, sometimes agreeing with you, sometimes not.

I proposed DDR II instead of a dual channel DDR bus, but since, according to you, this standard isn't near broad industry adoption, a dual DDR channel configuration paired to a significant increase of the FSB should give good results.

To utilize the B/W afforded by a dual channel PC266/PC333 (umm... About 4.2GB/s - 5.4GB/s) you'd have to either double the FSB freq to 532MHz (266MHz DDR) or to go QDR on 133MHz (essentially P4-style bus). That would drive mobo prices up significantly (prolly 6 layers at least, etc. Even the more expensive Intel are still at 100MHz QDR equivalent!) Going 128bit on an FSB of a 32bit CPU sounds very doubtful, not to mention the price increase in this case!

But let's stay realistic here, ok? I personally don't see too much of a reason to double the FSB B/W of Athlon line of CPUs. There's just no use for it yet. On Hammers (i.e. server/WS-grade CPUs) - yeah, sure, that's why folks are thinking about on-die memory controllers and inter-chip HT. But on desktop Athlons? Naah... "Not economically viable!"

NRG's graph seems to proove this (higher FSB = big boost).

I'll bet you're a McDonalds lover! Synthetics are nice, but let's see the beef first!!!

How the Athlon architecture would benefit from a big boost from the DRAM memory bandwidth? We'll have to see, but I'm optimistic.

Me too. I'm also realistic, so to speak, so I'm willing to settle for a 166MHz DDR FSB.

 

[HellDiver]

HD: Could you dum down your reply so I can understand what you are saying?

I'm intrested in the topic, but, you are over my head on this one, and, it's an area that is VERY important to me.

Well, lately I'm saying awfully lots , could you be more specific? You know I'll swim across the sea for ya, Greg - just name it...

And what abbrevs got you confused? AFAIK = As Far As I Know, IIRC = If I Recall Correctly, IMHO = In My Humble Opinion, B/W = Bandwidth, S/B = SouthBridge (of a chipset), N/B = Northbridge...

 

[Santilli]

Thanks HD: That helps...

Now to go over it again...

gs