Was Prescott really that bad?

[el etro]

Cedar Mil was real nice(not just because the 65nm process), but it lost its qualities in 3.6Ghz+ Pentium D forms. More or less what happens nowdays with the Bulldozers.
All other Pentium 4s were bad.

 

[HiroThreading]

Yeah, Prescott was awful. The 31 stage pipeline meant that it had worse IPC than the Northwood chips it was replacing due to a hefty branch misprediction penalty. I suppose Intel engineers were anticipating to clock to speeds near or above 5GHz, but that never came to fruition. It gets points for hitting a stock speed of 3.8GHz on a 90nm process though.

Overclocking now is much easier no FSB to deal with limiting the overclock. Back then the Prescott did not overclock better than they do now, also many more Hz with my cheap motherboard GA Z170 HD3, i5 6600k 4.5GHz on air, that is great.

I'm going to have to disagree. I think overclocking with the FSB was way easier than it is now.

 

[2blzd]

Prescott 3.2ghz was the first chip I ever bought to build a system. I was convinced the 2mb l3 cache and 925-E chipset would blow everything away...


It was good on paper....

 

[pantsaregood]

Yeah, Prescott was awful. The 31 stage pipeline meant that it had worse IPC than the Northwood chips it was replacing due to a hefty branch misprediction penalty. I suppose Intel engineers were anticipating to clock to speeds near or above 5GHz, but that never came to fruition. It gets points for hitting a stock speed of 3.8GHz on a 90nm process though.



I'm going to have to disagree. I think overclocking with the FSB was way easier than it is now.

this is what I'm getting at. Prescott utterly failed to compete with K8, but it did bring the clock scalability that Intel was after. Getting a Prescott (or Cedar Mill) Pentium 4 to 4.5 GHz is not at all uncommon. Someone mentioned that they are curious of what Tejas could've done - perhaps it would've made breaking 5 GHz somewhat common. Intel's strategy of "let's push clock speeds to ridiculous levels" could've worked out if thermals hadn't caught them off guard.

There was also some difference in Netburst vs. K8 and the current state of Intel vs. Bulldozer: Intel was clearly beaten once K8 was releases, but performance was at least competitive - an Athlon 64 3400+ almost always beat a 3.4 GHz Netburst CPU, but it never doubled its performance. The situation now is different - it isn't uncommon to see an FX9590 sharing company with an i3-6100. In the case of Netburst, an extra 500 MHz per-SKU would've likely taken K8 to a draw, as K8 had a much more difficult time ramping up clocks. Bulldozer, on the other hand, seems to hit a wall around 5 GHz, and even without said wall it would still need to clock >7 GHz just to consistently match something like a 6700K.

Also, please remember that in 2003-2005, most Prescott overclocks were thermally limited. In my experience, they were interesting CPUs in that if you could throw enough voltage at them and cool them, they would scale.

I also recall the 2.8 GHz Pentium D with a 533 MHz FSB being a steal - it would almost always hit 667 MHz FSB for a 3.5 GHz clock, and making it to 800 MHz for a 4.2 GHz clock wasn't uncommon.

That said, Pentium D had considerably more issues. It was effectively two dies taped together.

 

[SPBHM]

Overclocking now is much easier no FSB to deal with limiting the overclock. Back then the Prescott did not overclock better than they do now, also many more Hz with my cheap motherboard GA Z170 HD3, i5 6600k 4.5GHz on air, that is great.

i5 6600K and z170 is not really that cheap, you are paying the Intel OC tax, mid 2000s you could overclock cheaper stuff very easily and extract a ton of performance, and back then cheap and expensive CPUs used to have the same amount of cores, so you could get some really interesting performance out of $100 CPUs (and lower), beating CPUs with over 2x the price was really easy most of the time with Celerons, basically all chipsets allowed for some OC, not all motherboards exposed decent options, but many did.

FSB overclocking could be a pain with PCI and memory clocks, also Chipset stability, but, if you had a CPU with lower default FSB than the maximum supported by the platform it used to be pretty painless, also at some point most chipsets had a PCI clock lock and allowed a lot of freedom with that,
Celeron D normally had an FSB 133 while the platform was made for FSB 200

 

[IntelUser2000]

I suppose Intel engineers were anticipating to clock to speeds near or above 5GHz, but that never came to fruition. It gets points for hitting a stock speed of 3.8GHz on a 90nm process though.

They were apparently aiming way higher. Early rumors indicated 5.6GHz parts with quartely 500MHz increases. Even the project guys must have had little idea what could be done to fall that short.

I don't believe we'll ever see CPUs beyond 5GHz , even with overclocking. That's a physics limitation, not incompetence. No stable overclock on air has done much beyond 5GHz and that's a telling sign. Across manufacturers, ISAs, process.

That is unless you can cool a 1000W CPU. Perhaps we'll see it break the 10GHz barrier, but realistically after 5GHz the thermals/power use skyrocket.

 

[krumme]

I don't think, that NexGen had the K7 in preparation. More likely, their engineers, together with AMD's original ones (29K, K5), and those acquired from DEC (Alpha), did the K7 from ground up (1995 - 1999 sounds about right).

I think the point is also k7 was made using external buy up. If zen is good then its the first successfull inhouse ground up design in their history...

 

[Denithor]

Yes, Prescott really was that bad. Couldn't keep up with the X2 chips out at the same time, then the entire Netburst architecture got simply destroyed when Core 2 launched. Literally, destroyed. A given C2D chip running half the clock speed of a Netburst chip would do more work. At far lower power consumption. So, yeah.

I don't believe we'll ever see CPUs beyond 5GHz , even with overclocking. That's a physics limitation, not incompetence. No stable overclock on air has done much beyond 5GHz and that's a telling sign. Across manufacturers, ISAs, process.

That is unless you can cool a 1000W CPU. Perhaps we'll see it break the 10GHz barrier, but realistically after 5GHz the thermals/power use skyrocket.

Well, that target keeps going up. Back in the Nehalem days, chips would typically see drastic power increases going above ~3.5GHz. SB/IB would do 4GHz pretty easy, spike in power above that point. Now we see 4.5GHz easily, pushing toward 5GHz. Given enough time and process improvements, I think clock speeds will continue climbing, albeit not as fast as during the 90's - mid 00's.

 

[Hi-Fi Man]

I used to have a Pentium 4 540 with an Asus P5WD2 Premium. The motherboard was great but the CPU was just terrible, Prescott ran hot but it was actually an easier CPU to cool then my i7 4790K! I eventually put a G0 stepping 65nm Pentium D in that board which I was able to clock to 5GHz on regular air; I was really only limited by the northbridge!

My socket 939 MSI K8N Neo2 Platinum board (nForce 3 Ultra!) and Athlon 64 X2 4400+ combo was great and performed far better.

 

[Atari2600]

[QUOTE="pantsaregood, post: 38610426, member: 288765"
Was Prescott really that bad?
[/QUOTE]


No.

It was worse than that.

 

[dave_the_nerd]

Just going to leave these here:

http://www.anandtech.com/show/1610
http://www.anandtech.com/show/1650
http://www.anandtech.com/show/2045

 

[daxzy]

P-M was clearly better than P-4 while using way less power. It was very competitive with AMD as well but I wouldn't say it "blew away everything". The P-M had a somewhat weaker FPU than it's competition (from AMD and Intel). It was Core 2 Duo that corrected this while completely taking the performance crown back.

http://www.anandtech.com/show/1610/7

The other problem was the price/performance ratio wasn't particularly great. Here is a list of prices of various CPU's from when AMD did their P-M Desktop experiment:

http://www.anandtech.com/show/1610/3

Like I said previously, due to internal corporate political strife, Bianas/Dothan/Yonah wasn't allowed to be competitive in the desktop.

1. Dothan die size was 87 mm^2 (with 2 MB cache). Prescott/1MB was 112 mm^2; Prescott/2MB was 135 mm^2. From a fab perspective, they were actually cheaper to make than Prescott's.
2. Dothan TDP was 21W. Prescott was anywhere from 89W to 115W. If Dothan was given more TDP headroom (e.g. in a desktop environment), it could clock MUCH higher. Tomshardware has clocks it at 2.6 GHz with the stock dinky CT-479 cooler. There are many reviews you can look up where 2.5 GHz+ was not uncommon with the stock CT-479 cooler. The only real limiting factor was the FSB.

3. Gets worse at Yonah generation. Yonah die size was 90 mm^2. Pressler (65nm Pentium-D) was 162 mm^2. You could literally make 2 Yonah's for 1 Pressler. Not even going to talk about TDP and potential overclocking headroom.

Dothan was a pretty solid chip, but blowing everything else away might be a bit of an overstatement. http://www.anandtech.com/show/1610

Sorry, I meant for games. It was definitely weak for general workstation tasks/FPU intensive ones. However, Yonah corrected most of this by adding better SSE support. With a 25W TDP @ 2 GHz, it could trade blows with a X2-3800 (2.0 GHz/89W TDP).

http://www.anandtech.com/show/1900/14

But what about the bigger picture? What does our most recent look at the performance of Intel's Core Duo tell us about future Intel desktop performance? We continue to see that the Core Duo can offer, clock for clock, overall performance identical to that of AMD's Athlon 64 X2 - without the use of an on-die memory controller. "

 

[bbhaag]

I think the point is also k7 was made using external buy up. If zen is good then its the first successfull inhouse ground up design in their history...

Is that really true? I'm being honest. I don't follow cpu tech that closely especially newer post 2002 designs.

 

[whm1974]

Is that really true? I'm being honest. I don't follow cpu tech that closely especially newer post 2002 designs.

I doubt that's true.

 

[sirmo]

I think the point is also k7 was made using external buy up. If zen is good then its the first successfull inhouse ground up design in their history...

No that was K6.. K5 was a failure, K6 was NexGen.. K7 design was done at AMD (and it borrowed a lot of ideas from the Alpha) and it was led by Dirk Meyer who came from DEC. K8 (Hammer) was also done at AMD..

 

[whm1974]

No that was K6.. K5 was a failure, K6 was NexGen.. K7 design was done at AMD (and it borrowed a lot of ideas from the Alpha) and it was led by Dirk Meyer who came from DEC. K8 (Hammer) was also done at AMD..

The K7 and K8 CPUs certainly give Intel major headaches at the time. And reportedly Intel even sold the first Celerons at a loss to keep people from buying K6-2 processors.

 

[sirmo]

The K7 and K8 CPUs certainly give Intel major headaches at the time. And reportedly Intel even sold the first Celerons at a loss to keep people from buying K6-2 processors.

Aye those were the years.. the megahertz wars. I had a K6-2, Athlon Thunderbird, and later the first Opteron. All great products.

 

[whm1974]

Aye those were the years.. the megahertz wars. I had a K6-2, Athlon Thunderbird, and later the first Opteron. All great products.

Yeah the first system I built was an Socket A Duron later upgraded to the Thunderbird Athlon.

 

[krumme]

No that was K6.. K5 was a failure, K6 was NexGen.. K7 design was done at AMD (and it borrowed a lot of ideas from the Alpha) and it was led by Dirk Meyer who came from DEC. K8 (Hammer) was also done at AMD..

Ehh. Wasnt that what i was writing?
Neither k6 nor k7 was internal projects only.
K5 was bad. Prior processors was copies.
K8 was an evolvement from k7.
BD that was inhouse was a failure
=
If zen is good its first time in history a successfull 100% inhouse made x86 cpu is good.
30 years...

 

[BigDaveX]

I'm not sure I subscribe to the idea that Intel's 90nm was such a dud. You only have to look at the Dothan Pentium M's to see what was possible.

Dothan was a good performer - very good for the time and power envelope, actually - but its TDP was no better than the older Banias model, and there were actually a few areas such as standby power usage and overall power leakage where it was actually noticeably worse than Banias. It's just that, well, Pentium M was such a well-engineered chip that Dothan bumped up said stats from "negligible" to "very slightly less negligible."

Intel's 90nm wasn't such a failure that they couldn't produce a decent product at all, but it was almost certainly the worst of Intel's main manufacturing processes.

Ehh. Wasnt that what i was writing?
Neither k6 nor k7 was internal projects only.

With K6 they purchased the design from NexGen and tweaked it to better suit their manufacturing processes and make it Socket 7 compatible, that much is true. However, K7 was almost all their own work, aside from the EV6 bus, which was originated by DEC.

In fact, in some ways K7 was actually an even more impressive achievement than the K8 family. With K8, AMD had an advantage in that Intel's main competitor was based on a much older design in the case of Gallatin, and horribly misdesigned in the case of Prescott. However, Intel didn't really make any major errors with the Pentium III's design; AMD just made a CPU that was better in virtually every department, with the only major weakness being worse cache performance.

 

[Insert_Nickname]

Dothan was a good performer - very good for the time and power envelope, actually - but its TDP was no better than the older Banias model, and there were actually a few areas such as standby power usage and overall power leakage where it was actually noticeably worse than Banias. It's just that, well, Pentium M was such a well-engineered chip that Dothan bumped up said stats from "negligible" to "very slightly less negligible."

Intel's 90nm wasn't such a failure that they couldn't produce a decent product at all, but it was almost certainly the worst of Intel's main manufacturing processes.

I seem to remember the biggest problem with power consumption was when Dothan was used with the newer i910/i915 chipset, which included an FSB bump to 533MHz and a new, more power hungry graphics core (GMA900). When used with the older 855 this wasn't an issue, also Dothan did have twice the L2 cache.

With regards to TDP, I'm scratching my head a bit. 400MHz FSB Dothans had a 21W TDP while 400MHz FSB Banias had a 24.5W TDP...?

 

[krumme]

With K6 they purchased the design from NexGen and tweaked it to better suit their manufacturing processes and make it Socket 7 compatible, that much is true. However, K7 was almost all their own work, aside from the EV6 bus, which was originated by DEC.

In fact, in some ways K7 was actually an even more impressive achievement than the K8 family. With K8, AMD had an advantage in that Intel's main competitor was based on a much older design in the case of Gallatin, and horribly misdesigned in the case of Prescott. However, Intel didn't really make any major errors with the Pentium III's design; AMD just made a CPU that was better in virtually every department, with the only major weakness being worse cache performance.

Its semantics but when you bring in an entire team that builds a cpu on the ev6 bus its in my book not remotely a 100% inhouse project.

Thats not to say its bad. On the contrary i think its excellent to buy persons, teams and ip to make your cpu. There is in my world not an inherent quality in doing all the stuff yourself. I would tend to go the other way.

But the purpose is to show it takes decades not years to build an internal org. That can make a huge compettitive x86 core from scratch.

We have some ip from rambus in zen mem controller but thats it as i can tell?

 

[krumme]

And btw look at Prescott die size, perf and efficiency vs core 2 in AT review and its evident its a hopeless idiotic product. No other way to tell it. It was worse than worse as some say

 

[sirmo]

Ehh. Wasnt that what i was writing?
Neither k6 nor k7 was internal projects only.
K5 was bad. Prior processors was copies.
K8 was an evolvement from k7.
BD that was inhouse was a failure
=
If zen is good its first time in history a successfull 100% inhouse made x86 cpu is good.
30 years...

K7 was an inhouse project. Not sure what you mean.. They didn't buy a firm with a K7 design. They designed K7 in house. What am I missing?

 

[krumme]

K7 was an inhouse project. Not sure what you mean.. They didn't buy a firm with a K7 design. They designed K7 in house. What am I missing?

I didnt say they bough a firm with K7 design. What a strawman.
K7 is build using technology from Dec and the alpha processors with Dirk Meyers as lead engineer.