Was Prescott really that bad?

[superstition]

Yeah, I mean by that logic no CPU has been made "completely in house" by anyone in that time frame.

In any time frame.

Check out the history of the Fairchild F8. Guess who Fairchild brought in to make that a reality? (Hint: It was a person who got to see what was happening at another firm that was developing a next-gen CPU.)

If mid-1970s CPUs needed poaching of ideas from other teams then good luck with anything after that.

 

[DrMrLordX]

Smithfield was just two Prescott cores glued together (quite literally, actually).

Oh I know. It's just that technically it was a different CPU, even if it was the same uarch and the same process.

I still remember Tom's trying to sell people in the idea that a Pentium D + water cooling was an acceptable alternative to an X2.

 

[JoeRambo]

I still remember Tom's trying to sell people in the idea that a Pentium D + water cooling was an acceptable alternative to an X2.

Define viable tho. As already mentioned in this thread, P4D 805 was dirty cheap ( ~$100 ) and we ran quite a few machines with them oced. Could not beat its price/perf when cheapest AMD dual core was hundreds.

But they chew electricity at insane rates and were heating office space heavily.

 

[Hi-Fi Man]

Pentium D 805 was a real bargain for those that wanted dual core on the cheap however, most people knew it was a hack job of a dual core. MCM chips aren't bad but Intel's quick implementation meant all core to core traffic had to pass through the FSB to the northbridge then back through the FSB to the other core which meant insane latency. I think Core 2 Quad fixed this if I remember correctly by allowing direct die to die communication.

 

[VirtualLarry]

MCM chips aren't bad but Intel's quick implementation meant all core to core traffic had to pass through the FSB to the northbridge then back through the FSB to the other core which meant insane latency. I think Core 2 Quad fixed this if I remember correctly by allowing direct die to die communication.

Nah, Core2Quad was still an MCM'ed CPU, and while the two dies were wired together, they still used FSB to talk to each other. I don't know if Intel optimized it. I think, that in a multi-CPU FSB-linked cluster, there were already ways for each CPU to talk to each other over the FSB directly, rather than data being routed from CPU 1 to northbridge over FSB, and then back over the northbridge to CPU 2.

 

[Phynaz]

\

Absolutely.Prescott was every bit as bad as Netburst.

Why is this conversation even happening in 2016?

Huh? Prescott was Netburst.

86 posts so far on the day you registered. I think that has to be some kind of record.

 

[Arachnotronic]

Prescott was bad. Let's face it, the strategy there was "People can't recognize performance, but they do recognize big Mhz numbers as a crude proxy for performance". I think in this case the old axiom about no one losing money underestimating the intelligence of the public didn't work. Or maybe it would have if they could have just scaled those things up forever.

Imagine AMD didn't exist during the prescott years. And now remember what has happened as they have lacked competitive products in the SandyBridge+ era.

The shift to mobile/power efficient computing in the data center would have ensured the death of Netburst, with or without AMD's influence. Remember that Pentium-M was developed in parallel to the Netburst line and was ultimately called upon to save the day.

 

[SlowSpyder]

Early P4's (Willamette + Rambus) were bad. Northwood was half decent, Prescott was cool engineering in that Intel made the pipeline sooooo long and yet was able to keep IPC pretty close. But Athlons were just more efficient and better performers in absolute terms with overclocking involved. I know it isn't technically Prescott, but by the time they were making 65nm dual cores I remember the P4 wasn't half bad.

 

[Torn Mind]

Those who could not overclock, be it for highly sensitive calculations or buying a "regular" prebuilt from an OEM, got less performance for the same clockspeed with Prescotts instead of Northwood. And AMD was faster for nearly everything but video editing but since most people don't know about computers and how to determine what is "faster" than the other, people just bought on guts based on spec sheet numbers.

 

[nismotigerwvu]

Early P4's (Willamette + Rambus) were bad. Northwood was half decent, Prescott was cool engineering in that Intel made the pipeline sooooo long and yet was able to keep IPC pretty close. But Athlons were just more efficient and better performers in absolute terms with overclocking involved. I know it isn't technically Prescott, but by the time they were making 65nm dual cores I remember the P4 wasn't half bad.

Oh dear lord, I haven't thought about RDRAM in a long time. Twice the expense, and half the speed/capacity. So many systems had their useful life shortened to that stuff. While just about everything else remained serviceable, they would run into a RAM wall that just wasn't cost effective to eliminate.

 

[Hi-Fi Man]

Oh dear lord, I haven't thought about RDRAM in a long time. Twice the expense, and half the speed/capacity. So many systems had their useful life shortened to that stuff. While just about everything else remained serviceable, they would run into a RAM wall that just wasn't cost effective to eliminate.

The initial benifit of RDRAM was that it allowed the P4 to breath a bit more and stay ahead of the P3 as SDRAM didn't have enough bandwidth to properly feed a P4. However, as soon as DDR capable chipsets were released RDRAM lost its purpose.

 

[superstition]

The initial benifit of RDRAM was that it allowed the P4 to breath a bit more and stay ahead of the P3 as SDRAM didn't have enough bandwidth to properly feed a P4. However, as soon as DDR capable chipsets were released RDRAM lost its purpose.

This.

Really high latency is what cost RDRAM its performance. It was able to almost be on par with DDR 1 for performance but the performance-per-dollar was awful, especially with 1066 speed RDRAM.

RDRAM, as he said, was a big improvement over SDRAM for the P4.

(What we really needed was a dual Tejas board with eight sticks of RDRAM and a Parhelia.)

 

[BigDaveX]

To be fair, a Northwood-B P4 with PC1066 RDRAM did make for a very potent (if expensive) combo, and was the top-performing setup for its day. What really killed it was that Rambus were pouring all their efforts into XDR - which didn't get commercially used in any products until the PS3 arrived in 2006 - and they didn't bother releasing any RDRAM speed grades faster than PC1066, which would have needed a tri-channel chipset in order to provide the bandwidth that Northwood-C needed.

Instead, Intel pushed JEDEC into ratifying DDR-400 as a standard. Which earned them about six months of Northwood-C smacking the Athlon XP silly, and then just short of three years of their line-up getting annihilated by the Athlon 64, which used the bandwidth of a dual DDR-400 setup far more effectively than any Netburst CPU could have ever dreamed of.

 

[daxzy]

Ok, as you've seen in my posts here I'm not a fan of Netburst, but some of these postings here are just revisionist history.

The initial benifit of RDRAM was that it allowed the P4 to breath a bit more and stay ahead of the P3 as SDRAM didn't have enough bandwidth to properly feed a P4. However, as soon as DDR capable chipsets were released RDRAM lost its purpose.

RDRAM was super expensive from the start (P3/i820 chipset). But by the time Northwood-A's were around, 256MB PC800 RDRAM was at price parity with DDR-266. Then mysteriously, there were RDRAM shortages and price increases once DDR was being pushed.

But lets not forget that the DRAM OEM's were conspiring against Rambus (I actually got a paycheck this year for $120 from that class action lawsuit). Basically, they were upcharging RDRAM and using those profits to pad DDR (which was selling at a loss early on).

The initial benifit of RDRAM was that it allowed the P4 to breath a bit more and stay ahead of the P3 as SDRAM didn't have enough bandwidth to properly feed a P4. However, as soon as DDR capable chipsets were released RDRAM lost its purpose.

P4 was designed to be fed by a high bandwidth memory solution. Initially, DDR-266 wasn't even that great on P4 boards. 850E with RDRAM@1066 had about a 5-10% performance lead over 845E with DDR@333. Unfortunately for Rambus, even if the stars aligned and DRAM OEM's didn't screw them over, the low clock scaling of Prescott would've made RDRAM@1333 and above useless anyways.

I still remember Tom's trying to sell people in the idea that a Pentium D + water cooling was an acceptable alternative to an X2.

Pretty sure a $40 air cooler + ($133) PD-805 was a viable solution for lots of non-gamers if you were interested in a modest overclock. If you did content creation, it was actually quite the bargain, being more energy efficient (over time) than any single core solutions. Cheapest X2 was the X2-3800+ and it ran > $400 due to supply shortages/retailer gouging.

 

[Insert_Nickname]

850E with RDRAM@1066 had about a 5-10% performance lead over 845E with DDR@333.

You're perfectly right, but once mainstream dual channel became available with the 875/865 RDRAM lost all purpose. This was also the reason the (Xeon) E7205 chipset was popular with enthusiasts, as it was available earlier then those, used cheaper memory, and performed on par with the 850E.

 

[Hi-Fi Man]

Ok, as you've seen in my posts here I'm not a fan of Netburst, but some of these postings here are just revisionist history.



RDRAM was super expensive from the start (P3/i820 chipset). But by the time Northwood-A's were around, 256MB PC800 RDRAM was at price parity with DDR-266. Then mysteriously, there were RDRAM shortages and price increases once DDR was being pushed.

But lets not forget that the DRAM OEM's were conspiring against Rambus (I actually got a paycheck this year for $120 from that class action lawsuit). Basically, they were upcharging RDRAM and using those profits to pad DDR (which was selling at a loss early on).



P4 was designed to be fed by a high bandwidth memory solution. Initially, DDR-266 wasn't even that great on P4 boards. 850E with RDRAM@1066 had about a 5-10% performance lead over 845E with DDR@333. Unfortunately for Rambus, even if the stars aligned and DRAM OEM's didn't screw them over, the low clock scaling of Prescott would've made RDRAM@1333 and above useless anyways.



Pretty sure a $40 air cooler + ($133) PD-805 was a viable solution for lots of non-gamers if you were interested in a modest overclock. If you did content creation, it was actually quite the bargain, being more energy efficient (over time) than any single core solutions. Cheapest X2 was the X2-3800+ and it ran > $400 due to supply shortages/retailer gouging.

Revisionist? How so?

DDR made RDRAM irrelevant (as you said thanks in part to OEMs) and as I stated earlier, P4 needed a lot of bandwidth to perform at peak (bandwidth that DDR was able to provide even if RDRAM was a bit faster).

 

[BigDaveX]

P4 was designed to be fed by a high bandwidth memory solution. Initially, DDR-266 wasn't even that great on P4 boards. 850E with RDRAM@1066 had about a 5-10% performance lead over 845E with DDR@333. Unfortunately for Rambus, even if the stars aligned and DRAM OEM's didn't screw them over, the low clock scaling of Prescott would've made RDRAM@1333 and above useless anyways.

Any hypothetical PC1333 RDRAM would still have been too slow for Northwood-C or Prescott; it'd only have supplied enough bandwidth to supply a chip with a 667MHz FSB, in which case a dual-channel DDR333 chipset would have done the job just as well. They'd have needed PC1600 RDRAM to provide the 6.4GB/s of bandwidth the chip needed, and either the design didn't scale up that far or Rambus were too focused on XDR.

 

[superstition]

I hope people focusing on bandwidth are factoring RDRAM's high latency into it.

 

[Hi-Fi Man]

I hope people focusing on bandwidth are factoring RDRAM's high latency into it.

Another reason why DDR made RDRAM irrelevant.

 

[superstition]

Another reason why DDR made RDRAM irrelevant.

I think that's the primary one.

If it had had the latency of DDR with its higher bandwidth it would have outperformed DDR instead of mostly tied it.

People talk about 1066 RDRAM but from what I recall 800 was vastly more common and 1066 was vastly more expensive, especially in the highest capacity.

 

[Hi-Fi Man]

I think that's the primary one.

If it had had the latency of DDR with its higher bandwidth it would have outperformed DDR instead of mostly tied it.

People talk about 1066 RDRAM but from what I recall 800 was vastly more common and 1066 was vastly more expensive, especially in the highest capacity.

I've never even used RDRAM 1066MT/s. When I had my 478 board (can't remember the chipset but it did have AGP Pro) I only used RDRAM 800MT/s.

 

[daxzy]

You're perfectly right, but once mainstream dual channel became available with the 875/865 RDRAM lost all purpose. This was also the reason the (Xeon) E7205 chipset was popular with enthusiasts, as it was available earlier then those, used cheaper memory, and performed on par with the 850E.

Again, the reason SDRAM was cheaper than RDRAM was because DRAM OEMs' purposely made it that way. There was a collusion between the big makers (Samsung, Elpida, Hynix, Infineon, Toshiba) to lower increase costs for RDRAM to push SDRAM. From a manufacturing standpoint, there was little reason for a RIMM to be more expensive than a DIMM one, even factoring that whopping 1% royalty that Rambus charged.

DDR made RDRAM irrelevant (as you said thanks in part to OEMs) and as I stated earlier, P4 needed a lot of bandwidth to perform at peak (bandwidth that DDR was able to provide even if RDRAM was a bit faster).

Because RDRAM had lower latency in high bandwidth scenarios (at least on the P3 platform, which RDRAM was terrible on, I might add). This is comparing single channel RD-PC800 to SD-PC133. Anand thinks on the next page that DDR-SDRAM would close or be better than RD-PC800, but that's just a hypothesis.

http://www.dewassoc.com/performance/memory/rdram_v_sdram.htm
http://www.anandtech.com/show/551/4

This can be shown in real world results as well. Intel's i875p (flagship 800 FSB P4 platform) barely edges out the i850E on benchmarks, when using the same 533 FSB CPU's. This is despite the enormous 6.4 GB/s (DC-DDR400) vs 4.2 GB/s (DC-PC1066) bandwidth increase.

http://www.anandtech.com/show/1094/10

I've never even used RDRAM 1066MT/s. When I had my 478 board (can't remember the chipset but it did have AGP Pro) I only used RDRAM 800MT/s.

Rambus yields were actually quite good by the P4 timeframe. Pretty much any PC800-45 (or -40, I forgot, it's been awhile) could overclock to PC1066 without issue.

 

[dark zero]

Prescott wasn't that good as expecting... Their performance depends a lot on their clock speed and they weren't as efficient as expecting.

Now a measly ARM A53 could reach the same performance with lesser wattage and less clock speed (and smaller node)

 

[IntelUser2000]

This can be shown in real world results as well. Intel's i875p (flagship 800 FSB P4 platform) barely edges out the i850E on benchmarks, when using the same 533 FSB CPU's. This is despite the enormous 6.4 GB/s (DC-DDR400) vs 4.2 GB/s (DC-PC1066) bandwidth increase.

http://www.anandtech.com/show/1094/10

You are missing the point. CPUs with FSBs needed memory that matched its frequency/bandwidth. Memory bandwidth going over FSBs were essentially useless. 800MHz FSB CPUs were quite a bit faster(Actually, its a remarkable engineering feat on the 865/875 "Springdale" chipsets to beat it by few % despite it having same FSB and bandwidth. I've calculated in CPU benchmarks you needed 33% increase in bandwidth for 5-6% gains)That's the point. RDRAM stopped advancing after a point while DDR flew past them. Then again, DDR memory had 5-6 manufacturers going for it while RDRAM had just Rambus.

 

[Hi-Fi Man]

You are missing the point. CPUs with FSBs needed memory that matched its frequency/bandwidth. Memory bandwidth going over FSBs were essentially useless. 800MHz FSB CPUs were quite a bit faster(Actually, its a remarkable engineering feat on the 865/875 "Springdale" chipsets to beat it by few % despite it having same FSB and bandwidth. I've calculated in CPU benchmarks you needed 33% increase in bandwidth for 5-6% gains)That's the point. RDRAM stopped advancing after a point while DDR flew past them. Then again, DDR memory had 5-6 manufacturers going for it while RDRAM had just Rambus.

Technically Rambus wasn't a manufacturer because they were fabless. Samsung and another company were the manufacturers I believe.