Blavius wrote on 2022-12-03, 10:04:

Do dual pentium 4 motherboards exist?

No and I assume Intel did this to increase the sales on Xeons

Probably something related to the numbers of pins. Socket 603/604 had more pins for a reason. And well yeah, market segmentation is a thing - lowest end workstation/server motherboards had single Socket 478, just like regular desktop.

I have a Pentium 4 rig running a SL8PZ and a dual Xeon rig with a couple of SL8P3's. What's interesting about this is that both chips have the same CPUID/stepping (0F4Ah/R0). So it's basically the same CPU in both cases but just in a different packaging (LGA775 vs. S604)

Errius wrote on 2022-12-03, 13:01:

I have a Pentium 4 rig running a SL8PZ and a dual Xeon rig with a couple of SL8P3's. What's interesting about this is that both chips have the same CPUID/stepping (0F4Ah/R0). So it's basically the same CPU in both cases but just in a different packaging (LGA775 vs. S604)

That's very interesting, that supports the idea of forced market segmentation. Perhaps there are some pins required to enable SMP that are not present on the pentium 4 package, so there is no way to put more than one by design.

Earlier this was not the case, as wikipedia (sorry) states: "When first introduced, Slot 1 Pentium IIs were intended to replace the Pentium and Pentium Pro processors in the home, desktop, and low-end symmetric multiprocessing (SMP) markets. The Pentium II Xeon, which was aimed at multiprocessor workstations and servers, was largely similar to the ordinary Pentium II, being based on the same P6 Deschutes core, differing by offering the choice of L2 cache capacity of 1024 or 2048 KB besides 512 KB". Seems like for pentium 4 they decides 'low-end symmetric multiprocessing' also belonged under the multiprocessor umbrella.

It's just marketing.

Since P54 (Socket 5), Intel has designed CPUs with native SMP. They have however chosen whether to enable that functionality on a per CPU and per chipset basis, depending on what they wanted to sell to whom. You see that more clearly when you look at architecture code names rather than commercial names.

P4 CPUs were all based on the Netburst architecture, which most definitely supported SMP. Intel chose to completely separate the marketing of the SMP-disabled CPUs (called "Pentium 4") from the SMP-enabled CPUs (called "Xeon") despite the chips being otherwise virtually identical - as Errius points out, the CPUIDs and steppings were identical between Willamette vs Foster, Northwood vs Prestonia and Prescott vs Nocona. Prestonia did have HT support where Northwood did not, but this was not different silicon, just functionality enabled or not.
Intel did get a bit more creative in the high end, with Xeon MP (supporting 4-way SMP) having more cache by default, and there also being bigger-cache Xeon DP Gallatin CPUs - which again got re-used as non-SMP Pentium 4 Extreme Edition.

The restriction of P4 to non-SMP vs Xeon for SMP was for purely commercial reasons, on par with things like multiplier locks, disabled cache and technically unnecessary socket changes. So if you want to follow Intel's marketing, "P4" did not support SMP, but if you look from a technical point of view, Netburst most definitely did and only lacked it where intentionally disabled.

As multi-core became more common, this distinction was even more artificially moved from SMP vs non-SMP to SMP-in-single-socket vs SMP-over-multiple-sockets, and multi-socket designs were relegated to more niche high-end applications, as for all concerned multiple cores on a single die were cheaper, easier and generally preferable to the same number of cores on separate sockets.

By looks of the netburst Xeons it looks like they slapped an P4 on top on an interposer. Have saved some Xeons after we scrapped some HP Proliant DL380 G3 servers and im curious if its possible to desolder/remove the CPU core from the s603/4 pcb.

I wonder what was the retail price difference between these otherwise identical CPUs

I wonder if you could maybe take a 603/604 cpu and put a bga 478 socket on it? Then put whatever pentium 4 you had laying around on it?

Thanks for the detailed writeup dion, very interesting to read.

Sphere478 wrote on 2022-12-04, 04:25:

I wonder if you could maybe take a 603/604 cpu and put a bga 478 socket on it? Then put whatever pentium 4 you had laying around on it?

Heheh 😀 I guess you could, if the socket doesn't check for the missing pins. However, unless all extra pins are dummies, I doubt you'll be able to build a dual P4 this way so it's kinda pointless as Xeons (at least these ones) are cheap.

dionb wrote on 2022-12-03, 14:35:

Prestonia did have HT support where Northwood did not, but this was not different silicon, just functionality enabled or not.

Didn´t late Northwoods support HT?

Yes

Blavius wrote on 2022-12-05, 18:19:

Thanks for the detailed writeup dion, very interesting to read.

Sphere478 wrote on 2022-12-04, 04:25:

I wonder if you could maybe take a 603/604 cpu and put a bga 478 socket on it? Then put whatever pentium 4 you had laying around on it?

Heheh 😀 I guess you could, if the socket doesn't check for the missing pins. However, unless all extra pins are dummies, I doubt you'll be able to build a dual P4 this way so it's kinda pointless as Xeons (at least these ones) are cheap.

Dual extreme editions!

I can only think about one later "consumer oriented" multi socket system.
Intel SkullTrail in 2008.
But it's very high end.
And mostly like a server board (Dual socket 771 Core2Extreme with FBDIMM required)

I really want a 604 to 478 interposer now 🤣

Someone do this 🤣

Sphere478 wrote on 2022-12-05, 20:58:

I really want a 604 to 478 interposer now 🤣

Someone do this 🤣

Aren't you the one doing that? 🤣

🤫

But seriously, does anyone have a bga socket and a old xeon laying around that they can try?

Blavius wrote on 2022-12-03, 10:04:

-Do dual pentium 4 motherboards exist? If so, do you have experience with them? If no, why do you think that's the case?

Motherboards not.

But the Pentium D is special:
Smithfield based Pentium D's use 90µm process and have one package with one die with 2 cores that share the FSB that also goes to the motherboard.
Presler based Pentium D's use 65 µm pcoress and have one package with two dies that share fhe FSB that also goes to the motherboard.

code name, process, cores, threads, remark
Pentium 4, Prescott, 90µm, 1, 2 (hyperthreading)
Xeon, Nocona (no XD Bit) + Cranford + Irwindale (bigger L2), 90 µm, 1, 2 (hyperthreading)
Pentium D, Smithfield, 90µm, 2, 2 (2 cores on a single die; hyperthreading disabled)
Xeon, Paxville, 90 µm, 2, 4 (2 cores on a single die; hyperthreading)
Pentium 4, Cedar Mill, 65 µm, 1, 2 (hyperthreading)
Pentium D, Presler, 65 µm, 2, 2 (2 dies on a single package; hyperthreading disabled)
Xeon, Paxville MP, 65 µm, 2, 2 (2 dies on a single package; hyperthreading)