I'd say Socket A - 650 MHz at introduction, finishing off at 2200 MHz with the 3200+.

I'm thinking LGA 775: started with slow Prescott P4's and ended with significantly faster Core 2 Quad 9xxx. If you're looking for single core power, the Wolfdale E8600 is faster than the Core 2 Quads, and still much faster than single threaded performance of Prescotts.

If we're talking about upgradeability, you should include slockets. They are non-intervening upgrades.

Socket 3: 20 MHz (i486 SX) to 133MHz (5x86)
Socket 7: 75 Mhz (iP54C) to 570 (AMD K6/2+)
Slot 1: 233 (iPII) to 1,4 Ghz (iPIII-s)
Socket 370: 333 (iCeleron) to 1,4 Ghz (iPIII-s)

Absolute or relative increase?
Socket 3 is probably up there from Intel 486DX-25 (1989) to AMD Am5x86-P75 (1995) (Or P100 if it actually released).

RandomStranger wrote on 2022-05-09, 18:16:

Absolute or relative increase?
Socket 3 is probably up there from Intel 486DX-25 (1989) to AMD Am5x86-P75 (1995) (Or P100 if it actually released).

I'm not sure at all and too lazy to look it up, but we'd have to take the fastest CPU available at introduction, not the slowest.
Wasn't the DX-33 threre from the start?

So, Socket 3 would be 33 to 133 or even 160, if rare on paper.

7, apart from 4 and 5, started off with what? P90, 100? Up to K6-III 570?
That might be quite a bigger leap in actual computing power as the 486 mainly stayed the same while the K6-III is quite another beast than the humble P5?

I suppose Socket 1 is a strong contender if you do count later upgrades and CPU interposers, from 486SX-16 to AMD 5x86-133, so more than 8x the difference just in clock, not to mention the FPU and bus speed increase.

Without interposers you have Socket 3, from 486SX-20 (though more commonly they start at 25MHz) to the rare AMD 5x86 150, so 7.5x the difference. I don't think there's a Soc3 mobo that supports 16MHz CPUs.

Cuttoon wrote on 2022-05-09, 18:28:

I'm not sure at all and too lazy to look it up, but we'd have to take the fastest CPU available at introduction, not the slowest […]
Show full quote

RandomStranger wrote on 2022-05-09, 18:16:

Absolute or relative increase?
Socket 3 is probably up there from Intel 486DX-25 (1989) to AMD Am5x86-P75 (1995) (Or P100 if it actually released).

I'm not sure at all and too lazy to look it up, but we'd have to take the fastest CPU available at introduction, not the slowest.
Wasn't the DX-33 threre from the start?

So, Socket 3 would be 33 to 133 or even 160, if rare on paper.

7, apart from 4 and 5, started off with what? P90, 100? Up to K6-III 570?
That might be quite a bigger leap in actual computing power as the 486 mainly stayed the same while the K6-III is quite another beast than the humble P5?

I think the DX-33 wasn't available at the launch.

Also, did the K6-III 570 actually exist? Maybe I'm just uneducated, but wasn't the 550 the top? And even that is so rare I don't know it actually had commercial release or just some leaked engineering samples.

brian105 wrote on 2022-05-09, 18:11:

I'm thinking LGA 775: started with slow Prescott P4's and ended with significantly faster Core 2 Quad 9xxx. If you're looking for single core power, the Wolfdale E8600 is faster than the Core 2 Quads, and still much faster than single threaded performance of Prescotts.

That made me think. What about AMD? Their sockets from AM2 to AM3+. Does the + version becomes a new socket even if it maintains backward compatibility? My AM2+ board back then could go from Athlon 64 3800+ (AM2 from 2006) to Phenom II X6 1065T (AM3 2010) and other boards that could handle the power draw 1100T. I started that board with a Athlon X2 5050e then upgraded to Phenom II X3 710, then X6 1055T.

But OP prefers to ignore multi-core.

Socket 1 or Socket 7 seems probable for vintage sockets, if interposers are allowed. If we include modern sockets, AM4 is probably the most upgradable socket ever.

Bristol Ridge "dual"core bulldozer 3.4ghz turbo freq APUs -> 5950x 16core/32thread 4.9GHz boost freq monster.
GHz increase isn't insane but 8x (16x? AMD got in hot water for describing Bulldozer "cores" as real cores, so divide any Bulldozer by 2 to be legally accurate) as many cores on a dramatically better architecture is insane for performance.

Ignoring multicore seems unreasonable, especially for a socket as modern as AM4.

AM2/AM2+/AM3 are worth examining too, but I don't think AM3+ is worth looking at as a result of how miserable Bulldozer is, unless we're working with purely frequency numbers.

775 performance uplifts were a lot for a modern socket, especially on the singlethread, but for multithread they're not as insane as AM4.
Passmark:
Celeron 420 -> Q9650
471 ST 235 MT -> 1292 ST 2411 MT
~2.75x singlethread, ~10.3x multithread

A6-9500E -> 5950x
1478 ST 1841 MT -> 3298 ST 46198 MT (not a typo)
~2.2x singlethread, ~25.1x multithread

Socket 3 was released for new DX4 CPUs, so it has very short upgrade path. It's not correct to include pin-compatible CPUs which are older than socket itself.
Now Slot 1 is a strong contender, but only if third-party mods for BIOS and adapters are considered. Same goes for LGA775, because it had pre-Core 2 and post-Core 2 revision, latter shuffled some pins.

flupke11 wrote:

Socket 370: 333 (iCeleron) to 1,4 Ghz (iPIII-s)

Actually 300 Mhz

The Serpent Rider wrote on 2022-05-10, 05:21:

Same goes for LGA775, because it had pre-Core 2 and post-Core 2 revision, latter shuffled some pins.

Any source of that info? I don't remember such change and there were mobos which would support core duos/quads with just a bios update. Not to mention late 865 chipset boards with lga775 socket that supported practically everything you threw at them up to 1066 fsb. Voltage regulation requirements changed significantly, however.
On the other hand, LGA1151 came in two versions to enforce fake incompatibility between skylake 6/7 and coffee lake 8/9 series.

Anyway, socket 370 could be a strong contender but it came in different, incompatible revisions. Therefore, looking purely at pre-multicore sockets and disregarding chipsets, I think that the choice is between either socket 7 or socket A.

http://www.intel.com/Assets/PDF/datasheet/306382.pdf
http://download.intel.com/design/processor/da … shts/318732.pdf

https://en.wikipedia.org/wiki/Socket_604 also lasted a very long time - from 2002 to 2008 (Dunnington), and compatible with Socket 603 (early P4 Xeon).

The Serpent Rider wrote on 2022-05-10, 05:21:

Socket 3 was released for new DX4 CPUs, so it has very short upgrade path. It's not correct to include pin-compatible CPUs which […]
Show full quote

Socket 3 was released for new DX4 CPUs, so it has very short upgrade path. It's not correct to include pin-compatible CPUs which are older than socket itself.
Now Slot 1 is a strong contender, but only if third-party mods for BIOS and adapters are considered. Same goes for LGA775, because it had pre-Core 2 and post-Core 2 revision, latter shuffled some pins.

flupke11 wrote:

Socket 370: 333 (iCeleron) to 1,4 Ghz (iPIII-s)

Actually 300 Mhz

Thanks for the update! Another to hunt down 😀

The Serpent Rider wrote on 2022-05-10, 07:12:

http://www.intel.com/Assets/PDF/datasheet/306382.pdf
http://download.intel.com/design/processor/da … shts/318732.pdf

Damn, should have thought of these myself 😀

Anyway to summarize changes for curious:
There are a lot of changes among "FCxx" signals which are "signals that are available for compatibility with other processors."
If something was used on P4 and is not used on Core series (marked as FCxx) I do not consider it a change. But reverse is considered a change in my book.
This leaves us with these, Land name - P4 function - Core function:

1AL3 - VSS - VRDSEL: "This input should be left as a no connect in order for the processor to boot. The processor will not boot on legacy platforms where this land is connected to VSS. "
2AM5, AM7, AN7 - FCx - VID6,7,VID_SELECT: New voltage VIDs
3B13 - FC19 - COMP8: "COMP8 must be terminated to VSS on the system board using precision resistors."
4D23 - FC9 - VCCPLL: "VCCPLL provides isolated power for internal processor FSB PLLs."
5G2 - FC1 - COMP2: "COMP2 must be terminated to VSS on the system board using precision resistors."
6G5 - FC7 - PECI: "PECI is a proprietary one-wire bus interface."
7H2 - FC6 - GTLREF1: P4 had only one GTLREF, Core series need 2.
8L2 - TESTHI13 - SLP#: "When asserted in Extended Stop Grant or Stop Grant state, causes the processor to enter the Sleep state" Optional
9P1 - TESTHI11 - DPSLP#: "When asserted on the platform, causes the processor to transition from the Sleep State to the Deep Sleep state." Optional
10R1 - FC2 - COMP3: "COMP3 must be terminated to VSS on the system board using precision resistors."
11T2 - FC4 - DPRSTP#: "When asserted on the platform, causes the processor to transition from the Deep Sleep State to the Deeper Sleep state." Optional
12V2 - LL_ID0 - RESERVED: "All RESERVED lands must remain unconnected."
13Y3 - FC17 - PSI#: "Processor Power Status Indicator Signal"

D23, V2 and B12,G2,R1 can be considered significant and are most likely source of incompatibilies in early revisions of motherboards, which support core series in later revisions.
L2 and P1 could be ignored as long as sleep states were not used and VIDs could likely be worked around.

All in all pretty interesting. Surprising that it did not get the same treatment as LGA1151.

Would a POD-83 count as a socket 3 path? i.e. 486sx-16 (if it' possible to find a board that supports 16MHz) up to POD-83 MHz

Just considering the socket alone is a bit tricky to me.

Take socket A for instance, a board from the the KT133 chipset era would not let you go above Palomino.
Similar happens to Slot 1, with an unmodified board and/or special slotket, you cannot aim for top Pentium III (Tualatin 1.4). Not even all 370 boards would help you in that regard (Socket 370 boards/chipset with Tualatin support is more like an exception than the rule).
Socket 775 had a long life, but depending on the chipset, you could also be limited to PIV or low end Dual Core.

a board from the the KT133 chipset era would not let you go above Palomino.

Nope, some boards will refuse to work with any Athlon XP and some will work with Barton. But you're stuck with 133 Mhz bus at best (overclocking) and need a workaround for CPU multiplier limit.

Well, we're far from approaching the issue scientifically here 😀.

According the OP's title, I'd just take the physical socket into account. As an interposer does fit the physical socket, why not take that into the equation. If a contraption exists to run Itanium on socket 5, why not 😉.

Which parameter do we use to define the "biggest increase in CPU horsepower"? Speedcheck? Prime95?

My vote still goes to the physical socket 7.

S370 had a good run.
From our beloved Celeron 300A all the way up to Tualatin 1400S.
An impressive jump both frequency and performance wise.

S7 beats it in frequency increase, but I'm not sure it will in performance.

Socket A also had a nice life.
Didn't scale that much when it comes to frequency but performance? well maybe.

S775 on the other hand has no chance in frequency increase.
I mean it was introduced at 2.6 GHz ish and topped out at 3.8 GHz or so.
Performance on the other hand did a far more impressive jump.
From a lowly single core Pentium 4 all the way to the mighty Core 2 Extreme QX variants.

My money is on S775.