Question Intel had a 7 GHz CPU years ago

[igor_kavinski]

https://www.reddit.com/r/hardware/comments/wir8ay

213 mm2 die size, 150 Watts, 50 pipeline stages all for one core at >7 GHz.

I'm getting a strong feeling of lust for this CPU. 150W ain't that much. Imagine what this CPU could have been if it had been developed further, with modern branch predictors and huge caches.

 

[511]

Those were Otellini's fault. Remember Haswell was 2013 and Broadwell was 2014, while BK came in at 2013. They are all Otellini designs. So straight after success with Core 2 and "eliminating" competition, they went back to complacency. And they claimed they needed to tame advancements on Haswell to get that lower power but I don't believe that fully.

You are right so looks like after barett everyone was screwing around Barrett has few flaws like Itanium though.

 

[DavidC1]

You are right so looks like after barett everyone was screwing around Barrett has few flaws like Itanium though.

Itanium came out under Barrett, but development started in 1989.

 

[Thunder 57]

But igor is saying that they should have stuck to the dead end architecture and that would have somehow resulted in success for Intel.

It's entertaining but it's a dead wrong conclusion.

And they did this because they were relying on N-1 process for making GMCH, and N-2 for their IOH and moving to integrated memory controller would mean moving GMCH to the CPU die and less able to fill fabs. Internally they were considering moving to integrated memory controller in the late 1990's. That's when they should have started considering fabbing for outsiders as well. Their failures are entirely on them. Pentium III Coppermine could have had IMC and P2P interconnect.

LoL. they were winning on cache without an IMC and they knew it. I very much doubt they were working on am IMC in the late 90's and allowed AMD to catch up that way. K7 and even K8's L2 was piss poor.

 

[Thunder 57]

Itanium came out under Barrett, but development started in 1989.

VLIW has never worked. I don't know why the Russians are trying with Elbrus.

 

[igor_kavinski]

Itanium came out under Barrett, but development started in 1989.

Anything HP seems to be cursed anyway. Remember the memristor and their "Machine"? I guess still under development...

Intel has this stupid culture of internal competition between the design teams. Not like AMD where they collaboratively design stuff. They build on what the other team has done, rather than find faults in it and then try to go their own way.

 

[igor_kavinski]

VLIW has never worked. I don't know why the Russians are trying with Elbrus.

Vodka must've this effect of looking at very long instructions and seeing eternal bliss.

 

[DavidC1]

LoL. they were winning on cache without an IMC and they knew it. I very much doubt they were working on am IMC in the late 90's and allowed AMD to catch up that way. K7 and even K8's L2 was piss poor.

There were many that wanted to but filling the fabs were priority, hence never materialized. They were greedy, not stupid.

Intel has this stupid culture of internal competition between the design teams. Not like AMD where they collaboratively design stuff. They build on what the other team has done, rather than find faults in it and then try to go their own way.

It can turn out to be an advantage. People can easily fall into a collective mindset and struggle to get out of it. That's an issue too. But yea I agree with you on the big picture.

 

[Doug S]

Itanium came out under Barrett, but development started in 1989.

The research began inside HP (as PA-RISC 3.0 / PA-WW) in 1989, but Intel didn't get involved until five years later.

 

[LightningDust]

The research began inside HP (as PA-RISC 3.0 / PA-WW) in 1989, but Intel didn't get involved until five years later.

Indeed. Intel canceled its own 64b RISC project to hop on to HP's proposal, and infected it with silly stuff like x86 compatibility (which was, prior to roughly 2000, intended to include operating system compatibility.) The bones of the Itanium architecture were essentially complete long before Intel joined up.

 

[gdansk]

Indeed. Intel canceled its own 64b RISC project to hop on to HP's proposal, and infected it with silly stuff like x86 compatibility (which was, prior to roughly 2000, intended to include operating system compatibility.) The bones of the Itanium architecture were essentially complete long before Intel joined up.

Oh, so the requirement for a wundercompiler didn't come from Intel? That's interesting.

 

[LightningDust]

Oh, so the requirement for a wundercompiler didn't come from Intel? That's interesting.

The wundercompiler thing is massively overblown. HP's compilers mostly did fine, especially with PGO.

 

[Quintessa]

Indeed. Intel canceled its own 64b RISC project to hop on to HP's proposal, ...

On paper, IA-64 could've been revolutionary, but x86 baggage and compiler reliance killed it. The "7 GHz" you're recalling was more PR than production

 

[LightningDust]

On paper, IA-64 could've been revolutionary, but x86 baggage and compiler reliance killed it. The "7 GHz" you're recalling was more PR than production

I assure you, I recall no 7GHz anything.

IPF was "killed" by nothing so much as a secular decline in the RISC/UNIX industry once x86 got 64b addressing and competent RAS features - as well as internal feuds within (and between) HP and Intel.

 

[igor_kavinski]

as well as internal feuds within (and between) HP and Intel.

Always lovely to hear about those if you can recount a few or ALL of them please

 

[igor_kavinski]

I assure you, I recall no 7GHz anything.

So you would never wish a chimera child of Itanium and P4?

 

[LightningDust]

Always lovely to hear about those if you can recount a few or ALL of them please

I don't know how to tell the stories in a way that renders them coherent. I will simply say this - Intel and HP had fundamentally different goals with IPF. HP's was to provide a direct evolution of the HP 9000 family (with a port of HP-UX and maybe, if the userbase was very lucky, MPE/iX) and a compelling enough story to retain the Compaq base (Alpha + MIPS.) They were successful at the former and somewhat successful at the latter. Intel wanted a high-end counterpart to Xeon that would let them grab share from entrenched RISC players due to their superior economies of scale. This relationship was complicated by a number of things. This is all from memory, augmented a bit by the Oracle documents, so bear with me.....

• HP was dependent on Intel for lithography and uncore designs but Intel was dependent on the HP Fort Collins design center (until... 2005? That sounds right...) for actual Itanium microarchitecture work. The Itanium2 core, used in Itanium generations from 2002 to 2012, was an HP design. Intel had their own high-end Itanium design - Tanglewood - being worked on out of the ex-DEC/Compaq Hudson design center, but HP looked at it with extreme suspicion. Intel then killed it and replaced it with Tukwila, which had - shall we say - a troubled development process, despite not actually including all that much new. (It was essentially four I2 cores with QPI, and if it shipped in 2007-2008 as initially planned, at the projected 2GHz, it would have been excellent - a recurring theme with Itanium gens.)
• By 2006, it was clear that Itanium was not going to be what Intel wanted it to be, while HP was quite happy with it in most regards. Non-HP IPF vendors had niche success - mainly SGI, Bull, and the Japanese domestic vendors - but their combined shipments were still far less than HP alone, and most of them had no true ties to Itanium since they were largely shipping Windows Server, Linux, or proprietary systems. (The exception here is NEC, which used HP-UX and stayed committed to Itanium well into the 2010s.) Intel started making unhappy "we want out" noises around this time. HP responded with "well, uh, maybe IPF would be doing better in the market if you guys tried shipping one on time? Just a thought..." Bad feelings all around. Around this point, most of the other OEMs started getting nervous and eyeing the exits. Only HP, NEC, and Inspur ever shipped an internally-designed Itanium system after 2010. Beckton was so much better than Tukwila that the Itanium value proposition mostly disappeared, unless one was tied to HP operating systems.
• Bad feelings got worse as HP started doing internal prototyping of a large Opteron system (Octane.) Intel did not like this very much.
• Intel briefly flirted with canceling Tukwila (they had already canned its x86 counterpart, Whitefield, which would have included Core2 cores and QPI) and exiting Itanium entirely - resulting in a deal where HP would directly fund Intel for future generations. HP was in a very bad situation at this point, since they did not have their own lithography, or ownership of the IPF intellectual property, or their own processor design center since the transfer of the Fort Collins design center a couple of years earlier.
• RISC/UNIX took a collective beating after the 2008 recession, even Power, which at the time was doing great technically. HP started seriously looking for a way out of their situation at this point. HP-UX got experimentally ported to x86, which was never released. The plan, for a little while, was that UX would be ported to x86, Itanium would get new generations through the late 2010s in a compatible socket with Xeon, and you would be able to use both Xeon and Itanium cell boards in a single Superdome chassis.
  • ... then Oracle bought Sun, and promptly de-supported Itanium - supposedly based on private statements from Intel that they wanted to exit IPF, but in reality, a pretty clear move to boost SPARC, which had been doing poorly for some time. This cratered Itanium's sales, even after HP eventually won in court, and HP gutted the roadmap. No more UX port, no more mixed-processor Superdomes, no more extended Itanium roadmap. Xeon support in the Superdome eventually materialized in the Superdome X, while other parts of the projected "Kinetic" mission-critical Xeon lineup - notably the HydraLynx scalable blades - never saw the light of day. Nonstop, being a beloved cash cow for HP, got ported to x86, VMS got parted out, and UX got taken out behind the barn. (It's still supported, though only for another four months. Lots of customer angst about the impending EOL.)

 

[poke01]

I don't know how to tell the stories in a way that renders them coherent. I will simply say this - Intel and HP had fundamentally different goals with IPF. HP's was to provide a direct evolution of the HP 9000 family (with a port of HP-UX and maybe, if the userbase was very lucky, MPE/iX) and a compelling enough story to retain the Compaq base (Alpha + MIPS.) They were successful at the former and somewhat successful at the latter. Intel wanted a high-end counterpart to Xeon that would let them grab share from entrenched RISC players due to their superior economies of scale. This relationship was complicated by a number of things. This is all from memory, augmented a bit by the Oracle documents, so bear with me.....

• HP was dependent on Intel for lithography and uncore designs but Intel was dependent on the HP Fort Collins design center (until... 2005? That sounds right...) for actual Itanium microarchitecture work. The Itanium2 core, used in Itanium generations from 2002 to 2012, was an HP design. Intel had their own high-end Itanium design - Tanglewood - being worked on out of the ex-DEC/Compaq Hudson design center, but HP looked at it with extreme suspicion. Intel then killed it and replaced it with Tukwila, which had - shall we say - a troubled development process, despite not actually including all that much new. (It was essentially four I2 cores with QPI, and if it shipped in 2007-2008 as initially planned, at the projected 2GHz, it would have been excellent - a recurring theme with Itanium gens.)
• By 2006, it was clear that Itanium was not going to be what Intel wanted it to be, while HP was quite happy with it in most regards. Non-HP IPF vendors had niche success - mainly SGI, Bull, and the Japanese domestic vendors - but their combined shipments were still far less than HP alone, and most of them had no true ties to Itanium since they were largely shipping Windows Server, Linux, or proprietary systems. (The exception here is NEC, which used HP-UX and stayed committed to Itanium well into the 2010s.) Intel started making unhappy "we want out" noises around this time. HP responded with "well, uh, maybe IPF would be doing better in the market if you guys tried shipping one on time? Just a thought..." Bad feelings all around. Around this point, most of the other OEMs started getting nervous and eyeing the exits. Only HP, NEC, and Inspur ever shipped an internally-designed Itanium system after 2010. Beckton was so much better than Tukwila that the Itanium value proposition mostly disappeared, unless one was tied to HP operating systems.
• Bad feelings got worse as HP started doing internal prototyping of a large Opteron system (Octane.) Intel did not like this very much.
• Intel briefly flirted with canceling Tukwila (they had already canned its x86 counterpart, Whitefield, which would have included Core2 cores and QPI) and exiting Itanium entirely - resulting in a deal where HP would directly fund Intel for future generations. HP was in a very bad situation at this point, since they did not have their own lithography, or ownership of the IPF intellectual property, or their own processor design center since the transfer of the Fort Collins design center a couple of years earlier.
• RISC/UNIX took a collective beating after the 2008 recession, even Power, which at the time was doing great technically. HP started seriously looking for a way out of their situation at this point. HP-UX got experimentally ported to x86, which was never released. The plan, for a little while, was that UX would be ported to x86, Itanium would get new generations through the late 2010s in a compatible socket with Xeon, and you would be able to use both Xeon and Itanium cell boards in a single Superdome chassis.
  • ... then Oracle bought Sun, and promptly de-supported Itanium - supposedly based on private statements from Intel that they wanted to exit IPF, but in reality, a pretty clear move to boost SPARC, which had been doing poorly for some time. This cratered Itanium's sales, even after HP eventually won in court, and HP gutted the roadmap. No more UX port, no more mixed-processor Superdomes, no more extended Itanium roadmap. Xeon support in the Superdome eventually materialized in the Superdome X, while other parts of the projected "Kinetic" mission-critical Xeon lineup - notably the HydraLynx scalable blades - never saw the light of day. Nonstop, being a beloved cash cow for HP, got ported to x86, VMS got parted out, and UX got taken out behind the barn. (It's still supported, though only for another four months. Lots of customer angst about the impending EOL.)

this right here is gold, thanks for sharing.

 

[Quintessa]

IPF was "killed" by nothing so much as a secular decline in the RISC/UNIX industry once x86 got 64b addressing and competent RAS features

Exactly, once AMD64 hit the scene, IA-64 had no reason to exist outside niche HP-UX boxes. x86-64 steamrolled IA-64's niche.

 

[LightningDust]

Exactly, once AMD64 hit the scene, IA-64 had no reason to exist outside niche HP-UX boxes. x86-64 steamrolled IA-64's niche.

And SPARC's, and Power's.

RISC/UNIX performance (especially Power, but occasionally Itanium) was often higher than server x86, but with Opteron and later EM64T eating the entry level, the economies of scale weren't there to stay cost-competitive.

 

[igor_kavinski]

People being cheap (especially those with lots to spend, you know, the multimillionaires and billionaires) ruins the good stuff. Pentium 4 died undeservedly because it was just easier and cheaper to have a lower clocked architecture. I keep hearing about some cockroach called ARM that's threatening the vast x86 library.

It's always annoying that all these new CPU architectures can't be bothered to include a teeny tiny hardware based emulator or chip to keep old software libraries alive, much like the PS3 launch model had a PS2 chip in it. If the Itanium had an x86 hardware co-processor, it probably may have fared better, assuming Intel and HP had brought its price down to Xeon levels.

 

[igor_kavinski]

System manufacturer P5K-E vs System manufacturer System Product Name - Geekbench

P4 Presler scores one ST win against Thuban!



See? My fascination with P4 isn't unwarranted.

 

[DavidC1]

It's always annoying that all these new CPU architectures can't be bothered to include a teeny tiny hardware based emulator or chip to keep old software libraries alive, much like the PS3 launch model had a PS2 chip in it. If the Itanium had an x86 hardware co-processor, it probably may have fared better, assuming Intel and HP had brought its price down to Xeon levels.

What?!

Itanium and Itanium 2 did have one. It spent significant amount of die area and transistors to do so. The 1GHz Itanium 2 could outperform 1.5GHz Pentium 4 significantly but the x86 hardware unit performed like a 300MHz Pentium II - an absolute disaster and a token support.

With the 0.13u shrink of Itanium 2, they abandoned it and moved to dynamic translation like they do on Apple and ARM and it improved performance for a 1.5GHz Itanium 2 equal to a 1.5GHz Pentium 4.

 

[igor_kavinski]

With the 0.13u shrink of Itanium 2, they abandoned it and moved to dynamic translation like they do on Apple and ARM and it improved performance for a 1.5GHz Itanium 2 equal to a 1.5GHz Pentium 4.

That's kinda impressive for dynamic translation, in a world where they could've iterated quickly and moved beyond 5 GHz for the Itanium in a few years.

 

[DavidC1]

That's kinda impressive for dynamic translation, in a world where they could've iterated quickly and moved beyond 5 GHz for the Itanium in a few years.

Why is that impressive? Apple's dynamic translator does even better at 70-80%. The Itanium's translator is on the low side closer to 50%.

 

[Doug S]

Why is that impressive? Apple's dynamic translator does even better at 70-80%. The Itanium's translator is on the low side closer to 50%.

Rosetta 2 is a static translator. It supports dynamic translation for stuff the static translator can't handle, but there isn't much where that's true.