Jon Masters Leaving NUVIA, Returning To Red Hat

[zyxxel]

So why did Intel fail to conquer the mobile space then? They burned quite a lot of money trying to bring x86 into phones, but they failed miserably. Of course the ISA is not 100% alone to blame for this, but x86 implementations need to deal with highly complex old tech which could hurt their competetiveness with ARM implementations. Jim Keller saw some value in reducing the complexities but I wonder if he had enough time at Intel to fix this (e.g. to invent a x20 ISA which deprecates old cruft within x86).

When I bought my first Atom motherboard, it came with a nice cooler and fan. For the chip-set. Intel decided to mate a 2.6W dual-core processor with a 13W Intel P4 chipset. Lots of reviewers just assumed the HSF was for the processor. In short - Intel have a problem with their management, i.e. with actually putting full focus on doing low-power x86 solutions. It isn't the instruction set that blocks Intel x86 processors from competing real hard in the low-power segment - it's the company focus that just isn't there. Right or wrong is a completely different debate - maybe it just isn't meaningful for Intel to have a significant market presence in the really low-power segments.

[nokipaike]

However, honestly, it would be nice if 2-3 other competitive CPU manufacturers would emerge in addition to the usual ones on the market.
maybe all with their unique and avant-garde characteristics .. I know I know, I like to dream.

[PerformanceExpert]

Because Intel was not / is not used to design for the mobile space. x86 CPUs has traditionally a completely different SoC layout with accelerators and whatnot.
Again. Nothing to do with the ISA. x86 design methodology for SoC on PCBs was a PITA. Very large memory interfaces, IO-pins, accelerators, SPI, I2C etc, etc- Shitloads of pins, southbridge, external phy's, etc, etc. How many ARM embedded CPUs support 40-128 PCIe I/O lanes that are still "battery"? None.

A server/desktop CPU shrunk to Embedded is usually a bad fit.
Esp. if the parent company thinks it can just "shrink stuff" and off we go. It has nothing to do with the ISA.
Just as ARM were bad fits for servers and desktop, x86 was a bad fit for embedded. Does not mean that ARM ISA can not do servers, or vice versa.

Intel didn't shrink a server chip and tried to fit it into a mobile phone. They created a new microarchitecture called Atom. You might want to look it up sometime so you sound less stupid when you say things like this.

[PerformanceExpert]

Yes. Find one vendor graph. Claim something and be completely wrong about it while you're at it.
Performance is not the same as power efficiency. Performance is not linear with power.
The graph shows IDLE-normalized consumption. So again. What the hell are you claiming?

You're claming that a ARM MOBILE SoC tailor made for a use has lower idle-normalized power than an generic desktop CPU?
And then somehow, this relates to absolute performance?

POWER EFFICIENCY IS NOT THE SAME AS PERFORMANCE.

Better power efficiency implies better performance.

The graph clearly shows an Arm core beating x86 mobile cores on raw performance. But what is really interesting is that it also uses a fraction of the power. It is 4-5 times more power efficient while still beating x86 on performance. So clearly x86 implementations cannot compete even when using the same 7nm process (or Intels 10nm process).

Servers are power limited, for example Graviton 2 uses less than 1.5W per core. Arm's higher power efficiency means an Arm server can either achieve higher performance at the same power or the same performance while using less power. So better power efficiency implies better performance. Got it now?

[bridgman]

I love how one man's personal preference to return to his former employer sets off all kinds of irrelveant theories and debates. Instead of trying to read between the lines and connect the dots that aren't there, how about just taking the guy at his word and not seeing his decision as some sort of industry microcosm?

That's what happens when you have bad coffee at the office... you lose good people and the world goes crazy trying to understand what happened.

[milkylainen]

Funny then how this graph shows how a 4.5W core slaughters high-end x86 cores using 4-5 times as much power. How do you explain that exactly? Similarly, Graviton 2 shows that at 110W you can beat every Intel server in existence and compete with EPYC 2 using less than half the power and a third of the silicon area.

Is your point that AMD/Intel are totally incompetent and despite decades of experience still don't know how to design good CPUs? Arm is tiny compared to Intel and manages to design about 5 new microarchitectures per year, of which ~2 are wide, high-end OoO designs showing 20-30% performance gains. Every year. So is it possible that the x86 ISA is a tiny bit more complex than you're claiming?

Yes. Find one vendor graph. Claim something and be completely wrong about it while you're at it.
Performance is not the same as power efficiency. Performance is not linear with power.
The graph shows IDLE-normalized consumption. So again. What the hell are you claiming?

You're claming that a ARM MOBILE SoC tailor made for a use has lower idle-normalized power than an generic desktop CPU?
And then somehow, this relates to absolute performance?

POWER EFFICIENCY IS NOT THE SAME AS PERFORMANCE.

[milkylainen]

So why did Intel fail to conquer the mobile space then? They burned quite a lot of money trying to bring x86 into phones, but they failed miserably. Of course the ISA is not 100% alone to blame for this, but x86 implementations need to deal with highly complex old tech which could hurt their competetiveness with ARM implementations. Jim Keller saw some value in reducing the complexities but I wonder if he had enough time at Intel to fix this (e.g. to invent a x20 ISA which deprecates old cruft within x86).

Because Intel was not / is not used to design for the mobile space. x86 CPUs has traditionally a completely different SoC layout with accelerators and whatnot.
Again. Nothing to do with the ISA. x86 design methodology for SoC on PCBs was a PITA. Very large memory interfaces, IO-pins, accelerators, SPI, I2C etc, etc- Shitloads of pins, southbridge, external phy's, etc, etc. How many ARM embedded CPUs support 40-128 PCIe I/O lanes that are still "battery"? None.

A server/desktop CPU shrunk to Embedded is usually a bad fit.
Esp. if the parent company thinks it can just "shrink stuff" and off we go. It has nothing to do with the ISA.
Just as ARM were bad fits for servers and desktop, x86 was a bad fit for embedded. Does not mean that ARM ISA can not do servers, or vice versa.

[PerformanceExpert]

5W CPU will never compete with a 100W CPU if spending the same transistors, same fabrication process etc. Something has got to give.
A 5W CPU will be absolutely slaughtered by anything spending 100W in the same fabrication technology.

Funny then how this graph shows how a 4.5W core slaughters high-end x86 cores using 4-5 times as much power. How do you explain that exactly? Similarly, Graviton 2 shows that at 110W you can beat every Intel server in existence and compete with EPYC 2 using less than half the power and a third of the silicon area.

Is your point that AMD/Intel are totally incompetent and despite decades of experience still don't know how to design good CPUs? Arm is tiny compared to Intel and manages to design about 5 new microarchitectures per year, of which ~2 are wide, high-end OoO designs showing 20-30% performance gains. Every year. So is it possible that the x86 ISA is a tiny bit more complex than you're claiming?

[ms178]

So. ARM can compete on high end servers. But they will be spending the same power as x86 to do so.
And x86 can compete on low end battery devices. But they will be as slow as ARM while doing so.
It's just a question of implementation.

No magic sauce.

So why did Intel fail to conquer the mobile space then? They burned quite a lot of money trying to bring x86 into phones, but they failed miserably. Of course the ISA is not 100% alone to blame for this, but x86 implementations need to deal with highly complex old tech which could hurt their competetiveness with ARM implementations. Jim Keller saw some value in reducing the complexities but I wonder if he had enough time at Intel to fix this (e.g. to invent a x20 ISA which deprecates old cruft within x86).

[DanL]

It would be nice if you could take corporate statements at face value..

But this looks to be Jon's personal blog rather than a corporate statement (by either Nuvia or RH). Sorry, but I'm not buying your "steaming heap of crap" conspiracy theory. I will buy that your conspiracy theory is a heap of steaming crap though.