Intel Releases Arrow Lake and Lunar Lake Instruction-set Reference Guide

[btarunr]

In a bid to prepare its ISV ecosystem for emerging technologies with future processor microarchitectures, Intel periodically releases instruction-set reference guides. The latest of these was leaked to the web, making their first references to the upcoming "Arrow Lake" and "Lunar Lake" client processor microarchitectures. From the looks of it, Intel is planning a massive push into the client AI acceleration space, starting with the upcoming "Meteor Lake" architecture that debuts later this year. The processor is expected to feature hardware acceleration for AI, with the new AI Boost feature.

The company could build on AI Boost with even more capabilities in the subsequent "Arrow Lake" and "Lunar Lake" microarchitectures. Among the instruction sets relevant to AI deep-learning neural net building and training, are AVX VNNI with INT8, AVX VNNI with INT16, AVX-IFMA, and AVX-NE Convert. There are several new security-relevant instructions, including SHA512, SM3, and SM4. "Lunar Lake" will introduce TSE-PBNDKB (total storage encryption). The ISA Reference Guide can be accessed here.



View at TechPowerUp Main Site | Source

 

[A Computer Guy]

It will be interesting to see how AMD responds. Will we see an AI chiplet or AI stacked die like x3d cache in the future?

 

[Fierce Guppy]

I had a look at the CPUID instruction in that manual. Just the massive amount of information that can be returned by the CPUID instruction given the value in EAX really does draw a picture of the sheer abundance of accumulated changes in Intel CPU micro-architectures over the years. So long as there's still an instruction set common to all CPUs since the very first Pentium then it matters not. It's just fascinating to me as an assembly programmer from way back.

 

[bug]

I had a look at the CPUID instruction in that manual. Just the massive amount of information that can be returned by the CPUID instruction given the value in EAX really does draw a picture of the sheer abundance of accumulated changes in Intel CPU micro-architectures over the years. So long as there's still an instruction set common to all CPUs since the very first Pentium then it matters not. It's just fascinating to me as an assembly programmer from way back.

I remember when I was dabbling with assembly, I was overwhelmed by the changes going from 16 to 32bit. I don't imagine me keeping track of whatever came since.

 

[aaa社保了啊]

It will be interesting to see how AMD responds. Will we see an AI chiplet or AI stacked die like x3d cache in the future?

In the future, 3D caching can be widely applied, with the addition of L4 caching from the 14th generation Core. Currently, AMD's caching technology X3D product line is one step ahead of Intel's

 

[Snoop05]

AVX512 where?
Intel shooting itself in foot once again...

 

[A Computer Guy]

In the future, 3D caching can be widely applied, with the addition of L4 caching from the 14th generation Core. Currently, AMD's caching technology X3D product line is one step ahead of Intel's

My only issue with the 3D cache is it seems very limited at least it's marketed toward games and many games that are not cache bound it doesn't help.
On the flip side the current limitations of integrated 3d cache end up forcing AMD to release incredibly power efficient CPU's because they just can't juice them to edge out intel without them exploding.
Perhaps I just don't know what it's being used for if anything other than boosting CPU bound cache sensitive games.

 

[bug]

My only issue with the 3D cache is it seems very limited at least it's marketed toward games and many games that are not cache bound it doesn't help.
On the flip side the current limitations of integrated 3d cache end up forcing AMD to release incredibly power efficient CPU's because they just can't juice them to edge out intel without them exploding.
Perhaps I just don't know what it's being used for if anything other than boosting CPU bound cache sensitive games.

Well, you've just described the limit for every cache: if your stuff fits, you get stellar performance, if it doesn't, you're back to slow mode.
The difference here is that when a new level of caching is added to a CPU it usually helps 80-90% of the workloads, whereas AMD's 3D cache is squarely aimed at gaming. It seems to be doing its job well, even if it comes with a few drawbacks, so I don't have anything against it.