News Intel's Lunar Lake CPUs to use on-package Samsung LPDDR5X memory

[Admin]

Samsung reportedly scores a big design win with Intel to supply LPDDR5X memory for Lunar Lake CPUs.

Intel's Lunar Lake CPUs to use on-package Samsung LPDDR5X memory : Read more

 

[usertests]

and up to a six-tile NPU 4.0 AI accelerator

Help. I don't know squat about Intel's NPUs. How many TOPS do they claim for Meteor Lake, and how many tiles is it? Why is this NPU version 4.0?

Lunar Lake is supposed to have triple the AI performance, so I'd guess 2-tile in Meteor Lake.

 

[bit_user]

Help. I don't know squat about Intel's NPUs. How many TOPS do they claim for Meteor Lake,

Dunno, but it's not faster than their 128 EU iGPU (which lacks XMX cores, IIRC):

Source: https://www.tomshardware.com/news/i...meteor-lake-architecture-launches-december-14
​

and how many tiles is it?

Based on this slide, I'm guessing 2?

(same source as above)
​

Why is this NPU version 4.0?

Intel's original AI accelerator was called GNA (Gaussian/Neural Accelerator), which underwent 2 revisions. I guess they must be including those, in order to count the upcoming one as generation 4.

Lunar Lake is supposed to have triple the AI performance, so I'd guess 2-tile in Meteor Lake.

Yes. Given Gelsinger's recent comments about Lunar Lake having 3x AI performance, it lines up nicely with scaling from 2 engines to 6.

 

[bit_user]

It is set to come with either 16GB or 32GB of LPDDR5X-8533 memory-on-package, reducing the platform's footprint and improving performance compared to traditional platforms featuring either memory modules or soldered-down memory chips.

The bandwidth will be nice, but the real benefits are cost & power savings. I think you could probably achieve LPDDR5X-8533 off-package soldered (or LPCAMM modules?), but possibly at greater cost & power.

Note that LPDDR5(X) has significantly worse latency than regular DDR5. So, it's not a pure win, especially if they limit the memory datapath to just 128-bit. The main benefit for Apple of using on-package memory is their ability to scale up to 512-bit, but that certainly won't happen in a "thin & light" x86 laptop.