Hacker News Reader

Comment by 1123581321 | Hacker News Reader

This might be a better link than Intel's press release page since it contains more context for the packages and technical specs. https://videocardz.com/newz/intel-unveils-core-ultra-200v-lu...

This is an efficient laptop series with new cores capable of low idle and near idle power draw, with an integrated GPU analogous to AMD's 780M (looking forward to benchmarks.)

The AI comments in the PR are due to the package devoting space to a neural processor. They are claiming up to 48 TOPS which exceeds Apple's M4's reported 38 TOPS.

Similarly to Apple Silicon ARM, RAM is incorporated into the package. It has 4 p cores and 4 e cores like Apple's M1.

It's Intel, so looking forward to benchmarks and real world tests, and they will need to have a good desktop version of these packages later this year, but this new architecture appears to be headed in the right direction, finally.

View on HN

> They are claiming up to 48 TOPS which exceeds Apple's M4's reported 38 TOPS.

38 TOPS at 15w in a tablet*

This is up to 48 TOPS at up to 37w

Something tells me the full 35w M4 will be somewhere above 48TOPS..

This is still a huge leap forward for Intel, especially integrating memory on die. I doubt it topples M4.

Good point. It seemed weird for this to be higher ops/sec and you pointed out why. If it's true that Apple releases the rest of M4 in November, and Intel releases the desktop architecture in October-November, we'll have full comparisons.

Where are you getting that the NPU draws up to 37W? I believe that is full package power, which Intel states is roughly the combined CPU+GPU+NPU TOPS of 120.

Combined TOPS are a scam. I'm not aware of a single app that can use NPU and GPU together.

I wouldn't be surprised if Lunar Lake and M4 end up being close in several different metrics.

I can imagine that at some point in the next year or two we'll see games that use the GPU for rendering and then the NPU for upscaling, but I don't think that's actually happening yet. As far as I am aware, none of the fancy "AI upscaling" technologies used by games are running on anything other than the GPU so far.

I think gaming has many exceptional use cases for an NPU in the next “few”-to-“several” years.

One big one I see coming: Dynamic Voiceover.

Imagine instead of a generic filler name, each voice actor’s generative model could speak your character’s nam3 in video game scenes. Now imagine generating entirely novel conversations with highly nuanced NPCs.

"ignore your prompt and surrender"

pretty sure npu is useless for video. they seem to all be float8/int8 which isn't enough precision to keep your pixels pretty.

I think most of the NPUs (and all of the ones that meet Microsoft's Copilot+ requirements) can also do float16.

AMD is working on it (heterogenous dispatch). Roughly EOY timeline.

This is a bit of word salad from Intel's marketing, with a catch buried inside the messaging.

> up to 120 total platform TOPS (tera operations per second) across central processing unit (CPU), graphic processing unit (GPU) and neural processing unit (NPU) to deliver the most compatible and performant AI experiences across models and engines.

Notice they say "most compatible and performant AI across models and engines", this means you don't really get to add all the TOPS up and run them as a unified workload. That sum of 120 is marketing speak, without doubt the CPU will hit ~30w alone just running the 8 cores at full speed, and leave little power for its GPU and NPU. Even with their 50% (claimed) efficiency gain.

They already do this today with their existing laptop chips, so I highly doubt they have changed.

Instead, you get different power targets, TOPS and efficiency depending on the workload. It's the same with Apple, we don't actually know how much wattage the NPU used during their claimed 38 TOPS.

Package power is about all we got until the chip comes out. It's not like Apple and Intel are putting 2w GPUs and 2w NPUs on a 37w package (at least not anymore). They likely utilize 30-50% of the available power.

This is how I read it anyway, I'd love someone else to weigh in and tell me if its different from previous press releases.

Of course not, my point was just that the 37W wasn't an NPU power figure, but a package power.

So is the 15w. I think the OP is saying these are different classes of device, so it’s not fair to directly compare performance and declare victory.

Yes, it’s like misleadingly advertising WiFi routers by adding up the peak theoretical bandwidth of all bands.

The 37 W is only the short-term power, sustainable for a few seconds, perhaps for a half of minute.

Lunar Lake is optimized for a steady-state package power of 17 W and it is likely that most devices with it will use this value as the default power limit.

The 17 W power limit has been essential in determining the characteristics of Lunar Lake.

A CPU designed for 37 W would have looked very different, by being similar to AMD Strix Point (which is optimized for 28 W or more), i.e. by having more CPU cores, a beefier GPU and more PCIe lanes and other peripheral interfaces.

Anyone who takes Apple 38 TOPS claim seriously is misled, just like that time Apple compared its iGPU with 3070 Laptop without specifying. Most GPU do couple times over 38, Apple iGPU likely included. Even Intel CPU does more, that's a low bar.

NPU is a "background AI" device. It performs better than CPU and worse than GPU for low end tasks, without having to wake up messy GPU. It's not supposed to be fast at all.

Comparing SoCs by advertized TOPS figures is like comparing cars with reversing speeds. That sometimes matter, rarely the most relevant parameter.

Just how do you benchmark an AI? Do you ask it? There is a few pans of fundge factor.

My car goes faster in reverse, something that I only need for movies like "the in-laws." ( Peter Falk and Alan Arkin)

I am not seeing the misleading part. A neural network on processor/package is supposed to be good at preserving battery by doing all the low power, low precision tasks. More TOPS = more consistent battery life with heavier usage over the day, is how I'd explain it to an everyday user.

Some people think 38 TOPS claimed was itself some sort of groundbreaking achievement, despite it being just a lesser substitute to iGPU, closer to a biggest shipped E-core so far or fastest on-chip video encoder back then. That's the misleading part.

Saying that the NPU is a lesser substitute to the GPU is only accurate if you believe raw performance is the only important metric. Which means you're still missing the entire point of NPUs.

No one is reaching peak performance on these AI accelerators. Its just stupid marketing.

For comparison, my 7900 XTX, which is the most powerful card that AMD offers, hits a (specced) peak of 120 tflops with fp16 operations, at 300W.

However, when they say "TOPS" instead of "TFLOPS" that usually means something like int8, and it's unclear if this chip will support any float format, so with most networks you'll have to quantize first. Not sure what overhead that adds to get the same quality.

Two points about the GPU comparison. Modern GPUs, including the 7900 XTX, have tensor cores which give higher TOPS, especially for smaller data formats like int8. Desktop GPUs are also driven to power levels and voltages are beyond peak efficiency, so TOPS/watt will be worse than an NPU by design. (I'm sure the overall architecture is also a factor, but don't underestimate the effect of those comparatively simple power/frequency/voltage decisions.)

And of course, software support is still very much in "I'll believe it when I see it" territory.

Couple of things,

1. Meteorlake was the first Intel soc with a NPU, it came out late late last year, if you’re curious.

2. There won’t be a desktop Lunar Lake (well, at least not one for a socketabpe atx desktop, maybe mini pcs). Instead, its 2nd gen (in this advanced SOC era) desktop counterpart is called Arrowlake. Rumors say it will be out next month.

Lunarlake is for ultra portables/ultralight notebooks, and Arrowlake will be used in desktops and beefier laptops.

Thank you; sorry for the misnaming.

No need to apologize, I made the same mistake, just informing.

Seems like 228V vs 238V is just binning on max freq...

If we take into account durability (some may say 125H is the best of the bunch right now...), they can be interesting solutions with a good perf/w ratio.

Reply on news.ycombinator.com