AMD Ryzen AI 9 HX 370 12-Core APU Is Faster Than M3 Max | Hacker News Reader

AMD Ryzen AI 9 HX 370 12-Core APU Is Faster Than M3 Max

wccftech.com

37 pointsdoener16 comments

Here‘s the interesting bit:

> It wasn't mentioned if the chip was running at its full 54W TDP

The 14 core (which are 2 more than the Ryzen) M3 Max has a TDP of 78 watts, so if the numbers are correct this could mean a very impressive performance per watt (which is what actually matters, why does anyone care about raw power?).

Better yet, the chip appears to not even be made with the latest process (TSMC 4nm vs Apple’s 3).

TDP is mostly meaningless between manufacturers. You can (usually) compare TDP within a manufacturer but even then it’s a bit dicey across generations.

Not worthless but.... note that TDP is a thermal design spec.

It's not an electrical spec. The main difference is that a chip can overclock itself to 200W on 45W TDP, as long as the overclock period is short enough to not fill up the heatsink with heat.

And it can take minutes to fill up the heatsink in practice. Heat travels slower than electricity after all, even across copper and heat pipes.

Not through diamond, which is basically a thermal short.

I want to grow a single piece diamond heat sink.

All the R&D into lab-grown diamonds is meant for endgame use in computer chips, the jewelry trade is just a short-term means of financing the way there.

You can buy it: https://www.msesupplies.com/products/diamond-heat-sink-therm...

Even single crystal ones: https://www.msesupplies.com/products/single-crystal-diamond-...

> which is what actually matters, why does anyone care about raw power?

Raw power determines whether games run smoothly, how fast things compile, how fast a render takes etc. The performance per watt determines battery life but for some people that isn't particularly important. I can only run the Unreal Engine Editor for about an hour before the battery runs out on my laptop (5800H/16GB/3070). The raw performance when plugged in is much more useful to me than trying to squeeze another 20 minutes out of it. Of course I prefer using my desktop for those tasks, but sometimes I need to travel or demo so laptop (with charger) it is.

It would be TSMC 3, no?

The passmark numbers are almost identical to a 12th-gen intel i7-12700K, however the still-modern 12-core intel cpu has to pull about 4x as much power to do it.

Very cool if true! Unfortunately AMD and Intel have shown to not be able to sustain benchmark numbers over multiple runs. Apple with M-series has been able to do that more or less on the pro series macbooks. So I wouldn't be surprised if this chip beats the M3 Max on short term burst but will lose on real workloads like compilation.

How does one use the NPU features on such a chip from like a normal C++ program?

You need to install the XDNA and XRT drivers, then install the Vitis AI compiler (needs a free license) and then you can compile your own custom kernels using mlir-aie using plain old C++. They give you an easy to use API for vectorization. The hardest part is organizing the data transfer through the tiles.

In principle this can accelerate anything you can run on a normal CPU as long as it is a streaming workload i.e. single pass or your entire data fits in the memory tiles.

The vector registers are at least 1024 bit wide and you get 32 tiles/cores.

i don’t really know what Passmark is, but my gut feeling is to pay attention to single core performance + number of cores, rather than one [task] that perfectly uses all cores at once, no?

True, single core+ (maybe some minimal multicore work) and compute per watt are the two benchmarks that matter. Apple understands this...

Faster or not, API availability with no BS FW blobs or tooling and direct to hardware access will be winner here, otherwise no one will be able to program for this chips.

Here is some grounded discussion I could find about AMD Ryzen AI NPU programming https://github.com/ggerganov/llama.cpp/issues/1499

From there, here is list of Ryzen AI engine enabled laptop/mini PC models https://github.com/amd/RyzenAI-SW/issues/18

gather all that, I wonder how on earth these benchmarks comparing oranges to apples.

wccftech has repeatedly been unreliable source. IMHO should really be banned.

Reply on news.ycombinator.com