Is The Beelink HX 370 Mini PC AI Ready?

Can you really cram an AI machine into a mini pc form factor? Wendell investigates the Beelink SER9 AMD Ryzen AI 9 Mini PC!

You can find us...

For all our links, check out our link tree!

Thank you for watching!

-------------------------------------------------------------------------------------------------------------

Music: "Earth Bound" by Slynk

[FrenziedManbeast]

Whew Wendell at $1, 000 USD this doesn't feel very disposable to me! Good gravy that's some serious dough for NUC.

[hugevibez]

These Ryzen AI HX-series chips don't support SO-DIMM memory, so user-serviceable RAM is a no go on this platform. We might get something with CAMM-modules in the future.

[perhapsjames_]

Everything is AI ready. Even my socks are

[isbestlizard]

I decided to go with the ser8, being able to stuff 96GB of memory in it is way more useful for AI for me than having a few more mythical TOPS on the soc

[jongranada]

Take a shot for every time he says “Zen 5”

[obnoxiousvf]

Somethings I would like to see are USB-C power and some form of study plug for an external GPU (something more rugged than a bare PCIe connector)

[konic]

"Background AI tasks" grumble grumble

[TheHalo]

LPDDR5 3750 in AIDA is LPDDR5 7500, the fastest strix point supports, soldered
strix Zen5 cores are no the same as Zen5 desktop
strix is AVX512 double pumped (like Zen4) and is not full fat AVX512 load and store.

Turin Zen5(c) and Ryzen 9000 are full AVX512

[Sunlight]

Imagine if AMD would support the IGPU and NPU with ROCm for AI generation. Then with 64+ GB of memory this thing would a beast.

[jag_g]

$999 and it's 'basically disposable', this statement makes me a little sad with where things are in the PC space 😂

[deegarkie]

Hear me out, We've got 8 cores with the same 8MB of cache (1MB/core) as Zen2 4700G(at least this time all 8MB is usable by all 8 cores instead of only 4MB available to any one core), and then 4 cores with the same ammount of cache per core as a 7700x(4MB/core), how about unifying the cache, then dropping those 4 cores and trading them for more cache, say a total of 48MB of cache for those 8 zen4C cores, this frees up memory cycles and bandwith for those 16 RDNA 3.5 CUs
Heck if its more efficient to have 8MB of L3, use that freed up space for 48MB of slightly more dense but slower L4

Sure more cache uses more power, but also, you're giving up 4 cores which uses more power than cache, you're going to a unified cache pool instead of a split pool of 16+8 (sort of like in Zen2 mobile 4+4), getting more cache per core, and while not gaming, hitting memory less, going off package less, using less power.

[MikeBob]

Amazing! Beelink never disappoints! Thank you, Wendellman! 🙏🏼👍🏼🙃

[ShowXTech]

Already hyped for the next Gen Intel MiniPCs.
Also haven't seen SO-DIMMs on Strix Point yet.

[AshtonCoolman]

I would have liked to have seen Oculink and/or Thunderbolt 4 on this thing. Using it with a dGPU would be nice.

[dozerd]

3:18 this is what I envisision Wendell listening to on the daily. The quirkiness fits

[rgalik]

My grandma doesn't need Zen 5c and a top of the line GPU ;D
Not sure about Wendell's

[erelpc]

NPU chips are great but it's a shame we're moving further away from general processors to everything being handled by discrete processors like NPUs.

[dinckelman]

The only thing AI is ready for is the burst of its bubble

[FrenziedManbeast]

Anyone know which version of Zen this PC uses?

[euduvda]

I just hope that at some point any NPU hardware would be usable by endusers.
Bonus points if it's not just LLMs!