Thoughts on AMD’s Ryzen AI Max 395+ processor? Does anyone own a device with one of these?
How is the LTX 2.3 performance?
I don’t understand what an AI processor is. How does it help me use GPT in my browser
>>108905654>I don’t understand what an AI processor is.It's just like any other processor but twice as expensive because it has "AI" in the name.
>>108905654Whatever it is, Intel and AMD are making iGPU that are as capable as 2018 budget gaming dGPU even on business laptop. I wouldn't complain much.
>>108905576i do, its not worth it for ai stuff. but as a powerful computer, pretty good. wide memory bus is useful for running lots of 'traditional' work in parallel e.g. compiling etc.
>>108905789If i can't drop it into a desktop motherboard i don't care.
@grok please make picture of cat wearing flannel pajamas for funny 4channel post
it's just for ai slop, the gpu is extremely overpriced given that it loses to laptops at half the price
>>108905576i would buy one if we hadnt the ram price crisis.now i wait for the next model and cheaper prices
>16c/32t>on package GPU is equivalent to a desktop RTX 4060>total package power consumption is lower than any equivalent two die laptop or minipc>CPU perf is still near desktop equivalents as wellIts the best thing AMD has ever made
>>108905576is nice. i've been happy with my laptop. there may or may not be better options if you just want to do ai, but it's good as a regular machine besides. prices fucked atm though, so sucks for you.
>>108905576I've got an HX370. Not sure what's so AI about it other than the fact that you can allocate shared memory whichever way you like. I currently have it split 50:50 between CPU and GPU so that I can run a local model on it.
>>108905576It's $1,000 more expensive every time I see a company put it in their product. You could get a 64 core Threadripper build for the price of that CPU now.
>>108905576I've got a 395+ with 128gb of ram/vram for 'cheap' (i.e. for less than 1700) back in November. It's pretty neat, and a good all-round device imo. Does Vidya / Local models etc. quite well. Set it up with fixed 96gb of vram, seemed less hassle. Fits gpt-120 easily, although I definitely haven't set it up to full potential yet.With current prices (I've seen that spec reach 3k from vendors like hp) I absolutely can't recommend one though.
>>108909759Yes, current prices are absolutely insane for these and not worth it
>>108908429The have the "AI" branding because they feature a Xilinx NPU
>>108909814did you get it as a laptop or minipc?
>>108905576underrated af, but overpriced af. I really want one... I LOVE IT.
>>108905576Too expensive and AMD has godawful iGPU support. Only actual use is if you have massive boners for SFF machines and enough money for it, or if you absolutely need to run 96GB AI models.>>108905961You can drop a ryzen 8600g or 8700g into an AM5 board just fine. With some decent RAM (~7000MHz) it's about as fast as a RX 570. Upcoming HX 450 chips will be a bit faster but not much (RDNA 3.5 instead of 3).Unfortunately AMD gives zero fucks about APUs, so the big problem is that there's no upgrade path whatsoever. The successor Medusa Point still uses RDNA 3.5 (rebranded to 4m), and it's still 2-3 years away for desktop. After that they have Grimlock Point but we don't even know what iGPU that will use, but by the time that one gets released it'll be 2031 so it might not even make it to AM5.But the thing is, zen 6 will fix the horrible latency and power usage problems that made the APUs better, so there'll be no reason to use them once zen 6 is out. You can just use a RX7600 or something.
>>108910909>AMD has godawful iGPU support.Every single APU since Brazos and Llano have had great driver support. They just werk
>>108910909>Upcoming HX 450 chips will be a bit faster but not much (RDNA 3.5 instead of 3).Also replying to you again because you're a fucking retard and there is no change in GPU arch at all here as the HX400 series is just a refresh, not even a respin, its literally just improved binning of the exact same chips Strix Point and Strix Halo are the same RDNA 3.5, which is fundamentally exactly the same as the rest of RDNA3 but doesn't utilize the doubled ALUs per CU in order to keep xtor count and power consumption down.
How bad are the speeds vs a 90 class card
>>108910909>With some decent RAM (~7000MHz)amd shits itself on anything higher than 6000Mhz max
>>108911093The highest end SKUs with the fully enabled IGP are equivalent to a desktop RTX 4060 or mobile RTX 4070Only caveat being that you need to provide the chip a high enough power limit which some laptops featuring the chips do not
>>108905576Overpriced, underpowered. A mini PC price should be three digits, not four.
I have a Z13 flow with one inside. It's incredible. Basically the Linux / Windows version of the Apple M chips.>>108905654It's a buzzword, but the unified memory helps you run MOE ai models quickly.I use it to run 70-80GB ai models daily for local development. When you could get one of these 128GB mini pcs for $2k it was a good deal. For $4k now due to the artificial ram crisis its bs.
>>108911292What a fucking predatory scam.
>>108911566You can't clock without watts. Some vendors rather ship a thinner and lighter device.
>>108911292the real caveat is you have to deal with amd's dogshit software support
>>108912016back to plebbit you underage sperg
>>108905576Owned the 128 gig one, waste of money for anything other than gaming and general compute, gave it to my hobo fren. If you are planning to use it for llms you'll be treated to blazing fast grandma crawling across the kitchen floor speeds of 5 - 9 tokens per second and 27 + minute gen times in SDXL for a single 1024 x 1024 image.
>>108912617it will be good as a playground for agents though, with all that compute and memory bandwidth
Everything AI by AMD is shit
Mogged by Mac Studio
>>108905576its okits a marginally worse apple m4I have high hopes for zen 6 if it ever comes out
>>108912617gemma 4 31b with an mtp fork gets 10-15tok/s, qwen 3.5 122ba10 gets 30+ as an ngram copy pasting code monkey and 20 normally.sdxl is 45 seconds for 1024x1024. anima 60. klein 9b 90 in t2i and 3-4 mins in edit mode.anon.. be honest, you couldn't figure out how to use the gpu could you?
