AMD Halo Box (Ryzen 395 128GB) photos

[-]

-Akos-@reddit

Asking the real question..

In this economy, if you gotta ask, it's too much (insert cry emoji)

[-]

frightfulpotato@reddit

I'd be amazed if this was ever released to the public. More likely an engineering reference model. Anyone remember Project Quantum?

[-]

Fastpas123@reddit

Lmao... Yeahhh..... I'd love a decent little box like this for my AI needs but. If it costs 10k then it's better if I just forget it exists lol

[-]

knucles668@reddit

No way it’s $10K. DGX Spark is $5K with 128GB and the NVFP4 optimization. I don’t think AMD has any comparable special sauce. I say $3500 to be above costs and gain market share.

[-]

Exactly, I'll enjoy Claude via my work account, and at home I'll enjoy my LFM 2.5 1.2B which screams on my potato laptop, and fantasize have how I'll one day be able to run a big boy LLM on my own hardware..

[-]

MuDotGen@reddit

I am also just now hearing about LFMs. Seems they were designed to run on CPU, so curious how they perform against the competition for being CPU only.

[-]

MakerBlock@reddit

I've not used LFM before - how is it?

[-]

DataPhreak@reddit

Ooofff.... I got my strix halo for 2000$ in October. They are now.... $3299....

https://www.amazon.com/dp/B0F53MLYQ6 - same unit.

[-]

Fi3nd7@reddit

Literally anyone who bought hardware 6 months ago has made money. This isn't unique to the halo, everything is more expensive.

[-]

DataPhreak@reddit

Yeah, I thought that was mostly components though. I didn't realize it was affecting integrated systems like APUs. Also, this went up $1200 or 60%. The DGX Spark only went up $500 or 12.5%. I suspect price gouging here more than actual chip shortages, since these are printed directly on silicon.

[-]

SexyAlienHotTubWater@reddit

The spark was very overpriced up until quite recently (arguably it's still overpriced). It was designed as a dev box for people who make a lot of money designing training runs for larger clusters.

[-]

MarkoMarjamaa@reddit

I bought my Bosgame for about 2000e and not it's 2385e?

[-]

BannedGoNext@reddit

My guess is shy of 4000.

[-]

RoutineLengthiness32@reddit

For 4K, many people will not buy these, as you can already get a dgx spark at this price.We are currently using dgx spark for our physical AI on the cobolative robot btw.

[-]

BannedGoNext@reddit

Yea, but if they are at 3300 right now, and they are saying this is better I would guess they would charge more.

I don't think the halo is worth 3300 at all, and I have one. I got mine for 2000, that's an OK value, but even then it was marginal.

[-]

Infninfn@reddit

OEMs going for around $3300 now. By the time I got round to getting one it was $2300 but still glad I did.

[-]

BannedGoNext@reddit

got mine for 1600 on a sale :). It's so slow, but I can run a lot lol.

[-]

Fuzzdump@reddit

The Framework Desktop 128GB is currently ~$3100, so that's a ballpark.

[-]

DepthAggravating3293@reddit

Based on April 2026 reports, AMD Ryzen AI Halo (Strix Halo) systems featuring the Ryzen AI Max+ 395 are currently priced between $1,500 and over $3,399, with premium, high-memory configurations (128GB+ RAM) rising to around $3,399.99 due to component demand. These mini-PCs and laptops are designed for local AI inference and development, often compared to cost-effective alternatives for cloud GPU rentals. [1, 2, 3, 4, 5]

[-]

cafedude@reddit

$34.99

... oh, wait, you need to move the decimal point a couple places to the right.

[-]

ShelZuuz@reddit

34990

[-]

CatalyticDragon@reddit

One assumes about the same as other Strix Halo boxes. Which was 2k but now 3-4k.

[-]

1ncehost@reddit (OP)

Not released yet but probably about what the existing 395 partners charge.

[-]

Ell2509@reddit

How did you?

[-]

1ncehost@reddit (OP)

This one is at AMD's AI Dev Day

[-]

Ell2509@reddit

Oh wow you fortunate person! I would love to get the chance to attend something like that.

Must be work related?

[-]

darktotheknight@reddit

This is the only question that matters.

[-]

phreaqsi@reddit

If you have to ask...

[-]

lukaszpi@reddit

Stop cheering these unupgradeable, dedicated computers that just end up as waste. Make corpos to release components that you can interconnect, stack and mod to extend their life and purpose!

[-]

Both_Opportunity5327@reddit

Have you got one?

These things rock with their 16 cores 32 threads.....

No way this is going to be a door stop in even 10 years.

[-]

lukaszpi@reddit

Can you build upon it? Expand it? Replace soldered components that in bigger form factors are replaceable?

[-]

Both_Opportunity5327@reddit

That is not what it is for, these are small form factor devices. And mine has lots of ports and is under a TV in the living room, I have Ubuntu loaded and use moonlight to cast to my main computer, it run agents and compiles code.

Why would I need to upgrade it.

Also I have lots of desktop PC's five, two Enterprise Laptops, Macs and small handhelds like the Thor, SteamDeck.

Everything has its place.

And say I had a machine that was better for the tasks it runs now I would just turn it into a console for my kids....

[-]

beedunc@reddit

It’s a doorstop already.

[-]

Both_Opportunity5327@reddit

Is a 3090 a door stop? how about a 3060?

Ryzen 395 base computers sip power, and serve MOE models like a champ much faster than the equivalent Macs....

[-]

beedunc@reddit

Mac mini smokes this overpriced piece of junk.

[-]

PWCIV@reddit

And slower memory than a macbook lmao

[-]

Maximum_Transition60@reddit

this is gonna be wonderful for me to get once they are 100$ or in dumbsters waiting for me to pick them up

[-]

Quind1@reddit

This is my concern, too. At the rate new AI models are being released, this will become a doorstop a year from now.

[-]

Icy-Pay7479@reddit

It seems like the opposite is happening. At this rate it will be twice as useful in a year.

[-]

Late_Film_1901@reddit

And keep increasing in value!

I just got a refund on a 5 year old nvme drive that I RMAd recently and I can't even buy a replacement for the money. What an absurd timeline.

[-]

techno156@reddit

It might not be as good for bleeding-age models, but it'd still be a pretty decent computer.

[-]

DataPhreak@reddit

You don't understand. These have to be unupgradable in order for them to have unified memory architecture. They're actually really good regular computers that happen to be able to run the small-medium MoE's at a reasonable speed. You can still add a GPU cradle and if you can afford it you will probably want to. Basically, the idea is you put an a3b/a4b MoE on this, and run your TTS/STT and image/video gen on a regular GPU, as well as a tiny 4b for fast text gen that doesn't need to be smart.

[-]

createthiscom@reddit

you mean like the MZ73-LM0?

[-]

ViRROOO@reddit

Dont get me wrong. I own a framework desktop, I hope this is successful and they release a * proper * upgrade with medusa halo. But this feels super late to the party and not technically interesting anymore. What is even sad is that just now, 16 months after the chip release, rocm is kinda (emphasis in kinda) stable.

[-]

_derpiii_@reddit

rocm?

[-]

DataPhreak@reddit

ROCm is the AMD version of CUDA. Well, kinda. Not really. But it's a good way to think about it.

[-]

tecneeq@reddit

It is. In fact, it is that much, that when you link llama.cpp with ROCm and CUDA, the shared libs use the same function names and neither will work 😉

[-]

DataPhreak@reddit

Well, actually no. They both do similar things, but in different ways. ROCm is not CUDA. You don't have X number of ROCm cores like with CUDA.

[-]

tecneeq@reddit

I know ROCm is not CUDA.

ROCm is a copy so blatant, the shared libs function calls are 99% the same name.

[-]

superdariom@reddit

But I use Vulcan instead because it's faster bleugh

[-]

ViRROOO@reddit

Well, its not faster. Its stable, uses less memory, does not crash every inference restart, and does not take 25gb of disk.

[-]

spaceman3000@reddit

I use precompiled llamacpp-rocm from lemonade guys and 0 crashes. I use it 24/7

[-]

ViRROOO@reddit

Its fine if you use the pre-compiled versions. I'm not interested in that, I contribute to llama and other topics like tinygrad. Just as a point, AMD is the * only * GPU vendor that you have to full restart the system after the GPU crashes. So if you are just running inference on a month old model using lemonade is fine.

[-]

spaceman3000@reddit

Not lemonade. Llamacpp compiled by lemonade team. It's nightly based on nightly rocm.

I have strix halo for a year and never had a GPU crash. I also have 9070 t connected over oculink to the same pc and also no crashes. You're doing something wrong...

[-]

Sea-Attention-5815@reddit

What llms are using for voice assistant and image/video generation?

[-]

Middle_Bullfrog_6173@reddit

As your link shows, Vulkan is generally faster in tg and slower in pp. Personally I find the prefill good enough and generation limiting so that's an easy choice.

[-]

DataPhreak@reddit

I spent 2 days trying to compile fast-attn for the gfx1511, after finally succeeding, it doesn't work because of a bit architecture discrepancy on the hardware itself. AMD wants fp32, but fast-attn is native fp16. Which means I can't accelerate a lot of stuff like T2M or TTS. They'd have to build flash-attn from the ground up to get it to actually work on the strix halo.

[-]

LumpyWelds@reddit

I've got one two. Love it.

Mines stable as a rock. I installed 24.04 and then brought up ROCm to 7.2.1. Very happy so far.

But if I do have one complaint that is nobodies fault.

I trying to get vLLM to run for it's MTP/NEXTN speculation which is needed for Qwen3.6 as none of the speculation modes in llama.cpp work with Qwen's hybrid architecture.

I just can't get vLLM happy. So I'm hoping llama.cpp eventually supports MTP/NEXTN.

[-]

zrebar@reddit

Love mine! Dear AMD: please streamline the ROCm support.

[-]

rmyworld@reddit

Cool. How many FPS do you get on Halo Reach?

[-]

FoxiPanda@reddit

Every time I see one of these I just want to whisper in every AMD executive's ear "more memory bandwidth please"

[-]

ProfessionalSpend589@reddit

They’re already doing it: https://www.amd.com/en/developer/resources/technical-articles/2026/day-0-support-for-xiaomi-mimo-v2-5-pro-on-amd-instinct-gpus-.html

And it seems software support is great for their real customers.

[-]

FoxiPanda@reddit

Being one of those real customers, their software has greatly improved in the past 2 years. I won't claim it's on par with NVIDIA at this point, but they're decent. The problem with the Instinct cards is that they're 10x the cost of this thing just for one. A full system of them is north of $500K now. Two entirely different use cases.

The problem with these little boxes is they want to compare to the DGX Spark and they do a fine job at that, it's just that the DGX Spark is abysmal already memory bandwidth wise even compared to an RTX 5090/RTX Pro 6000 or a Mac Studio w/ an M3 Ultra (though Macs have their own issues with low PP).

128GB of unified memory is basically wasted when you can only get ~250GB/s of memory bandwidth (this is about the same as a Mac Mini). It's just too slow for any real LLM use at the 128GB memory size - go load up a 80GB-on-disk dense model and you'll see what I mean (hello 4-8tok/s depending on context). So, since they're that slow, the only real use case is MoE models (which are great for a lot of use cases), but even there, a Nemotron-3-Super-120B-A12B example only gets ~16tok/s TG early in context and still memory bandwidth limited. That speed not super usable even at token zero much less 100K.

Meanwhile, Apple also has unified memory and >800GB/s of memory bandwidth in 2025 (which is why they're sold out) but is hampered by the low PP tok/s (a serious pain point)... but it's pretty clear that they'll get to ~1.0-1.4T/s in 2026 with improved PP too thanks to the new cores/instructions in M5 (most likely to be 1228GB/s - TBD though).

RTX 5090/RTX Pro 6000s are both at 1.8TB/s and have excellent PP for their cost, but require extra infrastructure (a computer) and are expensive/difficult to come by (~$3500/~$9000) ... which then puts you in $5K / $11K territory by the time you have a functional system...which is not great but it's a lot better than 250GB/s memory bandwidth.

To get a 'usable' 128GB mini AI PC, I think you need to see about 1.0-1.5TB/s memory bandwidth and solid PP tok/s compute. If you can get there, then things really start to come together for dense models in the ~80GB-on-disk range. 500GB/s memory bandwidth on the roadmap for mid-late 2027 doesn't even cross my radar - I expect NVIDIA to deliver 3 things by then: A new DGX Spark that will probably be a lot better than 500GB/s, a new DGX Station (Rubin) that will be incredible (both in perf and expense), and something less-than-a-DGX-station in the $20-30K range that might be super appealing for serious local AI in that timeframe.

Anyway, I'm just ranting at this point, but the TLDR is still "more memory bandwidth please" lol.

[-]

_derpiii_@reddit

> Macs have their own issues with low PP

What does PP mean?

[-]

FoxiPanda@reddit

Prompt processing.

[-]

JohnBooty@reddit

Holy crap, amazing explanation. I never clearly understood those aspects. Thank you for taking that time.

Question: for the M3 Ultra, what might performance look like for subsequent prompts? When, presumably, a significant of the prompt is now cached (at least for agentic workflows with large system prompts, that remain invariant between prompts?)

[-]

FoxiPanda@reddit

It’s wildly better. So if your kv cache is hot and you only add in say 500 new tokens, then that now 20000+500 token prompt only computes the cache for the 500 new ones (there’s a bunch of caveats here too where you can mess up your cache lol but let’s say you did it right)… so 500/100 = 5 seconds to do the PP and then however many output tokens divided by the token generation rate (we said a flat 50 but that actually goes down as you get further into your context window) but lets say 200 output / 50 tok/s = 4 seconds…

So 9 total seconds. Not so bad right?

[-]

_derpiii_@reddit

Ah. IMHO, 'prefill' is less vague :)

[-]

FoxiPanda@reddit

Yeah the terminology game is silly. I prefer prefill/decode as well, but so many benchmarks use PP/TG now that I've gotten into the habit... and of course PP has other cultural connotations too /facepalm

[-]

JohnBooty@reddit

Yeah I was like, “low pp? there are pills for that now!”

But I didn’t want to derail the conversation with a top level reply hahahaha

[-]

_derpiii_@reddit

somehow it's slipped past my perusal. I was legit confused ahahah :)

[-]

Herve-M@reddit

I remember to have read the contrary about the software from AMD; especially for MI300X.

Didn’t they start very badly, due to low support from most software and having to manage forks etc?

[-]

FoxiPanda@reddit

Yeah two years ago it was pretty bad. It has improved a lot since then.

[-]

Mochila-Mochila@reddit

500GB/s memory bandwidth on the roadmap for mid-late 2027 doesn't even cross my radar

Preach it ! We must relentlessly campaign for more bandwidth 🙏

I'm a nobody... but you, as someone in the business, have at least a remote opportunity to make your voice heard from AMD's reps.

[-]

johnkapolos@reddit

Great post!

[-]

guai888@reddit

This kind of machine is good for verify AI workflow. Nvidia/AMD did this on purpose because they want to sell higher margin machine to Datacenter. Apple's machine is great for chat, agentic work but not good enough for image/video generation.

[-]

FinalCap2680@reddit

And more memory!!! 128 GB is a joke

And better speed. And better cooling. And better networking.

Some upgradeability would be nice too...

PS did I forget realistic pricing? I looked at those at \~1500-1600€ and passed.

[-]

JohnBooty@reddit

Thing is, they’ll never release something that will remotely encroach on the performance of their datacenter hardware.

Otherwise people would just fill datacenters with giant stacks of these things.

[-]

gh0stwriter1234@reddit

Medusa Halo is supposed to double it... so its on the roadmap.

[-]

FoxiPanda@reddit

Last I heard Medusa Halo wasn't until like... mid-late 2027? To get to like ~500GB/s memory bandwidth?

While it might be pretty decent on the low end of the spectrum, it needs to double again to be relevant in that timeframe I think...I really hope AMD is cooking up a fast follow on generation with a way bigger memory controller.

[-]

StevenSamAI@reddit

Sure, but I think the main thing is that it is closing the gap in a very real way.

On the hardware side we are setting a 128GB, low power consumption system, with a consumer price tag and increasing memory bandwidth.

On the software/model side, we are seeing efficiency improvements and very capable models with 5-20Bb active parameters that can fit into 128GB with room due a decent account of context.

Fire me, Claude Sonnet 3.5 was the first model I felt was really making a significant difference to my work, capability and productivity.

With that gap closing between hardware and model, is already pretty incredible what we can run on <$5K of hardware, and it's only getting better.

If we could be looking at anything closer to 1000TB/s of bandwidth, with 192GB+ ram for under $5k by 2030, then I think we'll be in a pretty crazy place

[-]

gh0stwriter1234@reddit

I mean they could just slap 128GB of HBM on there and call it a day instead.... assuming they get some fab to build a low cost variant. The expensive part of it is not so much the stacking, its the base die that adds extra cost... otherwise its no different than any other chips that are stacked like flash uses stacked as well as ultra high layer count chips.

[-]

Slasher1738@reddit

roughly the same bandwidth as DGX Spark. The real issue is that its RDNA 3.5 and not 4.

[-]

StardockEngineer@reddit

Lack of compute is the real problem.

[-]

Slasher1738@reddit

Has more than enough to run most models. Stop your whining and go prompt something

[-]

xornullvoid@reddit

Is there oculink?

The FEVM FAEX1 has oculink.

[-]

tecneeq@reddit

No. It will have USB4 like the Bosgame M5 as the fastest connector.

You will be able to use a NVME to oculink adapter as it has more than one NVME slots. Like the Bosgame M5 that came out a year ago.

[-]

xornullvoid@reddit

How many m.2 slots? The faex1 has 2 for nvme, and one more where the m.2 to oculink is attached. I think it wins on external connectivity. Also has dual ethernet, dual USB 4 type c.

[-]

tecneeq@reddit

Wins indeed. The Bosgame has 2 NVME, bunch of USB3, USB2, two USB4 and 2.5GB Ethernet as well as HDMI or Dipslayplort.

Not sure about the AMD yet, but i think it had two NVMEs.

[-]

hellomyfrients@reddit

pretty much a 1/1 from the gmktec strix 395/128gb pc i use

I fucking love it, I think these things will sell sell sell

[-]

RetiredApostle@reddit

No CD-ROM though.

[-]

PassengerPigeon343@reddit

How will we play Roller Coaster Tycoon?

[-]

gh0stwriter1234@reddit

https://openrct2.io ... with your iso you already archived a decade ago?

[-]

PassengerPigeon343@reddit

In all seriousness if anyone reading this is nostalgic for the game OpenRCT is amazing

[-]

gh0stwriter1234@reddit

It lets you do all the insane rollercoaster configurations the original game didn't allow also by enabling extra features in the settings.

[-]

Porespellar@reddit

Will it allow the https://en.wikipedia.org/wiki/Euthanasia_Coaster ?

[-]

gh0stwriter1234@reddit

You can in fact build similar things the game isn't designed to simulate park attendees dying due to g forces AFAIK though you can crash them to death though.

[-]

kyr0x0@reddit

Okay, enough Internet for today.

[-]

Hello_my_name_is_not@reddit

Only one way to find out!

[-]

manoman42@reddit

Insane nostalgia!

[-]

xmo3px@reddit

Just tell your local llm to build it for you.

[-]

nckai@reddit

And Theme Hospital?

[-]

floconildo@reddit

Nor floppy disk either smh

[-]

Maleficent-Ad5999@reddit

Japan recently decided to deprecate floppy disks.. maybe that’s why

[-]

Maximum@reddit

Can it even read punched cards?

[-]

ThisWillPass@reddit

No phono jack… at least it’s water proof…

[-]

Evening_Ad6637@reddit

Wait let me start emule first

[-]

kyr0x0@reddit

You didn't get the KaZaa update? Trillian anyone?

[-]

Jumpy_Fuel_1060@reddit

No parallel ports either? Pfft pass

[-]

dpenev98@reddit

That's a sexy box!

[-]

tecneeq@reddit

Have a Bosgame M5 since some time. Paid 1800€. Have programmable LEDs as well. What am i missing?

[-]

1ncehost@reddit (OP)

Amd badge lol

[-]

_lavoisier_@reddit

how is this different than other ryzen ai max+ minipcs?

[-]

tecneeq@reddit

It's pricier.

[-]

tamerlanOne@reddit

Dovrebbe avere un 10-15% in più di potenza cpu mentre il resto rimane invariato.

[-]

andy_potato@reddit

Could be a nice lobster home, depending on the price.

[-]

hejj@reddit

The floor for 128Gb 395s is $3k

[-]

tecneeq@reddit

$2400: Bosgame M5.

[-]

rorykoehler@reddit

I can't believe they did a press release for sticking a product they brought out a year ago in a new box.

[-]

TheBolypt@reddit

But can it run Crysis?

[-]

maxpayne07@reddit

Probably, 1080p, medium. 45 fps or so.

[-]

Due_Duck_8472@reddit

slow as a glacier right?

[-]

CheatCodesOfLife@reddit

but it's got programmable LEDs!

[-]

Due_Duck_8472@reddit

I rest my GPU

[-]

Own_Mix_3755@reddit

It will has to be significantly cheaper than Nvidia DGX Spark knockoffs - but eg in our country here in Europe you can get the Asus GX10 for little over the price you can get the Beelink minipc with Ryzen 395 and 128gb ram. And you get the whole Cuda magic packed with it and not to mention the possibility to connect more units together. So if this will be around the same price (which it wont) its dead on arrival for most (if they dont add any software magic).

[-]

Late_Film_1901@reddit

Yes it was attractive when GMKtec Evo X2 was at or below $2k and dgx spark was $3500. Right now the prices have almost equalized, with Asus ascent and strix halo basically at the same price, and in that case gx10 is a no brainer.

[-]

Slasher1738@reddit

agreed, gotta be 2000-2500.

[-]

Queasy_Asparagus69@reddit

Ok. Looks like a case…

[-]

Admirable_Flower_287@reddit

No PCIe slots?

[-]

OnkelBB@reddit

no fast port for clustering. meh.

[-]

DataPhreak@reddit

These used to be $2000 for the 128gb version. Now they are $3200. At this point you might as well buy a DGX spark anyway.

Still, the Strix Halo is good. But no fast-attn yet. Maybe one day. It runs star citizen tho.

[-]

bcredeur97@reddit

Might be 10gbit RJ45 at least? (I hope)

[-]

starkruzr@reddit

they could choose to toss a couple of the useless USB4 ports and give us at least an 8 lane slot. alas, no.

[-]

oxygen_addiction@reddit

The chip can't do 8 lanes, m2 and usb4. You have to pick.

[-]

starkruzr@reddit

I know; that's why I'm saying toss the USB4 entirely. that eats 8 lanes by itself -- there's your 8 lane slot right there.

[-]

gh0stwriter1234@reddit

Yeah if AMD is going to pitch this... they need to make it special in some way, AMD literally owns a ton of high speed network gear IP... no reason for this not to have dual 100Gbit/s SFP28 so you could link up a bunch of them in a poor mans cluster.

[-]

starkruzr@reddit

yes. the important thing is how low the latency can get with RoCE.

[-]

gh0stwriter1234@reddit

Which it probably doesn't even support lol... sad.

[-]

RoomyRoots@reddit

It's a toy. They would use that for servers and ask for much more money. I still don't really grasp why they build this.

[-]

gh0stwriter1234@reddit

Yeah alot of AMD chips have a 10GBe controller built into them that is never wired out I think they use it for debug testing but I think the only one I ever saw wired was on an embedded server variant of Zen 1 or 2 but most of these chips still have it I am pretty sure sitting there as dead silicon.

[-]

flockonus@reddit

Ran any models on these? What tk/s do you get?

[-]

truthputer@reddit

I'd been looking at getting one of these and it's basically not worth it if you already have a gaming PC with a decent video card. The Strix Halo is memory rich and speed / GPU poor: if you have a model that needs 96GB of memory to run, the GPU can only run it very slowly.

So that means running MOE models, but those already do really well on conventional graphics cards and spilling over into system RAM.

[-]

spaceman3000@reddit

Yeah we run it for 16 months. Google strix halo

[-]

1ncehost@reddit (OP)

This is a normal ryzen 395. Same everything to existing products with that chip.

[-]

RegularRecipe6175@reddit

Interesting that AMD decided to compete with their customers (HP, ASUS, Framework, etc.) who were already complaining about limited availability. Unless AMD increases production and allocates it to its integrators, this will drive prices up for consumers.

[-]

zackofalltrades@reddit

Someone sat on the Steam Machine.

[-]

RoomyRoots@reddit

It kinda looks both basic and pretty at the same time. The material looks cheapish.

[-]

MisticRain69@reddit

I like how it looks but I have a gmktec evo x2 and it has a big ass heatsync and is way bigger than the new AMD one and if you dont power limit it to 85 watts or so it can quickly thermal throttle so I wonder if they really nerfed the power limit on the new one so it doesnt melt or something

[-]

pulsar080@reddit

But what can I say, wiping dust off it will be a separate torture...

[-]

RoomyRoots@reddit

Gotta shake that mf every couple of months.

[-]

Longjumping-Bake-557@reddit

Why the hell can't we have this in a normal ATX form factor?

[-]

spaceman3000@reddit

Look at Framework PC

[-]

someguy@reddit

No PCIe lanes.

I think it has 9 or something.

[-]

count_dijkstra@reddit

It's disappointing that this product exists. RDNA 3/.5 has no dedicated tensor cores. The I/O die (that also houses the IGP) doesn't have dedicated pinouts for high-bandwidth, low latency networking.

It's just a gaming laptop solution. Very cool, but not interesting at all for ML. People were disappointed by the Spark, this is even weaker.

Now if they could somehow rebuild the I/O die with drop-in replacement using Navi 44, and price it significantly lower than the Spark... The compute will be an order of magnitude lower than the GB205 in the Spark, but it'd be a viable toy.

[-]

spaceman3000@reddit

This is strix halo that exists for a while. I'm using it for quite some time (ms s01 ultra) and I love it for llm.

[-]

shydrangeae@reddit

Gonna limit my enthusiasm and say this is just another supply chain for something functionally identical to the existing 395 boxes. But hey, maybe that keeps prices from going even higher.

(And then I can be very pleasantly surprised if this is somehow tuned a tiny bit better than the other 395 impls or they try to undercut the others on price. And hey, maybe the design is partly them working on a template for an eventual 495 box that is 10% faster and uses 10% less power. Gal can dream.)

[-]

_derpiii_@reddit

> this is just another supply chain for something functionally identical to the existing 395 boxes.

Yes! I don't understand why they're just iterating on the exact same chip set, yet making it seem like a big deal.

[-]

No_Ad_8807@reddit

How does this compare with the Nimo PC performance wise?

[-]

No-Trifle4243@reddit

price if available please?

[-]

toffee0_0@reddit

is it scalable ?

[-]

0utlookGrim@reddit

AMD engineers: finally, the note block pressure plate.

[-]

WithoutReason1729@reddit

Your post is getting popular and we just featured it on our Discord! Come check it out!

You've also been given a special flair for your contribution. We appreciate your post!

I am a bot and this action was performed automatically.

[-]

uti24@reddit

Someone already has mentioned it's like Ryzen 490

[-]

1ncehost@reddit (OP)

I confirmed with an AMD engineer and he said its 395.

[-]

fallingdowndizzyvr@reddit

For now, but the 495 is supposed to be a drop in replacement. Which makes sense, since the 495 is just an OC'd 395.

[-]

CommunityTough1@reddit

What do you mean drop in replacement? These are mini PCs that aren't designed to be serviceable. The CPU and RAM are soldered to the board.

[-]

fallingdowndizzyvr@reddit

I mean drop in replacement. As in you can use a 495 instead of a 395 on a board. It's a direct drop in replacement. Soldered or not.

[-]

Pitiful-Rip-5854@reddit

It allows the device manufacturer to release a new product with the faster CPU with minimal, or no, other hardware changes. It isn’t for the user to change.

[-]

gh0stwriter1234@reddit

Its a rebadge of the same CPU with modest clock speed bumps they do this all the because mini PC and laptop vendors like to have yearly product models.

[-]

LankyGuitar6528@reddit

It's lovely.

[-]

megadonkeyx@reddit

reminds me of sun hardware

[-]

funding__secured@reddit

Meh

[-]

im_just_using_logic@reddit

Yummy

[-]

nostriluu@reddit

I dream about a chunky convertible Thinkpad with Strix Halo because of its great overall performance, tons of shared RAM, ability to run large models (slowly), but it doesn't make much sense as a stationary system. The lack of a fast RAM expansion focused $2k-ish system in x86 land looks really bad tbh.

[-]

HIGH_PRESSURE_TOILET@reddit

They just landed their LED driver a couple days ago:

https://www.phoronix.com/news/AMD-Halo-Box-RGB-LED-Driver

[-]

gh0stwriter1234@reddit

I'd rather it look like a cardboard box... and have features rather than cheap bling.

[-]

HIGH_PRESSURE_TOILET@reddit

DGX Spark doesn't have a power LED and has great features such as ConnectX. It also kinda looks like a sponge or steel wool for cleaning your pots and pans from the front. Perhaps that's what you want.

[-]