AMD's 192 GB Gorgon Halo prices might leave you petrified

The first systems powered by AMD's Gorgon Halo system-on-chip (SoC) platform have arrived, boasting up to 192 GB of unified memory on board. That's enough to put DeepSeek V4 Flash on your desk if you can afford to pay a hefty premium for all that RAM. GMKtec's EVO-X5 Pro is among the first to feature the House of Zen's top- …

[xyz123]

192GB of ram. we're halfway to meeting the minimum spec to run windows 12 then.

[EvaQ]

"Nvidia's RTX Spark is a family of premium Windows systems"

Wow ... Windows? I would suspect AI is heavily developed and used on Linux?

Because of the Nvidia's hatred of Linux?

Ah, elsewhere: "Focus on Windows: Nvidia's initial launch for the consumer RTX Spark line (featuring Blackwell GPUs and Arm CPU cores with unified memory) was heavily co-developed with Microsoft for Windows on Arm"

Nice ...

[Geoff Campbell]

Re: Linux

I run a bunch of local LLMs on Windows, using the Windows Subsystem for Linux. Best of both worlds - Linux for the AI stuff, Windows for my day to day boring office work.

GJC

[that_dude, dude]

Re: Linux

interesting.. similar to Lemonade server, I imagine.

now if we could get a linux subsystem for windows that works flawlessly...

[elsergiovolador]

192

I will never buy a machine with RAM that is not a whole number like 256, 512, 1024.

[Throatwarbler Mangrove]

Evil forces

The strix (according to Wikipedia) was a bird of ill omen that fed on human flesh and blood. The gorgons turned people to stone. I guess they'll have plenty of code names, but someone is definitely being cheeky here. I'm holding out for Succubus Halo myself.

[TReko]

Re: Evil forces

That GMtec machine linked boasts "99.6% stability". Perhaps the remaining 0.4% instability is due to the demons?

I'm not sure what is being measured but at that price I'd prefer 100% stability or at least five 9's.

[Groo The Wanderer - A Canuck]

192 GB of extremely high speed memory has never been a "cheap" commodity in the IT world. Sure prices have come down a lot relatively speaking, but performance has improved by almost exactly the same margin with each generation. We're just truly reaching the limits of what can be done at any price point with our current computing approaches, and will have to undergo some technology shifts to see any real gains, such as optical switching.

[FIA]

Sure prices have come down a lot relatively speaking,

Not really.

The M4/M3 Ultra Mac Studios could be upgraded to 512G memory for $4000. They were released in March 2025.

The latest Mac Studio (released Sept 2026) can be upgraded to 256G memory for.... $4000.

512G pricing hasn't been announced yet.

[doublelayer]

It's never been cheap, but it has, and quite recently, been a lot cheaper. It was always in the thousands*, but now it's a lot more thousands. No, performance compared to that time (2024) is not significantly better; it's the same LPDDR5X as then and the speed is only slightly higher. This isn't a surprise; massive demand based on infinite money, but not for me, will do that. Prices were falling before then, and they will resume falling eventually, but your claims are incorrect for the recent short term and will be incorrect for the next year or two.

* Thousands in whatever currency you're using, at least for most readers. Yen, rupees, won, and a few others excluded.

[Groo The Wanderer - A Canuck]

You're missing the key point that we've pretty much reached the limits of the technology when you gripe that performance "hasn't really been improving." For many decades, it did; those days are behind us.

[doublelayer]

I disagree that that's the reason. Performance hasn't increased because 2024 was not actually very long ago and anyway, we're comparing the same tech. LPDDR5X is what's used here, not because that's still the best we can do, but because it's still the generally cheaper option which works with existing chip tech, again the cheaper option. We have faster RAM using better manufacturing. HBM works great. It isn't available to me because it's too expensive because LLM companies want all of it, and anyway it wouldn't work with hardware I can afford, but it is better. I don't think we have sufficient evidence that we've hit a limit. I think improvements that have been made are limited to hardware that nobody other than LLM companies is really using.

[Fido]

Not for me

With non-upgradable memory soldered on board there is no way to buy an underprovisioned system and top it off in a year if prices go down. Minisforum announced a similar computer but with better networking.

As much as I like the guaranteed availability and security of locally-hosted inference, I'd prefer a GPU that has reasonable 64-bit floating point for when I don't want a chatbot.

For these reasons I don't feel bad that I can't afford one either.

[openjunk2]

Re: Not for me

Ah just go with the flow, need more ram? then buy the next model with 265GB and throw the stupid 192GB model away.

[Anonymous Coward]

You may need to avoid using the toaster

Any idea how much this beastie draws in power? You may end up with an on-prem powerhouse, but that powerhouse needs, well, power. And cooling.

Interesting device, though, and I suspect if you have the money to buy it the support infrastructure will not pose a challenge either.

[Steven Raith]

Re: You may need to avoid using the toaster

I mean, I'm pretty sure the Strix Halo is about 150-200w - they come with 300w PSUs but that's for USB-PD/Thunderbolt power delivery overhead. I'd expect Gorgon Halo versions would be a bit more (because it has more, slightly faster RAM) but not by much. It's a mobile CPU and a tricked up mobile GPU (effectively, just scaled up to near desktop size) that has a looooaad more memory available to it than usual, nothing fancier than that.

Mac Studio M5 Ultra is under 400w too, but for significantly more performance due to it having more than double the memory bandwidth, which is still the main bottleneck for inference, once you get the tokens ingested and it starts returning data. I've got an M1 Max Macbook Pro that I use for light AI testing and messing, and it barely draws 70w while being quite usable.

The SOC based systems really really don't pull that much power at all - and certainly not compared to a system running enough discrete GPUs to match the VRAM availability.

Regards

Steven R

[abufrejoval]

Sigh, you youngbloods simply can't fathom the sacrifices we made...

My first computer was an 1MHz Apple ][ clone with 64, ...wait for it..., Kilobytes of RAM: it set me back around 4000 German Mark in 1982 or 83, easily €8000 in today's money.

My second computer came only two years later, an IBM PC-AT clone with an 8-MHz 80286, 640KB of RAM, 20MB of spinning rust, color EGA, a matching monitor and a NEC P5 dot matrix printer for a total of 20.000 Mark, €40.000 today: it almost has a pair of 96GB RTX 6000 look downright cheap!

I was just starting out studying computer science, but freelancing as a programmer to pay for that stuff (as well as food/housing). The alternative would have been a nice new VW Golf GTI or a used Porsche, but unfortunately those couldn't be used to make more money.

Yes, those rising prices are really bad news, especially since it's getting tougher to make money with programming.

But boy, have things gotten cheap in the mean-time!

A nice chat with my divorce lawyer today costs me more than a brand new laptop, while I can ill afford either: the world is out of whack in so many directions, RAM prices are still a pretty modest issue in my book.

But then I have around 1.4TB of RAM in my various home-lab machines, 222 CPU cores, 88TB of SSD and 122TB of HDD storage (mobiles excluded).

And that's just the current left-overs, I've pushed stuff to family and friends for decades.

Actually, I completely forgot to even count VRAM, 117GB of mostly GDDR6 and 7, only a lone GTX 1080ti with 11GB GDDR5X being the odd one in the bunch.

I just don't have all that in a single machine nor does anything feature HBM class bandwidth so in terms of running LLMs I'm stuck with a single RTX 4090 on a 96GB Ryzen 9 host.

I know, I know, poor me, especially since I am now retired and have you youngbloods pay for me!

But all of that stuff had to be earned, it paid for itself via extra bucks for extra smarts.

If I still had to pay 1980's prices, perhaps I'd have graduated to an 80486 with a 32MB of RAM, a 1024x768 VGA and 500MB of disk space.

So don't complain to me, ye spoiled brats!

[retiredFool]

Re: Sigh, you youngbloods simply can't fathom the sacrifices we made...

Know what you mean. I think in 95 I bought a sparc, forget the name, used the little square box. Probably had 16M of memory and small disk for I think 3K. I added a 1G external drive(1k buck) for more space. I then later bought an ultra-sparc pizza box with developers discount for 13K. Then the PC style ultra sparc for 6K. Then later I started buying linux pc's whoa, what a price drop. Whole systems were like 500 bucks that ran circles around the sparc's. Then AMD came out with the x86-64 and picked up one of those for 5k. That really blew away the sparc's. It had a whopping 12G of memory and dual core. The last high dollar was a quad x86-64 with 16G of memory for I think 6K. That was in I think 2005. And then prices just dropped, and we were living in good times. I've not spent more than 600 or 700 on a system since then, although always upgrades, so like a motherboard/processor not a whole system. I think the main thing that has happened is in the world of computers, they actually got a little more expensive instead of always more for less. Me, I'm hoping for a return to the normal more for less cycle.

[Mostly Irrelevant]

$6,799 seems like a good deal, I think the data center boon might have broken my brain.