Fair point but you get that only for 252GB of ram so it’s not even “totally better” than a Mac Studio ultra with 512GB RAM. It still the old “smaller but faster than a Mac” stuff nvidia sells.Maybe in 1-2 years they will match the Mac 512 but then Apple will release an even bigger Mac
I don't know about "useless" (it seems quite useful to me) but I do feel mislead. It's unified memory in the same way that my current dGPU has unified memory. I guess nvlink-c2c probably (?) doesn't introduce a bottleneck but it's still two distinct arenas with very different performance characteristics.
I find it just misleading by advertising 700GB RAM as AI headline. I could plug a 32GB GPU to my 1.5TB ram server and call it “AI station with 1.5TB+ RAM” just so that you find it actually useless compared to the headlone
7.1TB/s of HBM is not "useless" -- that's 30 times the memory bandwidth of my DGX Spark -- and nobody is expecting such a machine to run "Opus 5" on its own. For such large models even datacentre GB300 NVL72 are multiple trays linked together via NVlink etc. This machine has QSFP ports and ConnectX for linking up for larger models.
It's a workstation, not a rack. It's for AI researchers. I'd love to have one on my desk.
My point is this is “useless” because you can’t run large models. It’s like having a very fast and expensive SSD with a very small capacity.
It’s not useless but it is for “real work”. Apple has been providing 512GB ram machines for several years so to say you need a rack to run a large model for personal usage it’s missing the point.
You needed a rack of nvidia cards to match an 128GB ram Mac as well a while ago so it’s more of the same.
People who buy Macs to run local inference are not the target audience for this machine.
It's an AI research workstation for people whose ultimate work goes on production GB300 NVL72 data centre racks (and for large models you link them together.)
These are for billionaires/multi millionaires/cryptobros/stupidly rich people who can afford such “toys”.
I’d really love to see a geek/tech freak/gadget lover/IT/homelab billionaire buying large number of things we mere mortals can’t afford, but only dream of, and then he blogs about them ;-)
Since I'm a "retro computing" guy, I look forward to buying one of these on eBay or at a Vintage Computing Fair for a few hundred bucks 5 years from now.
Yes, which planet are you on? Australia has recently lowered power prices due to renewables, probably other places will too for similar reasons as the rollout continues.
There's a lot more going on here because the host machine has a boatload of expensive LPDDR5x (which is stupidly expensive) that can also be used as unified (but slower) memory to the GPU, 72 ARM64 cores, stupid fast NVlink/QSFP networking, etc. etc.
Basically it's the same as a tray in a GB300 NVL72, but in workstation/desktop form. Niche would be AI researchers.
They will... not sell a lot of these. But what a beast.
You can't build such a machine out yourself but if you could I suspect the price point would come out about the same.
> You can't build such a machine out yourself but if you could I suspect the price point would come out about the same.
That's kind of the nature of capitalism and price elasticity - the manufacture price only limited the minimum sale price, and sale price usually reflects how much is the market willing to pay for the good.
Where you might save a lot of money is on building something that's very targeted to your needs that matches them better than a GB300 workstation, for a lower price.
But the point of such a machine generally is to have the equivalent of a GB300 NVL72 tray on your desk. It's so that you can do the work that belongs in a production DC eventually. Or at least that's now NVIDIA would like you to use it.
You could build out a complicated multi GPU setup of your own, but the work you do on inference tuning for your kernels etc would not necessarily translate to the real world.
But yes, if you just want to run GLM 5.3 on your own machine, that's a whole other story.
> Or at least that's now NVIDIA would like you to use it.
Exactly. It's tailor-built to be a GB300 tray under your desk. Unless you really need a GB300 tray under your desk, you would be better served looking elsewhere.
Yeah, all these overpriced toys will be obsolete soon anyway, but regular memory for PC for actually doing some productive stuff is still overpriced. Pure madness.
At the expected price point it would be nice if it came with a free pied-à-terre in Kensington (https://en.wikipedia.org/wiki/Kensington). Just a small studio apartment, nothing special.
I have a $50,000 Tektronix 'scope on my bench that's Secured with a kensington lock. I am in a small office in a building I own and I rent out some extra space to another 3 person engineering company. One weekend they tried to remove my 'scope from the bench because they had something "important" to work on. They quickly discovered the lock. (The whole thing was caught on camera.)
They weren't trying to steal it, they just wanted to bring it over to their side. But the lock did its job! This is the perfect situation for a Kensington lock.
I yelled at them the next day telling them if they need a Series 5 'scope they could go rent or buy one.
Agree, the CPU memory is about 400GB/s vs. Apple M5 Ultra at 1.2TB/s. But the M5 only has that, not the 7TB/s portion.
So this box could* still be faster when using the full 748GB if both RAM types can be used without the slower part bogging everything down.
*That comes down to implementation. Maybe somebody can chime in on this.
I'm just barely starting to wrap my head around mapping model sizes and quants to hardware components and constraints.
I don't get what this device is for.
I a million percent understand wanting 252GB-VRAM, that would get me a 280-320B model like glm-5.3 or deepseek-v4-flash, which would be a massive improvement over my gpt-oss:20b 16GB toy. I would gladly pay a grand for this, I would never pay ten grand for this, and it seems to be priced around a hundred grand.
So obviously the customer is commercial not consumer.
Can anyone planning a project around one of these at work share what their workload is shaped like and how they're modeling price/performance?
For instance, I don't get the 512GB of system memory, I'd gladly drop that to 128 to save money. Am I missing something about commercial workloads? Is a 1T paramater model at 20 tok/s more important to your workload than a 300B one at 60? Is it simply a co-dependency of not the model but the other software you're running on the machine thats using/interacting-with/being-driven-by the model?
Whats your napkin math to justify 100k? Actually thats not even really the question, its more like - whats your napkin math to determine between the "dual linked GB10" use case vs this product's use case vs an 8U supermicro with 4 cards use case.
It's actually kind of nostalgic to see off the shelf computers have such a sky high price again. Back in the Silicon Graphics days the prices were aspirational! Give me something to lust over again... makes me feel young.
52 comments
[ 0.24 ms ] story [ 19.4 ms ] threadSo it has only 256GB of actual ”AI” memory making it “useless”/toy for actual real world AI workloads(I.e it can’t replace something like opus 5)
It's a workstation, not a rack. It's for AI researchers. I'd love to have one on my desk.
What even is this comment?
It’s not useless but it is for “real work”. Apple has been providing 512GB ram machines for several years so to say you need a rack to run a large model for personal usage it’s missing the point.
You needed a rack of nvidia cards to match an 128GB ram Mac as well a while ago so it’s more of the same.
People who buy Macs to run local inference are not the target audience for this machine.
It's an AI research workstation for people whose ultimate work goes on production GB300 NVL72 data centre racks (and for large models you link them together.)
What is it with those stupid names?
No price, so of course this is not for the smelly working class.
It just doesn’t make sense to buy this stuff at peak prices.
Basically it's the same as a tray in a GB300 NVL72, but in workstation/desktop form. Niche would be AI researchers.
They will... not sell a lot of these. But what a beast.
You can't build such a machine out yourself but if you could I suspect the price point would come out about the same.
That's kind of the nature of capitalism and price elasticity - the manufacture price only limited the minimum sale price, and sale price usually reflects how much is the market willing to pay for the good.
Where you might save a lot of money is on building something that's very targeted to your needs that matches them better than a GB300 workstation, for a lower price.
You could build out a complicated multi GPU setup of your own, but the work you do on inference tuning for your kernels etc would not necessarily translate to the real world.
But yes, if you just want to run GLM 5.3 on your own machine, that's a whole other story.
Exactly. It's tailor-built to be a GB300 tray under your desk. Unless you really need a GB300 tray under your desk, you would be better served looking elsewhere.
I think you might need a bit more than that at this price point ...
One listing is 410K for 280 sq/ft coming out at £1464 per square foot, almost exactly 10x the price per sq/ft we paid for our house a few years ago.
So $100K would be £74K which would get you ~50 sq/ft.
They weren't trying to steal it, they just wanted to bring it over to their side. But the lock did its job! This is the perfect situation for a Kensington lock.
I yelled at them the next day telling them if they need a Series 5 'scope they could go rent or buy one.
https://www.gigabyte.com/in/Enterprise/Tower-Server/W775-V10...
https://www.msi.com/Landing/NVIDIA-DGX-STATION
I don't get what this device is for.
I a million percent understand wanting 252GB-VRAM, that would get me a 280-320B model like glm-5.3 or deepseek-v4-flash, which would be a massive improvement over my gpt-oss:20b 16GB toy. I would gladly pay a grand for this, I would never pay ten grand for this, and it seems to be priced around a hundred grand.
So obviously the customer is commercial not consumer.
Can anyone planning a project around one of these at work share what their workload is shaped like and how they're modeling price/performance?
For instance, I don't get the 512GB of system memory, I'd gladly drop that to 128 to save money. Am I missing something about commercial workloads? Is a 1T paramater model at 20 tok/s more important to your workload than a 300B one at 60? Is it simply a co-dependency of not the model but the other software you're running on the machine thats using/interacting-with/being-driven-by the model?
Whats your napkin math to justify 100k? Actually thats not even really the question, its more like - whats your napkin math to determine between the "dual linked GB10" use case vs this product's use case vs an 8U supermicro with 4 cards use case.