Hopefully... If it doesn't, mere mortals like us consumers are going to be priced out of computers altogether. It'll be the end of the personal computing era and a return to big iron mainframes.
Which one is the fastest and how fast is it? Even the "fast" ones doesn't seem to even reach close to consumer NVIDIA GPUs released years ago when it comes to prompt processing.
indeed. make up your mind, and also "slow/fast" mean very little given plethora of models to choose from. fast for what, slow for what. fast with which harness, etc...
But more memory is available so a larger model can be loaded. It depends on the use case which is better. Any chat or voice model will have better UX with nvidia but document or code generation will be better with apple.
I would assume an H100 selling for less than $30k is a scam. I see one for 11,329€, and one for $25k but most are over $30k and I don't know how to judge these random storefronts. You've bought an H100 for $20k that was in working order recently?
Hopper is ~4 years old at this point though, compared to Blackwell which is ~2 years old, the difference isn't nil.
Depending on your use case, you might prefer native FP4 and FP6 low-precision support and 5th-generation Tensor Cores rather than what Hopper offers.
I think for training Hopper makes sense as it's generally a bit cheaper and the difference isn't that big, but for inference the difference widens a bunch makes a lot more sense to go with Blackwell.
You're correct, the difference isn't nil - H100 is data-center grade GPU with loads of bandwidth and compute while RTX PRO is a consumer grade GPU. The difference between the microarchitectures would make sense to point out if we had been comparing apples to apples and not apples to pits.
Kind of feels like if you spend $40K on your GPU setup, you don't want to cheap out on the PSU.
Still, you're right, I have a 1200W PSU for a single RTX Pro 6000, cost ~500 something, so still the PSU shouldn't be that large part of the total budget.
Funny thing is, you don't really want to just plug that into a wall socket. You want an uninterruptible power supply that can keep the machine running for the 5 to 8 minutes you need to get to it and shut it down safely during a power cut.
But you'll find most household power sockets aren't actually rated for the kind of power these things draw at peak. If you go for a dual GPU you're going to want to sustain a higher current. Think washer / drier / fridge kind of hookups for the UPS and that rig.
You spent $42K on the rig, you're also going to need to get an electrician in and have the electrical hookup seen to. The single 4090 I'm running put me right up to that edge where I could use a standard electrical hookup with a 1000w UPS.
So another $1K for the UPS, and $500 worth of electrical work (or more if they have to up the switch to 200 amps, for example).
$42k and a toy CPU at that. Threadripper or EPIC if it must be AMD camp, although I'd rather it weren't (_always_ some random issues on linux). TBH, if I were buying workstation at such prices, I'd probably first take a look at what HP has (since their cooling and immunity to dust is unprecedented), and then probably BOXX and Puget. For DIY there's always supermicro at such levels.
I've noticed that some companies have flat out dropped the high memory configurations from their lineup that were offered just months ago at inflated prices. OnePlayerX had a 128GB Strix Halo tablet, RedMagic had a 24GB Android phone - both have been removed from their sites. It's a similar situation for the new Thinkpad lineup.
That memory speed drop going from single DIMM per channel DDR5 to dual DIMM per channel is very much notable - 3600MT/s is barely any faster than what DDR4 could do in it's final days.
The CPU market seems to be missing the "HEDT" platform we used to enjoy.
Seems they might be selling a combo of dual RTX Pro 6000 workstation cards (non-refundable), but the setup they have, unless they change how the GPUs cooling work, isn't gonna work out thermally if you stack two of those workstation cards on top of each other. But it doesn't say WS or "Workstation", so I suppose could be the server variant too? But also doesn't say...
Hope they have good overall cooling the case, cooling doesn't seem to be mentioned, it'll be a noisy little machine no doubt :)
Thelio Mira AI
$3,299.00
Description
Specs
Warranty
Accelerate your AI development with Thelio Mira AI, System76's affordable, GPU-focused workstations, built for local AI development — so you can train, fine-tune, and iterate challenging AI workloads entirely on your own hardware.
Configure Thelio Mira with up to:
16-core AMD Ryzen 9000 Series CPU
192 GB DDR5 RAM
Dual NVIDIA RTX Pro 6000 GPU
192 GB GPU memory
this is miss leading. actually it is 4gb ram a400 price.
When I saw the 192GB I first thought they might have beat Framework to market with a Gorgon Halo machine with unified RAM that Framework has teased. This is just a GPU workstation.
I'd really like to just buy some collective shares of a GB300 NVL72 in a datacenter somewhere and have a daily token quota on a shared DeepSeek model or whatever was hot that week. I would just need to find ~100 like-minded folks at this $40k per share price to get started haha.
I've thought a little bit about this, really I suspect you'd actually want maybe 20 folks for about a 20k buy-in and a dedicated 8×B300. You could get some pretty sweet token/s. Probably a community/coop share structure. It'd be better than the GB300 which is fairly throughput optimized (not needed for 20ish folk)
People routinely buy cars that cost double that. This is a machine that people that typically earn six figure salaries use to do their work on. Exactly the target demographic that would buy/lease expensive cars as well. And then those get used mainly to commute to work.
I don't really need one. But I do have a 4.5K mac book pro for work. Yes it's a bit overpowered. But it's the main vehicle with which I earn a living and it routinely saves me time by blazing through builds and generally doing things quickly for me. I have slightly more GPU on that thing than I need. But it's nice to have the option to experiment with some of the open weight AI models.
> People routinely buy cars that cost double that.
Maybe it's the "routinely" or the "people" part, but "people" generally don't ever buy a 80K car, it's a small fraction of the population who can afford such luxury cars. And I'm sure the ones who buy cars for 80K, unless they're car enthusiasts or just plain billionaires, don't do so "routinely" either.
Are you sure about that? For the price of the top configuration here you can buy a 5-stack of 256GB Mac Studio M5 Ultras.
This dual NVIDIA RTX PRO 6000 setup has 192GB of VRAM at 1.8 TB/s, versus the 5-stack of Macs that collectively have 1.25 TB at 1.2 TB/second... I'm not convinced that equation comes out in Nvidia's favour.
Right, performance is more than just the aggregated memory speed of the hardware you have. I'm fairly sure, at least last time I looked, maybe Apple launched something new in the last 2-3 months that has completely changed the picture?
> performance is more than just the aggregated memory speed of the hardware you have
In other fields, sure, but for big LLMs it's a very significant part of the performance picture. That stack of Macs also is going to be able to natively run models 4-5x larger than the dual Blackwells can hold in memory - any model over about 128GB of weights isn't going to fit on the GPUs, and is going to be heavily performance constrained by moving data across the PCIE bus.
> maybe Apple launched something new in the last 2-3 months that has completely changed the picture
Indeed. The M5 Ultra (currently up for pre-order) has 50% higher memory bandwidth than its predecessor, and a claimed 4x improvement in prompt prefill.
> any model over about 128GB of weights isn't going to fit on the GPUs
Not sure if you misunderstand what GPUs we're talking about, one RTX Pro 6000 has 96GB of VRAM.
> Indeed. The M5 Ultra (currently up for pre-order) has 50% higher memory bandwidth than its predecessor, and a claimed 4x improvement in prompt prefill.
Exciting! Eagerly awaiting the benchmarks and comparisons then. Lets hope "4x improvement" had a good enough baseline so 4x actually ends up useful in practice compared to the current hardware they offer.
> 50% higher memory bandwidth
Seems this lands on ~1.2 TB/s if what Apple claims is correct. For reference, RTX Pro 6000 does 1.8 TB/s, so seems Apple is indeed getting closer incrementally.
Yeah, that's why the "any model over about 128GB of weights isn't going to fit on the GPUs" part doesn't make sense, you'll easily be able to run weights in that weight class on two of them. Or did I misunderstand what "on the GPUs" you meant?
Right, easily run 128-160 GB models, yes. Not easily run models much larger than that. Anything that doesn't fit in the 192 GB (including context, cache, etc) is going to have to be sparse/MoE, and the rather anaemic PCIE bandwidth is going to hurt.
By comparison, the 5x Mac cluster should be able to run a dense ~800GB-1TB model without a drastic slowdown.
Sold my computer + DLSR camera back in the day to afford moving countries. Can't imagine what life would have been if I didn't, very happy I did. Lived with a netbook for 2-3 years, but was before I was a professional programmer.
I absolutely love system76. Never owned one of their desktop machines, but I have a couple of their older laptops and they work fantastically for me. As in I can easily get a decade out of them and replace the batteries every few years. One of them is almost a decade old ironically enough.
Anyone know why they chose a consumer grade CPU on this? I’m a little surprised to see a 9950X as the top option. There’s not enough PCIe lanes available to run everything without bifurcation. I imagine the GPUs probably aren’t too bothered generally… but the NVMe drives are likely to slow down as a result.
A Threadripper ain’t cheap, but if you’re spending $45k on a workstation it seems like a weird place to skimp. Not to mention you’d then have the option of shoveling a few more 6000 pros in it when you want to run something larger (assuming your home/office electrical box can support it).
The GPUs might care quite a bit about bifurcation. The RTX Pro 6000s can communicate directly between themselves over PCIe 5.0 16x and saturate the bus doing so. This is used when doing tensor parallel LLM execution for the All-Reduce steps.
Someone can work around this by including a PCIe 5.0 switch like the Broadcom PEX89048. That switch can be used to, for instance, have the two GPUs talk to each other at full PCIe 5.0 16x speeds, while talking to the motherboard at say PCIe 4.0 16x, or even 8x. As an end user, cards using that switch seem to be an AliExpress only item, but I assume an OEM like System76 could get their hands on it to place directly on the motherboard. I imagine the price of adding that to a motherboard could be less than the price of switching to a Threadripper or Xeon 600 platform.
However, I can't find any evidence that System76 has added a switch, and in fact they refer to x8/x8 multi-GPU support, making me think they didn't add a switch. To my mind, this isn't a very satisfactory setup.
Aye but at $45k you’re so deep already is that really worth the savings? It’s so much easier to manage a single system, and the performance will be in absolute terms better.
Truly, I get your point, but $45k is obscene for a workstation (because of the RAM shortage). I can’t fathom it being a reasonable decision[1] unless you’re able to churn an immense profit from it. And… if you can, it’s an expense that saves time, energy, and effort. If it is only profitable when $8k in price difference makes it viable, is that really worth the effort? An $8k increase making a difference, when you’re wholly dependent on the frontier (or near frontier) lab not releasing a model that makes your work(station) irrelevant— seems reckless. Couple that with the doubling of DeepSeek’s parameters for their flash model… and well the math just don’t math for my naive brain.
[1] I’m absolutely incapable of spending that much on a workstation so my opinions may be irrelevant… but I cannot understand the “stepping over quarters to save a penny” mentality[2].
[2] I could be missing the forest for the trees. As a result, I’d love to know how I am being shortsighted. I just can’t fathom a situation where an $8k surcharge in system RAM doesn’t make sense. It’s not about system RAM, it’s about GPU RAM, margins, and useful lifetime of the system.
This is a link to their configurable workstation on consumer boards. They offer TR or Xeon based boards too; it's just a different link/SKU: https://system76.com/desktops/thelio-major
> An $8k increase making a difference, when you’re wholly dependent on the frontier (or near frontier) lab not releasing a model that makes your work(station) irrelevant
Hmm, why? I don't understand why better models being available remotely would change the usefulness of what you already have working locally? If it works for you, it works, regardless of what the frontier labs have, and if it doesn't, it doesn't, again regardless.
I remember losing sleep when I bought my RTX Pro 6000, but somehow they keep going up in price, and waddya know I’ve even done real work that appreciates the VRAM size.
Same. Had to convince the wife, initially aimed to get two of them, ended up with one. Now my wife is berating me for not convincing her that I should have bought four and then sell two later... Can't win :)
Haha, bloody hell, mate. $17k a pop? I bought these for $8k-ish. That's bonkers to buy this when newer flash models won't fit on this any more. DeepSeek V4 Flash seems the last in its line, and then you have to count on Qwen Next. Seems like a waste honestly.
I find it hard to believe any enterprise willing to drop over $40k on a workstation PC would choose `Pop!_OS 24.04 LTS with the COSMIC Desktop Environment` over Ubuntu.
Why? It is built from Ubuntu[1]. I am a very happy full-time daily user. Anecdotally it is more stable and productive than Windows ever was for me, and plenty of enterprises use windows…
The prefill rates are way faster on these cards compared to apple silicon, which affects TTFT and therefore usability for a lot of coding tasks, unless you fire and forget most of the time.
Mac GPUs supposedly being fastest thing on Earth obsoleting everything NVIDIA is just pure marketing. It was just faster once than some midrange laptop NVIDIA, which conveniently wasn't explicitly marked as different chip sharing the branding with the desktop variant(oof).
The real benefits to going Mac is its quiet and discreet, high wife/CEO acceptance design, and their huge GPU-assignable shared RAM.
If you're okay with your desk being the wing top of a flying airplane, and 3M Peltor or David-Clark is your favorite working time headphone brand anyway, then your build will be faster and cheaper than a maxed out Mac Studio.
For me the operating system (and ecosystem in general) is one of the biggest reasons I use a Mac, but the unified memory pool is a benefit too. I am still distraught about the gaming situation though.
I wonder why Apple has not really pushed their memory bandwidth numbers. They're starting to catch up to 2020 at this point -- cool, but still struggling to run years-old models. (FLOPS is even further behind by a year or two)
Maybe they're waiting on HBM? Given that they're skipping M6 Max to go straight for M7, I'd assume they're doing a redesign for it.
GDDR has higher latency and much lower capacity. HBM requires an interposer and also has lower capacity. It's not an easy choice. The M7 family will probably double memory bandwidth using LPDDR6.
There’s roughly 50% more memory bandwidth on an RTX 6000 pro versus the M5 Ultra. I imagine it really comes down to what you’re doing, I know my 5090 runs laps around the M5 Pro I have when it comes to local LLM performance (until it runs out of memory).
If it wasn’t for the fact the 6000 Pro is selling for like 100% ($9,000) over MSRP it would probably look a lot more reasonable.
If your goal is pure GPU compute, you should have got a 5090. It can bench comparable to the RTX 6000 Pro cards and is near guaranteed to outperform M5 Ultra.
Is this just a marketing stunt? From a completely ignorant user, doesn't AI reeuqires high bandwidth ram (like the GPU)? Is really DDR5’s bandwidth enough?
I thought it was a pricing mistake. But no, you just have to select the variations. $42,412.00
They missed an opportunity by not adding $8
They will add $27,008 in time.
I had to do a double take myself, scrolled down and arrived at a similar number.
Wish I built this a few years ago!
Yeah, I bought a Thelio Major three years ago. I'm kicking myself for not going dual-4090s for a mere $10K back then.
Hopefully it will come down to that price point or cheaper again in the next 3-4 years?
Hopefully... If it doesn't, mere mortals like us consumers are going to be priced out of computers altogether. It'll be the end of the personal computing era and a return to big iron mainframes.
"All of this has happened before, and will happen again."
Me three
And people still don't believe Apple is giving a good deal on Macs.
Prompt processing on Macs is VERY slow...
which engine? cause there are dozens already, and some are quite fast.
Which one is the fastest and how fast is it? Even the "fast" ones doesn't seem to even reach close to consumer NVIDIA GPUs released years ago when it comes to prompt processing.
large model and slow, or speed and a teeny model, make your choice
Or you get N of RTX Pro 6000 and get large and fast models :) Make your choice, and do it before the prices go up even more.
indeed. make up your mind, and also "slow/fast" mean very little given plethora of models to choose from. fast for what, slow for what. fast with which harness, etc...
well yes if money is no object, why not
none of them are Nvidia GPU FAST
But more memory is available so a larger model can be loaded. It depends on the use case which is better. Any chat or voice model will have better UX with nvidia but document or code generation will be better with apple.
> Prompt processing on Macs is VERY slow...
That sounds like a contradiction.
/J
I wonder what the target audience for this price range is? At that price point I would rather go for a single H100.
I would rather take this option than a single H100:
> 96 GB Dual NVIDIA RTX PRO 5000 w/ 1000W PSU (non-refundable) +$19,169
Though I don't think there's anything stopping you from trying to stuff an H100 into this machine if you want to BYO.
For that price you get the H100. The machine is overpriced, and I'd personally rather go for a second-hand workstation with server-grade components.
I would assume an H100 selling for less than $30k is a scam. I see one for 11,329€, and one for $25k but most are over $30k and I don't know how to judge these random storefronts. You've bought an H100 for $20k that was in working order recently?
What kind of comment is that? No, I have not bought the h100 obviously, and, yes, you can buy the new one for 20k. Used one for even less.
Hopper is ~4 years old at this point though, compared to Blackwell which is ~2 years old, the difference isn't nil.
Depending on your use case, you might prefer native FP4 and FP6 low-precision support and 5th-generation Tensor Cores rather than what Hopper offers.
I think for training Hopper makes sense as it's generally a bit cheaper and the difference isn't that big, but for inference the difference widens a bunch makes a lot more sense to go with Blackwell.
You're correct, the difference isn't nil - H100 is data-center grade GPU with loads of bandwidth and compute while RTX PRO is a consumer grade GPU. The difference between the microarchitectures would make sense to point out if we had been comparing apples to apples and not apples to pits.
So a new car, or an AI rig.
That's a bitter pill to swallow. Most of this is the cost of the Nvidia card, and the RAM.
Really just the GPUs. PSUs aren’t that expensive maybe 100-200 for a good 750W-1000W one
Kind of feels like if you spend $40K on your GPU setup, you don't want to cheap out on the PSU.
Still, you're right, I have a 1200W PSU for a single RTX Pro 6000, cost ~500 something, so still the PSU shouldn't be that large part of the total budget.
Funny thing is, you don't really want to just plug that into a wall socket. You want an uninterruptible power supply that can keep the machine running for the 5 to 8 minutes you need to get to it and shut it down safely during a power cut.
But you'll find most household power sockets aren't actually rated for the kind of power these things draw at peak. If you go for a dual GPU you're going to want to sustain a higher current. Think washer / drier / fridge kind of hookups for the UPS and that rig.
You spent $42K on the rig, you're also going to need to get an electrician in and have the electrical hookup seen to. The single 4090 I'm running put me right up to that edge where I could use a standard electrical hookup with a 1000w UPS.
So another $1K for the UPS, and $500 worth of electrical work (or more if they have to up the switch to 200 amps, for example).
Not for the faint of wallet.
Kind of feels like if you're noticing $1K here and $500 there, maybe $42K is too much to start with :)
$42k and a toy CPU at that. Threadripper or EPIC if it must be AMD camp, although I'd rather it weren't (_always_ some random issues on linux). TBH, if I were buying workstation at such prices, I'd probably first take a look at what HP has (since their cooling and immunity to dust is unprecedented), and then probably BOXX and Puget. For DIY there's always supermicro at such levels.
That's what I don't understand, they're shipping RTX 6000s that are PCI lane-starved. Buying a Maserati and driving it around in 3rd gear.
it's probably a fantasy setup, not built nor tested. Or worse, someone there things x8/x8 vs x16 GPUs are 'good enough' for this setup at that price.
I've noticed that some companies have flat out dropped the high memory configurations from their lineup that were offered just months ago at inflated prices. OnePlayerX had a 128GB Strix Halo tablet, RedMagic had a 24GB Android phone - both have been removed from their sites. It's a similar situation for the new Thinkpad lineup.
The honor 400 used to be sold for 300 dollars with 12 gb of ram. Now the replacement the honor 600 comes with 8 GB at twice the price.
SSD too. The priced have tripled...
It would be great if China had more lithography machine....
All memory. SSDs are expensive, microSD is expensive, even spinning rust has inflated.
hoist by our own strategic export controls, rip
Taiwan would disagree
That memory speed drop going from single DIMM per channel DDR5 to dual DIMM per channel is very much notable - 3600MT/s is barely any faster than what DDR4 could do in it's final days.
The CPU market seems to be missing the "HEDT" platform we used to enjoy.
Threadripper (4 or 8 memory channels)
Seems they might be selling a combo of dual RTX Pro 6000 workstation cards (non-refundable), but the setup they have, unless they change how the GPUs cooling work, isn't gonna work out thermally if you stack two of those workstation cards on top of each other. But it doesn't say WS or "Workstation", so I suppose could be the server variant too? But also doesn't say...
Hope they have good overall cooling the case, cooling doesn't seem to be mentioned, it'll be a noisy little machine no doubt :)
They're [Max-Q] blower cards and there's a big gap between them so the cooling looks fine.
So the two variants they offer is the Max-Q and the Server edition one, but only explicitly mentioned for one?
Edit: The Max-Q is one of the options, what's the other option then?
Seeing as it adds a second PSU, I'm guessing the workstation one.
That's bananas, unless they also water cool them, would easily overheat if it's just two bog standard workstation cards on top of each other.
Are they possibly selling setups they haven't actually tested practically in the real world?
Page mentions liquid cooling, for what it's worth.
> Accelerate your AI development with Thelio Mira AI, System76's affordable, GPU-focused workstations
From $3,299 (with just 64 GB RAM and a 4 GB GPU) to over $50,000. Very affordable indeed.
50K without a proper CPU, mind.
Insane
Thelio Mira AI $3,299.00 Description Specs Warranty Accelerate your AI development with Thelio Mira AI, System76's affordable, GPU-focused workstations, built for local AI development — so you can train, fine-tune, and iterate challenging AI workloads entirely on your own hardware.
Configure Thelio Mira with up to:
this is miss leading. actually it is 4gb ram a400 price.dual RTX6000 price is : $40,538.00
When I saw the 192GB I first thought they might have beat Framework to market with a Gorgon Halo machine with unified RAM that Framework has teased. This is just a GPU workstation.
I'd really like to just buy some collective shares of a GB300 NVL72 in a datacenter somewhere and have a daily token quota on a shared DeepSeek model or whatever was hot that week. I would just need to find ~100 like-minded folks at this $40k per share price to get started haha.
I've thought a little bit about this, really I suspect you'd actually want maybe 20 folks for about a 20k buy-in and a dedicated 8×B300. You could get some pretty sweet token/s. Probably a community/coop share structure. It'd be better than the GB300 which is fairly throughput optimized (not needed for 20ish folk)
Can't wait for in 10 years time to find one of these workstations on eBay selling for £100-150
Then I will do a video of me playing Crysis on it
Slightly related, Crysis running on a RTX Pro 6000 via Proton :) https://www.youtube.com/watch?v=M5XDaO6b0uY
It only costs $40,000.
People routinely buy cars that cost double that. This is a machine that people that typically earn six figure salaries use to do their work on. Exactly the target demographic that would buy/lease expensive cars as well. And then those get used mainly to commute to work.
I don't really need one. But I do have a 4.5K mac book pro for work. Yes it's a bit overpowered. But it's the main vehicle with which I earn a living and it routinely saves me time by blazing through builds and generally doing things quickly for me. I have slightly more GPU on that thing than I need. But it's nice to have the option to experiment with some of the open weight AI models.
> People routinely buy cars that cost double that.
Maybe it's the "routinely" or the "people" part, but "people" generally don't ever buy a 80K car, it's a small fraction of the population who can afford such luxury cars. And I'm sure the ones who buy cars for 80K, unless they're car enthusiasts or just plain billionaires, don't do so "routinely" either.
And you would buy it through your company so it won’t be quite as much net. Surely no one is buying these just for fun
> Surely no one is buying these just for fun
-_-
Some of us worked hard in life and have spare money, sue me!
the landing page showed "$3,299.00" and I had my candy store moment for a few secs.
5 years ago I could not afford top shelf coders. Now I cannot afford top shelf machines.
$37,000 in GPUs. Man, I never expected another PC manufacturer to make Apple's top configuration look cheap by comparison.
Calculate what performance you get per $ spent, and Apple again looks the most expensive out of probably anything else you could buy.
Are you sure about that? For the price of the top configuration here you can buy a 5-stack of 256GB Mac Studio M5 Ultras.
This dual NVIDIA RTX PRO 6000 setup has 192GB of VRAM at 1.8 TB/s, versus the 5-stack of Macs that collectively have 1.25 TB at 1.2 TB/second... I'm not convinced that equation comes out in Nvidia's favour.
Right, performance is more than just the aggregated memory speed of the hardware you have. I'm fairly sure, at least last time I looked, maybe Apple launched something new in the last 2-3 months that has completely changed the picture?
> performance is more than just the aggregated memory speed of the hardware you have
In other fields, sure, but for big LLMs it's a very significant part of the performance picture. That stack of Macs also is going to be able to natively run models 4-5x larger than the dual Blackwells can hold in memory - any model over about 128GB of weights isn't going to fit on the GPUs, and is going to be heavily performance constrained by moving data across the PCIE bus.
> maybe Apple launched something new in the last 2-3 months that has completely changed the picture
Indeed. The M5 Ultra (currently up for pre-order) has 50% higher memory bandwidth than its predecessor, and a claimed 4x improvement in prompt prefill.
> any model over about 128GB of weights isn't going to fit on the GPUs
Not sure if you misunderstand what GPUs we're talking about, one RTX Pro 6000 has 96GB of VRAM.
> Indeed. The M5 Ultra (currently up for pre-order) has 50% higher memory bandwidth than its predecessor, and a claimed 4x improvement in prompt prefill.
Exciting! Eagerly awaiting the benchmarks and comparisons then. Lets hope "4x improvement" had a good enough baseline so 4x actually ends up useful in practice compared to the current hardware they offer.
> 50% higher memory bandwidth
Seems this lands on ~1.2 TB/s if what Apple claims is correct. For reference, RTX Pro 6000 does 1.8 TB/s, so seems Apple is indeed getting closer incrementally.
> Not sure if you misunderstand what GPUs we're talking about, one RTX Pro 6000 has 96GB of VRAM.
Right, but the top option here is a pair of RTX Pro 6000s, hence 192 GB of VRAM in total. Should be enough for a 128GB model plus context, cache, etc.
Yeah, that's why the "any model over about 128GB of weights isn't going to fit on the GPUs" part doesn't make sense, you'll easily be able to run weights in that weight class on two of them. Or did I misunderstand what "on the GPUs" you meant?
Right, easily run 128-160 GB models, yes. Not easily run models much larger than that. Anything that doesn't fit in the 192 GB (including context, cache, etc) is going to have to be sparse/MoE, and the rather anaemic PCIE bandwidth is going to hurt.
By comparison, the 5x Mac cluster should be able to run a dense ~800GB-1TB model without a drastic slowdown.
What motherboard do they use? (Or do they design their own now?)
The ASUS logo is visible in the photo of the innards.
I'm guessing it'll depend on the GPU (and CPU) choice, putting the cheapest that is sufficient for the choice.
New car or a Thelio. Tough decision.
Second-hand car is more rational than new car.
I bought a spark instead of a motorcycle so it’s somewhat relative and it’s also somewhat relative.
There was a time when I sold my computer and bought a motorcycle.
Can honestly say that was one of the best trades I’ve ever done. I didn’t buy another computer for 3 more years.
Sold my computer + DLSR camera back in the day to afford moving countries. Can't imagine what life would have been if I didn't, very happy I did. Lived with a netbook for 2-3 years, but was before I was a professional programmer.
I absolutely love system76. Never owned one of their desktop machines, but I have a couple of their older laptops and they work fantastically for me. As in I can easily get a decade out of them and replace the batteries every few years. One of them is almost a decade old ironically enough.
This makes Mac Studio with 256GB memory look cheap in comparison. More memory to store weight at a quarter of the price!
It's gonna have WAY faster token rates than that Mac Studio though, enough to be a qualitative rather than quantitative difference.
At that price, it's competing against a cluster of M5 Mac Studios, which we don't have test data for yet.
Can't it be both?
EG As I gazed at Janice, joy with equal part contentment washed through my soul.
Nobody seems to have noticed that they dropped their ampere-based thelio.
Anyone know why they chose a consumer grade CPU on this? I’m a little surprised to see a 9950X as the top option. There’s not enough PCIe lanes available to run everything without bifurcation. I imagine the GPUs probably aren’t too bothered generally… but the NVMe drives are likely to slow down as a result.
A Threadripper ain’t cheap, but if you’re spending $45k on a workstation it seems like a weird place to skimp. Not to mention you’d then have the option of shoveling a few more 6000 pros in it when you want to run something larger (assuming your home/office electrical box can support it).
The GPUs might care quite a bit about bifurcation. The RTX Pro 6000s can communicate directly between themselves over PCIe 5.0 16x and saturate the bus doing so. This is used when doing tensor parallel LLM execution for the All-Reduce steps.
Someone can work around this by including a PCIe 5.0 switch like the Broadcom PEX89048. That switch can be used to, for instance, have the two GPUs talk to each other at full PCIe 5.0 16x speeds, while talking to the motherboard at say PCIe 4.0 16x, or even 8x. As an end user, cards using that switch seem to be an AliExpress only item, but I assume an OEM like System76 could get their hands on it to place directly on the motherboard. I imagine the price of adding that to a motherboard could be less than the price of switching to a Threadripper or Xeon 600 platform.
However, I can't find any evidence that System76 has added a switch, and in fact they refer to x8/x8 multi-GPU support, making me think they didn't add a switch. To my mind, this isn't a very satisfactory setup.
You need far more absurdly expensive RAM (RDIMM/ECC?) for a threadripper. Like 4x more expensive.
Aye but at $45k you’re so deep already is that really worth the savings? It’s so much easier to manage a single system, and the performance will be in absolute terms better.
Truly, I get your point, but $45k is obscene for a workstation (because of the RAM shortage). I can’t fathom it being a reasonable decision[1] unless you’re able to churn an immense profit from it. And… if you can, it’s an expense that saves time, energy, and effort. If it is only profitable when $8k in price difference makes it viable, is that really worth the effort? An $8k increase making a difference, when you’re wholly dependent on the frontier (or near frontier) lab not releasing a model that makes your work(station) irrelevant— seems reckless. Couple that with the doubling of DeepSeek’s parameters for their flash model… and well the math just don’t math for my naive brain.
[1] I’m absolutely incapable of spending that much on a workstation so my opinions may be irrelevant… but I cannot understand the “stepping over quarters to save a penny” mentality[2].
[2] I could be missing the forest for the trees. As a result, I’d love to know how I am being shortsighted. I just can’t fathom a situation where an $8k surcharge in system RAM doesn’t make sense. It’s not about system RAM, it’s about GPU RAM, margins, and useful lifetime of the system.
This is a link to their configurable workstation on consumer boards. They offer TR or Xeon based boards too; it's just a different link/SKU: https://system76.com/desktops/thelio-major
So they offer both options.
> An $8k increase making a difference, when you’re wholly dependent on the frontier (or near frontier) lab not releasing a model that makes your work(station) irrelevant
Hmm, why? I don't understand why better models being available remotely would change the usefulness of what you already have working locally? If it works for you, it works, regardless of what the frontier labs have, and if it doesn't, it doesn't, again regardless.
I remember losing sleep when I bought my RTX Pro 6000, but somehow they keep going up in price, and waddya know I’ve even done real work that appreciates the VRAM size.
Same. Had to convince the wife, initially aimed to get two of them, ended up with one. Now my wife is berating me for not convincing her that I should have bought four and then sell two later... Can't win :)
Speaking of AMD CPU’s, how is Strix Halo coming along?
As that’s unified memory à la Apple and I think up to 128gb
Haha, bloody hell, mate. $17k a pop? I bought these for $8k-ish. That's bonkers to buy this when newer flash models won't fit on this any more. DeepSeek V4 Flash seems the last in its line, and then you have to count on Qwen Next. Seems like a waste honestly.
k is single H100 money. The workstation form factor only makes sense if you're locked out of cloud GPU allocation.
Damn, what a misleading price.
I find it hard to believe any enterprise willing to drop over $40k on a workstation PC would choose `Pop!_OS 24.04 LTS with the COSMIC Desktop Environment` over Ubuntu.
Why? It is built from Ubuntu[1]. I am a very happy full-time daily user. Anecdotally it is more stable and productive than Windows ever was for me, and plenty of enterprises use windows…
[1]https://system76.com/support/difference-between-pop-ubuntu
Is it more powerful than Apple silicon, or why would anyone buy this? Just for Linux?
Oh, it's System76, that's why. Open hardware. Makes sense.
Edit: Also Nvidia
The prefill rates are way faster on these cards compared to apple silicon, which affects TTFT and therefore usability for a lot of coding tasks, unless you fire and forget most of the time.
Mac GPUs supposedly being fastest thing on Earth obsoleting everything NVIDIA is just pure marketing. It was just faster once than some midrange laptop NVIDIA, which conveniently wasn't explicitly marked as different chip sharing the branding with the desktop variant(oof).
The real benefits to going Mac is its quiet and discreet, high wife/CEO acceptance design, and their huge GPU-assignable shared RAM.
If you're okay with your desk being the wing top of a flying airplane, and 3M Peltor or David-Clark is your favorite working time headphone brand anyway, then your build will be faster and cheaper than a maxed out Mac Studio.
For me the operating system (and ecosystem in general) is one of the biggest reasons I use a Mac, but the unified memory pool is a benefit too. I am still distraught about the gaming situation though.
It is more powerful than Apple.
I wonder why Apple has not really pushed their memory bandwidth numbers. They're starting to catch up to 2020 at this point -- cool, but still struggling to run years-old models. (FLOPS is even further behind by a year or two)
Maybe they're waiting on HBM? Given that they're skipping M6 Max to go straight for M7, I'd assume they're doing a redesign for it.
GDDR has higher latency and much lower capacity. HBM requires an interposer and also has lower capacity. It's not an easy choice. The M7 family will probably double memory bandwidth using LPDDR6.
So the MacBooks might get roughly the M5 Ultra's bandwidth in a couple years. Not bad -- my M4 Max is currently stuck in 2016
Do I want one? Yes.
Can I afford it? Absolutely not.
How is this interesting compared to either a Mac Studio or Nvidia Spark? Did I miss something?
This is a tier above the best Mac Studio and two tiers above the Spark.
Big GPUs
if it goes over pcie to get to each other, it's not really "gpu memory"
wouldnt you be better off maxing out the new Mac hardware? the memory bandwidth is insane on those jobbies
Two RTX 6000 have more FLOPS and more memory bandwidth than a M5 Ultra (but they also cost more).
Depends on your use case
If running models then Apple is fine, if training or tuning or large number of users then Nvidia is king
There’s roughly 50% more memory bandwidth on an RTX 6000 pro versus the M5 Ultra. I imagine it really comes down to what you’re doing, I know my 5090 runs laps around the M5 Pro I have when it comes to local LLM performance (until it runs out of memory).
If it wasn’t for the fact the 6000 Pro is selling for like 100% ($9,000) over MSRP it would probably look a lot more reasonable.
If your goal is pure GPU compute, you should have got a 5090. It can bench comparable to the RTX 6000 Pro cards and is near guaranteed to outperform M5 Ultra.
Is this just a marketing stunt? From a completely ignorant user, doesn't AI reeuqires high bandwidth ram (like the GPU)? Is really DDR5’s bandwidth enough?
> 192 GB GPU memory
How did I miss that? I was still sleeping I guess
Hehe - yeah, I was pretty sure that you just overlooked it - pointing that out was irresistible :o)