Enjoying this issue?
Get tomorrow's AI & engineering digest in your inbox — hand-picked, summarized, and always spam-free.
TLDR
The 64GB DGX Spark at $4,999 is a poor value compared to the Mac Mini M5 Pro, which offers faster memory and costs roughly $2,000 less. The 128GB Spark's price has risen to around $6,950, making the 64GB model's price less of a hike than it seems, but the Mac Mini remains a better choice for most users unless CUDA is required.
Key points
The 64GB DGX Spark costs $4,999, while the 128GB model launched at $3,999 but now costs around $6,950.
The Mac Mini M5 Pro with 64GB costs between $2,699 and $3,200 and has faster memory bandwidth.
Two 64GB Sparks combined cost $9,998, which is more expensive than a single 128GB Spark.
The speaker recommends the 64GB Spark only if you need CUDA; otherwise, the Mac Mini is a better value.
Techniques
- low precision compute
- model quantization
- agent-based testing
Stop scrolling. Start reading smarter.
Receive the day's most important AI & engineering updates in one concise email. No spam.
Transcript (captions)
Nvidia wants $4,999 for the 64 GB DGX Spark. And I went through the price history of the 128 GB model it cuts in half. That one launched at $3,999.
Press reports now put it near $6,950. What I wanted to know was whether $4,999 is a bargain or a hike. If you've been weighing $4,999 for 64 GB against last year's $3,999
for 128, you've been weighing it against a price that stopped existing in February. I expected a plain price hike. Against launch, it is one. Against today's 128 GB price, it isn't. For just
running models, my pick is a 64 GB Mac Mini M5 Pro. And the rest of this is why if you need CUDA, Nvidia's own software layer that a lot of AI tools are built on, the answer changes. The popular take
is easy to find, even from titles alone. One ends in makes no sense. Another asks, "Is Nvidia insane?" And the reaction has real math behind it. The 128 GB Spark launched at $3,999.
This box has half the memory for $1,000 more. But that $3,999 is the old price. In February, Nvidia raised the 128 GB Founders Edition to $4,699,
citing memory supply with no hardware changes. Tech Insider now puts the 128 GB near $6,950 and XDA says close to $7,000. Press figures and retail prices vary, but none
is $3,999. By the end, you'll know what that $4,999 buys and which box fits which job. One question stays open until later. If you start with 164 GB Spark and add a
second, is that cheaper than buying 128 GB today? Nvidia wasn't alone. Apple raised Mac Studio prices on June 25th and cited the same AIdriven memory surge. The M4 Max went up $500. The M3
Ultra went up $1,300. And the M5 Ultra then launched at $5,499, which Macworld counts as $1,500 above the M3 Ultra's launch price. The RTX5090
shows it louder. Nvidia store listed it at $4,929.99 in August per Kami Labs against a $1,999 launch price. That's $146.5% above launch. It makes the launch price
look like a clearance sale. How long does this last? Kami Lab says Intel and Silicon Motion project a return to normal around 2028. And a PC Gamer headline says it could run past that.
One thing didn't change. Nvidia's blog says the 64 GB model uses the same GB10 chip as the 128 GB one. The memory is what got cut. My read is that the 64 GB Spark is what that shortage looks like
as a product. Less memory and a higher price than the launch one. Nvidia doesn't say so. And my reasoning is only the dates and the prices. Think of the Spark's memory as a workbench. Its size
decides what fits on the bench. Its speed decides how quickly you can reach across it. Size is a yes or no. Speed is how fast the answer comes out. Like any workbench picture, it only covers the
reaching, and the thinking is a separate job we'll get to. Does anything else do the job better? I expected one number for the Spark's memory speed. Nvidia's product page says 273 GB a second and
artificial analysis lists the 128 GB Spark at 256. The page marks the 64 GB option as coming soon and never says outright that it keeps that speed. Same chip, so I'm using Nvidia's 273. Either
way, the Mac Mini M5 Pro lists 307 and that's faster. Faster memory only helps if the model fits. So start with what fits. Quen 3.8 8's 27 billion model is dense. So the whole thing works on every
word. Stored at 16 bits, each number takes two bytes. So 27 billion* 2 is about 54 GB. That barely leaves room on a 64 GB bench. Squeeze it to four bits per number and it drops to around 17 or
18 GB, which fits with room to spare. There's also a third party squeeze of a Quen 3.8 coding model called Flash. Next, ISTALLab shrank its 176.9 billion parameters until the part that must stay
in memory is 29.6 GB per the model card, a 28.8 GB lookup table can sit on disk. It scores 75.60 on SWEBench verified, a test of fixing real software bugs against 82.80 for the unsqueezed model.
On paper, 29.6 6 GB in a 64 GB box sounds roomy until the operating system and the conversation asked for seats too. No Spark run of it showed up in the sources. So, this is paper math until
someone runs it. What doesn't fit? Deepseek V4 Flash at 284 billion parameters and GLM 5.3 Flash at 320 billion. Nvidia's blog puts the ceiling at up to 100 billion parameters. Divide
each by 100 and you land at roughly three times that ceiling. Parameters aren't the whole rule, though. What counts is the gigabytes the squeeze build needs, and the heavily squeezed
builds of these two still don't fit in 64. Back to the Mac Mini, Apple Insider lists the 64 GB 15 core M5 Pro at $2,699. Retailer listings run higher up to about
$3,200. Set those against $4,999 and that works out to a gap of about $1,800 to $2,300. None of it comes from Apple's own page, so check there before
you buy. Memory speed isn't the whole race, though. The one independent benchmark I read is Artificial Analysis's Agent Test from September 29th. It tested the 128 GB Spark, not
the 64 GB one. The other machines were an AMD Ryzen AI Halo, a MacBook Pro with an M5 Pro and 64 GB, and an RTX 5090. The Spark was 1.4 4 to 1.7 times faster than the AMD box on three of four
models. On Quinn 3.5 9 billion, the AMD box tied it. The MacBook Pro landed within 2 to9% of the AMD box on two models and 21% slower on that 9 billion one. Why would the Spark come out ahead
when its memory isn't the fastest? Artificial analysis says it has more math hardware for small number formats called low precision compute, so it reads your prompt faster. Reading is
mostly calculation. Writing the answer means pulling the model through memory again for every new word. By that logic, the max faster memory should help there. That's my inference, and the test didn't
measure it. The workbench picture only covers that second job. Stitch those results together, and the 128 GB Spark looks ahead of a Mac laptop, but you only get there by joining two
comparisons through the AMD box. The test is eight agent tasks with up to about 56K of context, so it says little about other work. One video titled DGX Spark versus Mac Studio versus RTX5090
claims a hands-on run of all three. Its title points to the 128 GB Spark since the 64 GB 1 only ships October 23rd. The RTX 5090 was more than 3.5 times faster than every other machine on every model
that fits in its 32 GB. Subtract 29.6 from 32 and the Flash Next build leaves a couple of GB for everything else. And that's the card that was $4,929.99 on Nvidia store in August. Nvidia's
answer to the 64 GB limit is a second box. Two 64 GB sparks act as 128 GB combined with twice the memory bandwidth and up to 1.7 times the performance on Quen 3.8 27 billion. That's Nvidia's own
test compared with a single system. And the blog doesn't give the setup. Two units at $4,999 is $9,998. And Tech Insider states the same. One
128 GB Spark at about $6,950 gives you the same 128 GB. Take one from the other and the pair cost roughly $3,000 more. Though you do get two chips for it, so it's not quite the same 128.
Starting small turns out to be the expensive way to end up at 128. That $9,998 lands close to a 256 GB Mac Studio M5 Ultra. Apple's spec page lists 1.2
terabytes a second of memory bandwidth with 96, 256, and 512 GB options. If Nvidia's twice the bandwidth claim holds, a pair of Sparks sits around 546 GB a second, double 253, and you get
that. The Mac Studio has twice the memory of the pair and more than twice the bandwidth. The base is $5,499 with 96 GB and the 256 upgrade adds $4,000. So about $9,499
going by fellow AI and Dadnology. AD community piece puts the full chip with 256 at $10,799. So the tier matters and it runs Mac OS. So no CUDA. So, the question from the
start, 164 GB Spark now and a second later ends at $9,998. One 128 GB today is about $6,950 if you can find one near that price. The two-step path only makes sense if you
might never need the second box, or if $4,999 is what you can spend today. Divide each price by its gigabytes. The 64 GB Spark comes to about $78. The 128 GB Spark at
$6,950 comes to about $54. The 64 GB Mac Mini lands between $42 and $50. Next to the Mac Mini, the extra money buys Nvidia's software, CUDA, and a chip that looked
faster in one agent benchmark. That test used the 128 GB Spark in a MacBook, so treat it as a lean. Next to the 128 GB Spark, the 64 GB one is the same chip with half the memory and a lower ticket.
If your work has to run on CUDA, the 64 GB Spark is my pick. If the models fit in 32 GB, price a 5090 against it. If you know you'll want 128 GB, price the 128 GB box first because a second 64 GB
unit ends up costing more. If you only want to run models, skip it. A 64 gigabytes Mac Mini M5 Pro has faster memory. And divide 307 by 273 and you get about 12% more. It also costs
roughly $2,000 less depending on configuration. The exception is long prompts and agent loops where the Spark's compute looks stronger. That comes from artificial analysis putting
the 128 GB Spark ahead of an AMD box that a MacBook M5 Pro came close to on two models. It's a lean, not a head-to-head. and the test had no Mac Mini and no 64 GB Spark. For long
answers from short prompts, the Max faster memory should matter more. A 5090 is quickest on models that fit in 32 GB. And a 256 GB Mac Studio is for people who want the room and can spend about
$9,500 or more. So, the comparison from the start, $4,999 against last year's $3,999 was against a price that's gone. What are you running that needs more than 64
GB?