If they tried to make a “Blackwell with no HBM” they’d need to:
Redesign the package: swap out the silicon interposer that routes thousands of traces from GPU die to HBM stacks.
Rebalance the architecture: Blackwell’s shader/Tensor cores are tuned for terabytes/sec of memory bandwidth. Running them off LPDDR5X at ~273 GB/s would starve the chip; you’d lose efficiency so badly the part would look broken.
Split the supply chain: now you’ve got two kinds of Blackwell dies/packages — one with HBM (for DGX/servers) and one without (for Spark). That’s the exact opposite of “keep supply chains clean.”
NVIDIA has never done this for a datacenter GPU family. Since Volta (V100) in 2017, every flagship GPU tile has shipped with HBM. The only NVIDIA GPUs that use GDDR/LPDDR are GeForce/consumer and some embedded/automotive SKUs — but those are separate dies, not stripped-down versions of datacenter silicon.
I don’t have any belief or evidence one way or the other. I am not an insider. Just here reading the datasheet and trying to rationalize what I’m seeing without any other evidence. I didn’t even stay at a Holiday Inn Express last night - so I’m doubly disadvantaged!
I do agree - it would seem nonsensical to design and manufacture two separate chips for this - but some other companies (ahem a certain electric car maker) have been known to include hardware but disable in firmware unless it’s purchased. Not saying that’s what is going on here but something seems ‘off’ for sure.
We really are not going to know until we get devices in hand.
The whole point of Spark is as a developer box. NVIDIA’s pitch is: “Prototype here, then scale to DGX / GB200 racks without rewriting code.”
If Spark’s GPU were architecturally different — i.e. no HBM, LPDDR-only — you’d be developing on one type of system and deploying to another. That breaks the dev-to-scale story completely.
128 GB LPDDR alone cannot host 200B params. The only way NVIDIA can make that marketing claim is if Spark’s Blackwell GPU does carry HBM3e, and the Grace LPDDR augments it as a coherent pool.
That “200B” tagline is itself indirect evidence of HBM inside Spark — they’d never promise that if it were really LPDDR-only.
With LPDDR-only @ ~273gb/s you’re memory bound and can only produce approximately 30-40 TFLOPS of INT4 in ideal conditions.
NVIDIA advertises up to 1PFLOP of INT4 under ideal conditions.
If you swap the LPDDR with HMB3e at ~8-10TB/S, and use the same math used to get the initial 30-40 Tflops number, it comes close to the 1PFLOP of int4 advertised.
I personally don’t know the math involved - so I trust you on this. I hope it’s correct! I’m eagerly awaiting an upgrade to my Jetson AGX Orin 64GB which is very memory bound it seems sporting only LPDDR5 and a claimed memory bandwidth of ‘only’ 204.8GB/s
Well at this point its either going to be a huge hit with HMB3e with the Blackwell GPU or its going to be a major flop. Excited to see either way, would like it to be worth the money though.
This SoC isn’t primarily designed for workstation use. It’s the (still unannounced but widely leaked) N1X SoC under a different name. It’s an M4 Pro/Max and Ryzen AI Max alternative.
With a MUCH slower memory bandwidth. If true - the Mac Studio starts looking far more attractive with up to 819 GB/s on the M3 Ultra (over 3 times the bandwidth stated for the DGX Spark) it would surely be a much better machine for inferencing.. assuming you put enough ram into it.
I’m only seeing a very niche case for this product with only LPDDR5x. If true, it makes it look like NVIDIA wanted to compete in the unified memory space with AMD/Apple but executed it horribly. Although the Ryzen Max tops out at ~128GB total unified memory, performance wise it would be as good or better if the model fits in memory. Also cheaper. Only advantage to the Spark becomes the CUDA ecosystem. If thats all NVIDIA is betting on, then they are just hoping its CUDA 4 ever and no other ecosystem develops. ROCM couldn’t get refined any faster at this point.
Extremely valuable conversation here. 🙏 I’ve been anticipating this system for months, and am now seriously considering Mac Studio instead. Given I can’t 2x my pre-order quantity, and even if I could, it still couldn’t load the largest models, it’s looking much less appealing.
Yea I agree, unsure if it’s worth the wait anymore, specially with the recent shipments for Framework Desktop starting to arrive for preorders.. There is actually one listed now on ebay https://www.ebay.co.uk/itm/257078798502
I am starting to feel this is not a good deal anymore as it were once announced earlier this year
“The Jetson AGX Thor Developer Kit is listed at $3,499 and is available for pre-order from selected distributors, with shipments expected to begin on November 20, 2025.”
If this means anything it’s gonna be a while before the Spark ships.