With the announcement of Jetson AGX Thor availability for purchase, I’ve been re-reading the marketing materials and am a bit confused.
The specification sheet for the Jetson Thor lists 2070 TFLOPS of AI performance, while the DGX Spark lists 1 PFLOP (both FP4 sparse). The DGX Spark marketing materials also appears to use PFLOPS and AI TOPS interchangeably, but aren’t they different measures for floating point vs integer operations per second?
The first sentence of each products overview page seems to imply that the Jetson Thor has 2x the performance;
Jetson Thor: “NVIDIA® Jetson Thor™ series modules give you the ultimate platform for physical AI and robotics, delivering up to 2070 FP4 TFLOPS of AI compute and 128 GB of memory with power configurable between 40 W and 130 W.”
DGX Spark: “Powered by the NVIDIA GB10 Grace Blackwell Superchip, NVIDIA DGX™Spark delivers 1 petaFLOP of AI performance in a power-efficient, compact form factor.”
Is this a case of marketing terminology conflation, or might the Jetson AGX Thor provide better local inference performance compared to DGX Spark?
The RAM being similar but performance different could be down to scaling with power draw. The DGX Spark ships with a 240w USB-C brick, and I think it’s specced to draw significantly higher than the Thor at 170W.
These are also rather different markets and I believe they will run different OSes; those who want DGX Linux know who they are and which version they need. The Thor is the evolution of the Jetson platform. Though they potentially could be moving toward convergence…
I’m really “scared” about what we’re seeing right now with AGX Thor. I mean, great to see a board like this, but what really happens to our DGX Spark now?
We’ve suffered multiple delays, we still do not have a real datasheet to compare for example the number of tensor core and others techinical aspect, the only number provided is that the AGX is outperforming by a 2X the Spark in terms of FP4 TFOPS.
I’m feeling uncomfortable now with other placed through our partners for a total of 2 DGX Spark, even because we don’t have a clear shipping date yet, while Thor directly comes ouut of the shelf after it’s announcement. I don’t know….
The DGX Spark has 6144 CUDA cores, or just as many as RTX 5070. I believe I saw numbers claiming 1000 TOPS in FP4 sparse mode. I’m not sure how many tensor cores, or what type of tensor cores even, if any?
The AGX Thor has 2560 CUDA cores, with 96 fifth-generation Tensor cores; benchmarks I’ve seen so far indicate it performs on LLMs about as fast as an RTX 5070. I’m unsure if these were tests used in FP4 Sparse mode, as NVIDIA rates it for 2070 TOPS in FP4 sparse mode; half that for FP8 sparse mode, and lower still for dense mode. I’m not even sure if the benchmarks were even using the tensor cores – I can’t say I’ve been too impressed with the reviews so far.
I suppose I can see the Spark using more power due to using CUDA cores instead of Tensor Cores as the main source of processing, offering FP32 performance needed for precision during training and perhaps greatly versatility. For LLM inference, I’d imagine then the Thor might be more desirable, but that leaves me bewildered by the weak benchmarks I’ve seen so far; even the Nvidia provided LLM benchmarks.
It almost seems like unless you actually intend to use the Thor for building robots, you’d be better off just buying a Mac M4 Max 128GB or AMD AI Ryzen. As for the Spark, the 200G ConnectX port seems interesting, but are there any 200G switches available to connect more than 2 together? None that make financial sense probably. Are these going to be the same speed as a 5070; so the same speed as a Thor, but FP32 friendly?
I’m also very interested in understanding the architectural difference between Thor and spark and glad I’m not the only one confused
indeed, would like to confirm whether flops is really double for Thor than spark? Also, Both have the same kind and size of memory but I’m wondering how the gb super chip differs, is it the same unified memory performance? Is this about the trade off with Thor architecture which is not super chip but more energy efficient, because one is targeted to robotics and the other to desktop dev users (and Thor is more focused on inference while spark on training)? I’ve seen some numbers about what cores are in the Thor the spark doesn’t have the same info in the official spec. I’m really curious in seeing detailed internal difference between the two GPUs.
The NVIDIA Jetson Thor Developer Kit is a purpose-built developer platform targeted at developers creating robotics and physical AI solutions that deploy with embedded Jetson modules. DGX Sparkis a purpose-build compute to build and run AI, targeted at AI developers and data scientists who need to augment current laptop, desktop, cloud, or data center resources to provide large local memory and access to the NVIDIA AI software stack for their AI prototyping, fine-tuning, inference, data science, and general edge workloads.
Some other differences that are noteworthy ( doesn’t really apply to AI directly but perhaps how it is used in an AI workflow).
DGX Spark has one NVENC/ NVDEC chip. Thor has Two.
DGX Spark has a connect-x nic. Thor is not connect-x but a 4x25g nic. It doesn’t appear to support RDMA among other features you get with connect-x. which also means you probably can’t combine the thor modules very easily.