DGX Station vs 8x RTX Pro 6000: Which AI Workstation Wins?
DGX Station's HBM3e vs an 8x RTX Pro 6000 rig's GDDR7, compared on real LLM throughput, video generation, and coding workloads.
Neither machine wins outright. They win different jobs.
NVIDIA’s DGX Station and a workstation built from eight RTX Pro 6000 GPUs represent two different bets on how to run large models locally. DGX Station leans on 252 GB of HBM3e memory running at roughly 7 terabytes per second, paired with 748 GB of unified LPDDR5 memory for overflow. The 8x RTX Pro 6000 rig (referred to as “Grando” in the source testing) leans on raw GDDR7 capacity across eight cards, enough to hold a 614 GB model entirely in VRAM, but with GPUs talking to each other over PCIe Gen 5 at a comparatively pedestrian 64 GB/s per link. Which one is faster depends entirely on the model architecture and the job.
TL;DR
- DGX Station’s HBM3e bandwidth (about 7 TB/s across 252 GB) crushes single-user generation speed on models that fit inside it, with GPT-OSS 120B hitting what testers called an absurd tokens-per-second rate running entirely from HBM.
- The 8x RTX Pro 6000 rig holds more total GPU memory, enough to fit a 614 GB model like DeepSeek V4.1 Flash (FP8) entirely across its GDDR7 pool without spilling to slower memory.
- Mixture-of-experts architectures change the calculus, because MoE models can run their inactive experts from slower memory without a heavy speed penalty, which plays to DGX Station’s unified memory design.
- In head-to-head testing on DeepSeek V4.1 Flash, the RTX Pro 6000 rig only edged out DGX Station by a small margin (307 tokens per second range) despite one system running on HBM and the other on GDDR7 plus interconnect, showing the gap narrows once a model fits natively.
- Multi-user throughput matters more than single-chat speed for real deployments, with DGX Station posting around 3,800 tokens per second in one multi-user test and over 2,000 tokens per second on a Nemotron 3 Super 12B configuration.
- Video generation favored DGX Station in testing, with a model called H3 running noticeably faster on DGX Station’s HBM, producing 5-second clips in about two minutes.
- Neither machine is meant for solo chat sessions, the actual use case demonstrated was orchestrating multiple AI agents and coding tasks across the hardware simultaneously, which is a very different workload than asking one model one question.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
How does DGX Station’s memory architecture compare to GDDR7?
DGX Station’s defining trait is its HBM3e pool: 252 GB of memory running at roughly 7 terabytes per second of bandwidth. That is an order of magnitude faster than GDDR7, which the RTX Pro 6000 cards use. But 252 GB is a hard ceiling for the HBM tier. Once a model’s active weights exceed that, DGX Station spills into its 748 GB of unified LPDDR5 memory, which runs around 600 GB/s. Still fast for system memory, but nowhere near HBM speeds.
The 8x RTX Pro 6000 rig takes the opposite approach: eight GPUs’ worth of GDDR7 memory stacked together gives it enough total capacity to hold very large models (the testing showed a 614 GB DeepSeek model fitting entirely on the card pool). The tradeoff is that those eight GPUs need to communicate with each other, and they do it over PCIe Gen 5 at 64 GB/s per direction, a fraction of what GPUs on a tighter interconnect (like NVLink) would offer. For workloads that need to shuttle data between GPUs constantly, that interconnect becomes the bottleneck rather than memory bandwidth.
Which machine is faster for running large language models?
It depends on whether the model fits natively on one system and how its architecture distributes compute. On GPT-OSS 120B, which ran entirely from DGX Station’s HBM, generation speed was described as absurd, because nothing had to fall back to slower memory. That is close to a best-case scenario for the HBM design.
On DeepSeek V4.1 Flash, a much larger model at 614 GB in FP8, the dynamic flips. That model fits entirely on the RTX Pro 6000 rig’s combined GDDR7, while DGX Station has to split weights between its 252 GB HBM tier and its slower unified memory. In testing, the RTX Pro 6000 system came out ahead, around 307 tokens per second, but only by a small margin. That is notable: a model running partly on GDDR7 and partly through a 64 GB/s interconnect only modestly outpaced a system splitting the same job between HBM3e and LPDDR5.
The reason is architectural. Modern mixture-of-experts (MoE) models don’t need every parameter active on every token. Expert layers that aren’t firing on a given pass can sit in slower memory without tanking throughput, because they’re not on the critical path. That makes MoE models a good match for DGX Station’s two-tier memory setup: hot weights in HBM, cold experts in unified memory, and the system only pays a speed penalty on the portion that spills over.
Prompt processing (prefill) is a separate story. HBM wins prefill every time, by a wide margin, because that phase is almost purely bandwidth-bound and doesn’t benefit from the kind of expert-routing tricks that help with generation.
How do they compare for video generation?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
In testing with a video generation model referred to as H3, DGX Station pulled ahead clearly. H3 is a relatively small model, so it fits comfortably inside the HBM tier, and that’s where DGX Station’s bandwidth advantage shows up most directly. Five-second clips took about two minutes to generate on DGX Station, noticeably faster than on the RTX Pro 6000 rig, with the hardware running audibly quieter and running a bit hotter under the sustained generation load. For smaller, compute-dense models that fit entirely in fast memory, DGX Station’s architecture does exactly what it’s designed to do.
What does multi-user throughput look like in practice?
Single-user chat speed is the easiest number to benchmark but the least representative of how these machines actually get used. Anyone buying a six-figure AI workstation isn’t running one chat session, they’re supporting a team, running background agents, or keeping a privacy-sensitive workload entirely off the cloud. Multi-user throughput is the number that matters there.
In that mode, DGX Station posted around 3,800 tokens per second in aggregate, and just over 2,000 tokens per second running a Nemotron 3 Super 12B configuration. Those numbers matter because models at that size are genuinely useful for real development work: unit tests, integration tests, scaffolding, and other bounded coding tasks. The practical pattern demonstrated was using a larger, smarter model (like Claude Opus, accessed via API) as an orchestrator or manager, while a fleet of smaller local models handled the grunt work on-premises, burning electricity instead of cloud tokens. That division of labor is where local hardware like this earns its keep, not in replacing frontier cloud models outright.
Is either machine actually worth it?
That depends on what “worth it” means for the buyer. Cloud API tokens are, comparatively, not expensive, and the source testing was explicit that these machines only make financial sense if you can keep them busy close to 24 hours a day. That points toward teams running continuous agent workloads, privacy-constrained environments where data can’t leave the building, or developers running many parallel coding agents against local models instead of paying per-token for every background task. For a single person asking occasional questions, neither machine is a rational purchase. For a small team running orchestrated agent pipelines around the clock, the math changes, and the choice between DGX Station and a multi-GPU RTX Pro 6000 rig comes down to whether your models are big and MoE-shaped (favoring unified memory) or mid-sized and GDDR7-friendly (favoring raw GPU memory pooling).
Frequently Asked Questions
Is DGX Station faster than a multi-GPU RTX Pro 6000 workstation?
It depends on the model. DGX Station wins clearly on models that fit inside its 252 GB HBM3e pool, especially smaller or mixture-of-experts models. An 8x RTX Pro 6000 rig can hold much larger models entirely in GDDR7 memory and was faster in testing on a 614 GB DeepSeek model, though only by a small margin.
Why does mixture-of-experts architecture matter for this comparison?
MoE models only activate a subset of their parameters per token. That lets inactive expert weights live in slower memory (like DGX Station’s unified LPDDR5) without a major speed penalty, since they’re not constantly on the critical compute path. This plays to DGX Station’s tiered memory design in a way dense models don’t.
How much memory does DGX Station have compared to the RTX Pro 6000 rig?
Remy is new. The platform isn't.
Remy is the latest expression of years of platform work. Not a hastily wrapped LLM.
DGX Station has 252 GB of HBM3e memory plus 748 GB of unified LPDDR5. The 8x RTX Pro 6000 rig’s total capacity comes from pooling GDDR7 memory across all eight cards, enough in testing to hold a 614 GB FP8 model entirely in GPU memory.
What is multi-user throughput and why does it matter more than single-chat speed?
Multi-user throughput measures how many tokens per second a system can generate across concurrent sessions or agents, rather than one user typing into one chat window. It matters because workstations like these are typically bought to support teams or continuous background agent workloads, not single interactive sessions.
Can these machines replace cloud AI models entirely?
Not based on the workloads shown. The practical setup paired a strong cloud model as an orchestrator with smaller local models handling routine coding tasks like unit tests and scaffolding on local hardware, reducing cloud token spend rather than eliminating it.