An RTX PRO 6000 Blackwell Max-Q has 96 GB of VRAM, the card's own memory. Two give you 192 GB and four give 384 GB. The question is what each step lets you do that the step before could not.
The answer is not "everything gets faster". More cards add memory first. Speed only improves in some cases, and we will say which.
What more cards really add
- Memory. Cards pool their VRAM, so two cards hold a model of up to about 190 GB (the selector counts 95.4 GB usable per card). Three or four cards go further. This is what lets a very large model run fully in VRAM instead of spilling into slow system memory.
- Sometimes speed. With tensor parallel (the model split across 2 or 4 cards, which work on every word together, as in vLLM), the selector assumes about 1.45 times one card's speed on two cards and about 2.2 times on four. That helps a model that needs the extra cards. A model that already fits on one card does not get faster by adding more.
- More chats at once. Extra memory leaves room for more KV caches (the model's working memory for each conversation). This is where small models gain most.
Three cards are a special case. Tensor parallel needs 2 or 4 cards, so three only add memory. That is why this post compares one, two and four.
Which card is it?
All three builds use the Max-Q edition: 96 GB of GDDR7 with ECC, 1,792 GB/s of memory bandwidth and 300 W, per NVIDIA. The Workstation Edition has the same memory but is rated 600 W. In a tower with several cards, the 300 W card is the practical one: it keeps the power draw and the heat inside what one case and one circuit can handle.
Model sizes against the three steps
The chart shows the memory each model needs at the quant (file compression) named, with a 32,000-token context and one user. The three lines mark the usable memory of one, two and four cards.
Memory needed, 32K context, one user, against 1, 2 and 4 cards
Estimate. A bar that ends left of a line fits fully in VRAM on that many cards.
- 1 card (95.4 GB usable)
- 2 cards (190.8 GB)
- 4 cards (381.6 GB)
Source: EMD Lab AI selector estimate, calculated from model file sizes and configuration files. Not a measurement.
What each step unlocks
| Model and quant | 1 card (96 GB) | 2 cards (192 GB) | 4 cards (384 GB) |
|---|---|---|---|
| Qwen3.8 27B, 8-bit | about 42 tok/s, 27 chats | about 42 tok/s, 69 chats | about 42 tok/s, 151 chats |
| Llama 3.3 70B, 4-bit | about 26 tok/s, 4 chats | about 26 tok/s, 13 chats | about 26 tok/s, 31 chats |
| gpt-oss-120b, 4-bit | about 176 tok/s, 24 chats | about 176 tok/s, 102 chats | about 176 tok/s, 258 chats |
| Qwen3.5-122B-A10B, 4-bit | about 129 tok/s, 15 chats | about 129 tok/s, 114 chats | about 129 tok/s, 311 chats |
| Qwen3.8-Flash-Next, 4-bit | offload, about 46 tok/s | about 189 tok/s, 69 chats | about 220 tok/s, 269 chats |
| GLM-5.3, smallest quant | offload, about 9 tok/s | offload, about 28 tok/s | about 124 tok/s, 37 chats |
tok/s means tokens per second, roughly words per second. Two things stand out. The speeds repeat across a row when the model already fits on one card, so the extra cards buy chats, not speed. And the rows that change speed are the ones where the model did not fit before.
One card, 96 GB
Every model from 27B to 31B runs at 8-bit, which is close to the original quality. gpt-oss-120b writes at about 176 tok/s, Qwen3.5-122B-A10B at about 129 and Llama 3.3 70B in 4-bit at about 26, because the 70B model is dense and reads all its weights for every word. This is the right size for a department that wants a strong model for a handful of people. For more than that, count the chats: Llama 70B has room for only four conversations at 32K.
Two cards, 192 GB: the AI workstation
Two cards are where the 180B-class Qwen3.8-Flash-Next at 4-bit fits entirely in VRAM, at an estimated 189 tok/s. On one card it must offload and drops to about 46. Two cards also change the chat counts a lot, for example gpt-oss-120b from 24 to 102 conversations at 32K. The AI preset is built this way: 2 × RTX PRO 6000 Max-Q.
Four cards, 384 GB: the Team AI server
Four cards open the very large models, but with a limit. GLM-5.3 and DeepSeek V4.1 Flash fit only at their smallest quants, around 2-bit, which costs some quality. At 4-bit they do not fit: GLM-5.3 needs about 486 GB. Four cards also give the biggest chat counts, and with tensor parallel they run a model that needs them about 2.2 times faster than one card would, as the selector assumes. Choose four cards for these huge models or for a crowd. For gpt-oss-120b alone, two cards already give 102 chats.
The power point is part of the plan. Our configurator estimates about 1,880 W at the wall for the four-card Team build, so it needs its own circuit. Check your power supply and breaker before you order. How many people can chat at once also depends on the serving software, so we size this one together with you. The Team preset starts from the right parts.
The short version
- Each RTX PRO 6000 Max-Q adds 96 GB. More cards mean room for bigger models and more chats first, speed second.
- Tensor parallel on 2 or 4 cards helps only a model that needs them. Three cards add memory only.
- One card is for 27B to 120B models. Two cards add the 180B class and many more chats. Four cards add the largest models, at about 2-bit.
- All figures here are selector estimates, calculated and not measured. The four-card build needs its own power circuit.