Quick Answer: The RTX 5090 and the RTX PRO 6000 Blackwell use the same GB202 chip and the same 1,792 GB/s of memory bandwidth, so on any model that fits in 32 GB they generate text at almost the same speed; in Hardware Corner’s llama.cpp tests the 5090 was even slightly faster (50.92 versus 45.72 tokens per second on Qwen3 32B). The PRO 6000’s case is its 96 GB: it runs Llama 3.3 70B and gpt-oss-120b on one card, which the 5090 cannot load. Buy the RTX 5090 if your models fit in 32 GB, and the RTX PRO 6000 if you need 70B to 120B models, multi-user serving or ECC memory. At November 2026 prices ($6,919 for the 5090 in Tom’s Hardware’s price index, $16,000 for the PRO 6000 at NVIDIA) the PRO 6000 is also the cheaper card per GB of VRAM.
In 2025 this comparison had an easy answer: the 5090 cost $1,999 and the PRO 6000 about $8,565, so three 5090s gave you the same 96 GB for less money. The memory shortage of 2026 changed that. This guide compares the two cards on the numbers that decide a local-AI purchase: memory, bandwidth, published benchmark results and dated prices. We don’t test hardware ourselves. Every speed figure below comes from the publication or project named next to it.
Specs side by side
| GeForce RTX 5090 | RTX PRO 6000 Blackwell Workstation Edition | |
|---|---|---|
| GPU | GB202 (Blackwell) | GB202 (Blackwell) |
| CUDA cores | 21,760 | 24,064 (about 11% more) |
| Memory | 32 GB GDDR7 | 96 GB GDDR7 with ECC |
| Memory bus and bandwidth | 512-bit, 1,792 GB/s | 512-bit, 1,792 GB/s |
| AI TOPS (NVIDIA's FP4 rating) | 3,352 | 4,000 |
| Board power | 575 W | 600 W (300 W Max-Q version) |
| Partitioning (MIG) | No | Up to 4 instances of 24 GB |
| Drivers | GeForce Game Ready and Studio | NVIDIA RTX Enterprise |
| Display outputs | 3× DisplayPort 2.1b, 1× HDMI 2.1b | 4× DisplayPort 2.1b |
| Launch price | $1,999 (January 2025) | $8,565 (April 2025) |
The table explains almost every benchmark result below. Token generation on a local model is limited by memory bandwidth, because each new token reads the model’s weights from VRAM, and both cards read at 1,792 GB/s. The PRO 6000’s extra cores help with compute-heavy work such as prompt processing and serving many users at once. The memory size decides which models load at all.
Price per GB: the 2026 inversion
Both cards cost far more than at launch, but not by the same factor. Here is what we found, with the source and date for every price:
| RTX 5090 (32 GB) | RTX PRO 6000 (96 GB) | |
|---|---|---|
| Launch price | $1,999 (January 2025) | $8,565 single unit (April 2025, per Tom's Hardware) |
| Price found, November 2026 | $6,919 best US price (Tom's Hardware GPU price index, Nov 3 2026) | $16,000 at NVIDIA's US Marketplace; Newegg and B&H from $15,999 (Thunder Compute, Nov 3 2026) |
| Increase since launch | About +246% | About +87% |
| Price per GB of VRAM | About $216 | About $167 |
| Cost of 96 GB of VRAM | Three cards: about $20,760, up to 1,725 W | One card: $16,000, 600 W |
| Cloud rental instead | $0.65 per GPU-hour paid by CloudRift for its Oct 2025 benchmark | $1.69 to $5.50 per GPU-hour across 9 providers (Thunder Compute, Nov 3 2026) |
At launch the math favoured the consumer card: three 5090s at MSRP cost $5,997 for 96 GB, against $8,565 for one PRO 6000. In November 2026 one PRO 6000 is about $4,760 cheaper than three 5090s and draws a third of the power. NVIDIA never announced its price changes. Tom’s Hardware reported on August 12, 2026 that NVIDIA’s US store had moved the PRO 6000 from $13,250, set in June 2026, to $16,000, and tied the increases to the memory shortage, since the card carries 96 GB of GDDR7. The 5090 rose even faster in relative terms: Tom’s Hardware’s price index lists its lowest-ever US price as the $1,999 MSRP and its best price on November 3, 2026 as $6,919. Treat both figures as dated snapshots and check the live price before you buy.
Speed on models that fit both cards
Hardware Corner ran both cards through llama.cpp’s benchmark tool at 4-bit (Q4_K_XL) on Ubuntu 24.04 with CUDA 12.8, in a ranking last updated December 9, 2025. These are single-user results, the way most people run models at home.
| Model (4-bit, 16K context) | RTX 5090 generation | RTX PRO 6000 generation | RTX 5090 prompt processing | RTX PRO 6000 prompt processing |
|---|---|---|---|---|
| Qwen3 8B | 145.34 t/s | 140.62 t/s | 6,956 t/s | 7,588 t/s |
| Qwen3 14B | 102.68 t/s | 96.86 t/s | 4,473 t/s | 5,027 t/s |
| Qwen3 30B-A3B (MoE) | 141.63 t/s | 139.76 t/s | 4,669 t/s | 5,084 t/s |
| Qwen3 32B | 50.92 t/s | 45.72 t/s | 2,077 t/s | 2,374 t/s |
| gpt-oss-20b | 249.19 t/s | 237.92 t/s | 7,168 t/s | 7,479 t/s |
Source: Hardware Corner GPU ranking for local LLMs (llama.cpp llama-bench, updated December 9, 2025). t/s = tokens per second.
The 5090 wins every generation row by 1% to 11%, and the PRO 6000 wins every prompt-processing row by 4% to 14%. Neither gap is one you would notice in a chat window. Hardware Corner’s own summary puts the 5090 at 106% of the PRO 6000’s token-generation speed. In practice, if your models fit in 32 GB, the extra $9,000 for the PRO 6000 buys no speed for a single user. At long context the picture evens out: at 131K context the PRO 6000 was ahead on Qwen3 14B (39.83 versus 37.2 t/s) and gpt-oss-20b (118.44 versus 112.01 t/s).
What only the PRO 6000 can run
At 4-bit, plan for roughly 0.5 to 0.6 GB per billion parameters plus a few GB for context. That puts the 5090’s ceiling at about 32B dense models with long context, or larger mixture-of-experts models with part of the weights in system RAM.
| Model | Approx. memory needed | RTX 5090 (32 GB) | RTX PRO 6000 (96 GB) | PRO 6000 generation, 16K context |
|---|---|---|---|---|
| gpt-oss-20b (MXFP4) | About 13–16 GB | Yes | Yes | 237.92 t/s |
| Qwen3 32B, 4-bit | 20–24 GB | Yes | Yes | 45.72 t/s |
| Qwen3 32B, 8-bit | About 34–38 GB | No | Yes | Not in that test |
| Llama 3.3 70B, 4-bit | 42–48 GB | No | Yes | 28.24 t/s |
| gpt-oss-120b (MXFP4) | About 61–65 GB | No | Yes | 182.23 t/s |
| GLM-4.5-Air (106B MoE), 4-bit | About 60–73 GB | No | Yes | 82.87 t/s |
| Mistral Large 123B, 4-bit | About 70–75 GB | No | Yes | 17.73 t/s |
| Llama 3.3 70B, 8-bit | About 72–80 GB | No | Yes, short context | Not in that test |
| Qwen3 235B-A22B, 4-bit | About 130–140 GB | No | No (needs two cards or offload) | — |
Speed column: Hardware Corner (llama.cpp, Q4_K_XL / MXFP4, December 2025). The 5090 rows for 70B and larger are blank in that test because the models do not fit.
The standout number is gpt-oss-120b at 182 tokens per second on one PRO 6000. OpenAI sized that model for a single 80 GB GPU, and as a mixture-of-experts model it reads only about 5 billion active parameters per token, so it runs faster than a dense 32B model on the same card. That makes it the strongest single argument for 96 GB: a 120B-class model at chat speeds on one card. Hardware Corner’s separate PRO 6000 page (re-tested March 2026) recorded 210.1 t/s for gpt-oss-120b at 4K context and 99.8 t/s at 128K.
Serving many users: batched throughput
Single-user speed is close. Serving many requests at once is where the PRO 6000’s extra cores and memory pay off, because the spare VRAM holds a larger KV cache for more parallel requests. CloudRift benchmarked rented GPUs with vLLM’s serving benchmark (1,000 input and 1,000 output tokens per request, up to 400 concurrent requests) on October 9, 2025:
| Model (vLLM) | RTX 5090 setup | RTX 5090 throughput | 1× RTX PRO 6000 throughput |
|---|---|---|---|
| Qwen3-Coder-30B-A3B (AWQ 4-bit) | 1 card | 4,570 t/s | 8,425 t/s |
| Qwen3-Coder-30B-A3B (AWQ 4-bit) | 4 cards | 12,744 t/s | 8,425 t/s |
| Llama 3.3 70B (AWQ 4-bit) | 2 cards | 1,230 t/s | 1,031 t/s |
| GLM-4.5-Air (AWQ 4-bit) | 4 cards | 4,622 t/s | 3,169 t/s |
Source: CloudRift, “Benchmarking RTX GPUs for LLM inference” (Dmitry Trifonov, October 9, 2025). Throughput is aggregate tokens per second across all concurrent requests; multi-card 5090 setups ran as replicas or with pipeline parallelism.
One PRO 6000 served 1.8 times as many tokens per second as one 5090 on the same 30B model. Multiple 5090s can beat one PRO 6000 on raw throughput, but they need two to four cards, a server-class host and two to four times the power to do it. At the rental prices CloudRift paid ($0.65 per hour for a 5090, $1.29 for a PRO 6000), the cost per million tokens on the 30B model came out almost equal ($0.040 versus $0.043), and in the PRO 6000’s favour on GLM-4.5-Air ($0.113 versus $0.156 for four 5090s).
Power, cooling, drivers and the fine print
- Power. The 5090 is rated at 575 W and the PRO 6000 Workstation Edition at 600 W. Both use a single 16-pin 12V-2x6 connector, and Hardware Corner suggests a 1,000 W Gold power supply for the PRO 6000. The Max-Q version keeps all 96 GB but is capped at 300 W with a blower cooler, which is the version to stack two to four cards in one tower.
- ECC memory. The PRO 6000’s GDDR7 has error correction. It matters for long training runs and shared servers, much less for chatting with a model at home.
- MIG partitioning. The PRO 6000 can be split into up to four isolated 24 GB GPUs (Lenovo’s product guide lists “up to 4 MIGs @ 24GB”), so one card can serve four users or four models. The 5090 cannot do this.
- Licensing. NVIDIA’s GeForce driver license does not cover data-center deployment. That is irrelevant on a desk, but if you plan to rack cards in a server or rent them out, the PRO 6000 is the licensed option.
- Availability. Both cards are hard to find at a fair price in late 2026. Digital Trends reported on October 17, 2026 that NVIDIA denies the 5090 is discontinued but admits supply is constrained, while Micro Center has marked several partner models as “Not Carried”. On November 5, 2026 the three cheapest of 36 used and refurbished 5090 listings tracked by GPU Poet averaged $6,450, barely below new.
Who should buy which
- Buy the RTX 5090 if your models are 32B or smaller (Qwen3 32B, Gemma 3 27B, gpt-oss-20b, FLUX image generation) and you are one user. It is as fast as the PRO 6000 on those models and costs less per card.
- Buy the RTX PRO 6000 if you want 70B to 120B models on one card, need 8-bit quality on 32B models, serve a small team, or fine-tune larger models. Per GB it is now the cheaper card.
- Buy the Max-Q version if you plan to add a second or third card later.
- Buy neither if you mainly want to hold big models and can accept lower speed. A 128 GB unified-memory box costs less; see the alternatives below.
Cheaper ways to get the same result
- Two used RTX 3090s (48 GB). Enough for Llama 3.3 70B at 4-bit with full CUDA support, at about $3,290 for the pair (two cards at GPU Poet’s $1,643 three-cheapest-listing average, November 5, 2026). Slower than either card here and much louder. Our best GPU for local LLMs guide covers the used-card risks.
- A 128 GB unified-memory box. A Strix Halo mini PC, a Mac Studio M5 Max or a DGX Spark loads gpt-oss-120b for less than a PRO 6000, but at 256 to 614 GB/s instead of 1,792 GB/s. Our DGX Spark alternatives guide compares them with published benchmarks.
- AMD Radeon AI PRO R9700 (32 GB). The same 32 GB as a 5090 at a fraction of the price, with 640 GB/s and ROCm instead of CUDA. See our best GPU for AI guide.
The bottom line
The RTX 5090 and RTX PRO 6000 are the same chip with the same bandwidth, so the choice is about memory, not speed. For one user running models up to 32B, the RTX 5090 is as fast as the PRO 6000 in every published llama.cpp result we found. For 70B and 120B-class models on a single card, for serving several users, or for ECC and MIG, the RTX PRO 6000 is the only consumer-reachable option, and at late-2026 prices it is also the better deal per GB of VRAM. Whichever you choose, check the live price first: both cards have more than doubled in price since launch.
Sources: NVIDIA RTX 5090 and RTX PRO 6000 Blackwell spec pages; Lenovo Press product guide LP2364 (RTX PRO 6000 Max-Q); Hardware Corner GPU ranking for local LLMs (Allan Witt, updated Dec 9 2025) and RTX PRO 6000 benchmark page (March 2026); CloudRift LLM inference benchmarks (Oct 9 2025); Tom’s Hardware RTX PRO 6000 price reports (2026) and GPU price index (Nov 3 2026); Thunder Compute RTX PRO 6000 pricing (Nov 2026); Digital Trends (Oct 17 2026); GPU Poet (Sep and Nov 5 2026); OpenAI gpt-oss model card.