Dell Pro Precision 9 T4 Runs MiniMax M2.7 at 83.7 tok/s on Two GPUs

That figure is 2.8 times the 30.3 tokens per second the same tower managed with the model on one card and the remainder spilled into system memory. StorageReview ran both tests with the llama.cpp server, build b11716, on Ubuntu 26.04.1 with NVIDIA driver 595.99.02 and CUDA 13, using the 4-bit Q4_K_M quantization of the model and one request at a time. Its unit, an 86-core Xeon 698X with 256 GB of DDR5-6400 and two 300 W Max-Q cards, lists at about $66,600 on Dell.com as configured, the lab reported. Voltgrid found no other lab that had published local-AI measurements of the T4 as of Nov. 14, which is why this is a product file and not yet a review.

Dell Pro Precision 9 T4 Specs and Price

The T4 is what StorageReview calls Dell's flagship single-socket workstation tower. Dell announced it in March alongside the larger Pro Precision 9 T6, and Notebookcheck reported on March 16 that it would go on sale in May. According to that launch report, it takes Intel Xeon workstation processors up to the 86-core Xeon 698X, rated at 350 W, up to 4 TB of RAM with workstation graphics, and a 1,500 W or 2,400 W power supply depending on the build.

Dell gave the T4 three GPU layouts, Notebookcheck reported: one 600 W card, two 300 W cards, or four 50 W cards. For local AI that choice is the product. StorageReview found Dell's configurator lists either one 600 W RTX PRO 6000 or two 300 W RTX PRO 6000 Max-Q cards, each with 96 GB of GDDR7. Memory goes in 16 slots over eight channels, up to 4 TB of DDR5 RDIMM, and CPU options start at the 12-core Xeon 634.

Prices, as StorageReview read them on Dell.com: $5,925 for the base configuration, plus $8,300 for the Xeon 698X, $15,000 for 256 GB of memory and $30,820 for the pair of Max-Q cards, before storage, networking and a 36-month ProSupport upgrade that Dell adds automatically at that level.

Why Two Max-Q Cards Matter for MiniMax M2.7

A large language model generates each token by reading its active weights from memory, so the model must sit in fast memory to run quickly. One RTX PRO 6000 holds 96 GB in GDDR7 on the card. MiniMax M2.7 is a 230-billion-parameter mixture-of-experts model, and its Q4_K_M file needs 129.3 GB. On one card, the overflow lives in the host's DDR5, and StorageReview measured 30.3 tokens per second for one user with a 16.2-second wait for the first token. Split by layer across both cards, 192 GB of GDDR7 holds the whole model: 83.7 tokens per second, a first token in 3.1 seconds, and a peak of 127.3 tokens per second under concurrent load.

The host still counts once the model spills. In the same one-card-plus-RAM test, the T4's eight channels of DDR5-6400 put it at 30.3 tokens per second, against 22.6 on the four-channel Pro Precision 9 T2 and 15.8 on the dual-channel Pro Max Tower T2, which carries a 600 W card, StorageReview found. The T4 measured 328 GB/s in STREAM Triad, a CPU memory test, 53% above the Threadripper PRO 9995WX in Dell's Precision 7875. Yet the 7875 edged the T4 on the spill test at 32.8 tokens per second, and StorageReview said it had not isolated the cause, so bandwidth alone does not explain the result.

What One Max-Q Card Gives Up on Smaller Models

On models that fit in 96 GB, the T4's 300 W Max-Q cards trail Dell's 600 W card. With llama.cpp b11716 and one GPU, StorageReview measured gpt-oss 120B at 166.9 tokens per second for one request and 424.8 at the peak of 64 concurrent requests. The Pro Max Tower T2's 600 W RTX PRO 6000 reached 193.0 and 520.7 on the same model and tool, and was 13 to 25% faster at peak on every model that fits one card.

The second card buys back throughput rather than single-user speed on those models. Running one llama.cpp server per GPU lifted the T4's gpt-oss 120B peak to 787.9 tokens per second, 1.85 times one card, while splitting that model across both cards served one user at 170.6 tokens per second, within 3% of one card, and peaked at 539.2. Splitting did cut the time to a first token on 8,192-token prompts from 1,189 ms to 719 ms, StorageReview found. The Precision 7875, with the same two Max-Q cards, ran MiniMax M2.7 split at 85.0 tokens per second and gpt-oss 120B at a 905.8 peak with one server per GPU.

Why Tokens Per Second Needs a Test Label

A tokens-per-second figure describes one test, not one machine. The T4's 83.7 tokens per second is MiniMax M2.7 in 4-bit Q4_K_M, split by layer across two Max-Q cards, served to one user by llama.cpp build b11716. The same tower managed 30.3 with one card and the overflow in system memory.

The 787.9 tokens per second for gpt-oss 120B is a different kind of result: throughput summed across concurrent requests, with one llama.cpp server per card, not one user's speed. Only results from the same model, quantization, tool and number of users can be set against each other.

How it sits against the alternatives

Our RTX PRO 6000 review covers the card itself: 96 GB, a 600 W board power on the Workstation Edition, and a $16,000 listing on NVIDIA's US marketplace, shown out of stock when read on Aug. 10. Those vLLM serving figures were measured with a different tool and workload from the llama.cpp numbers here, so the two sets do not compare directly.

Above 192 GB, Dell's own route is the Dell Pro Precision with GB300, a DGX Station listed at $160,195 on Dell.com in its default configuration when read on Nov. 8, with 252 GB of HBM3e on the GPU and 496 GB of LPDDR5X on the CPU. For buyers weighing several cards against a GB300 tower, our look at DGX Station alternatives sets four RTX PRO 6000s against it on vLLM. The DGX Station review sets out NVIDIA's 1,600 W total system power rating and single-user speeds.

What is not known yet

Voltgrid found no lab other than StorageReview that had published local-AI measurements of the T4 as of Nov. 14, so every speed above comes from one lab, one llama.cpp build and the medians of repeated passes on one unit. No second outlet has confirmed the MiniMax M2.7 or gpt-oss 120B figures. The review does not give wall power under LLM load, so the cost in watts of the second card is not yet measured. StorageReview did not test vLLM or SGLang on the T4, which leaves its multi-user serving under those engines unknown.

The prices here are Dell.com's as StorageReview read them, and Dell's configured prices move. The next milestone is a second independent measurement, along with any change to the price of the dual Max-Q option; when those land, this file becomes the base for a full review.