Dell Pro Precision 9 T4 Review: Two RTX PRO 6000s and Memory Bandwidth for Local AI
Dell's Pro Precision 9 T4 with two RTX PRO 6000 Max-Q cards nearly doubles local AI serving. Its Xeon leads on memory bandwidth but trails on per-core work.

The Dell Pro Precision 9 T4 is a single-socket tower aimed at a narrow but growing group: teams that want to run, embed and fine-tune models on hardware they control, including models too big for a single 96GB card. The review unit pairs an 86-core Intel Xeon 698X with 256GB of DDR5-6400 ECC memory and two 300W RTX PRO 6000 Blackwell Max-Q cards. My read is that the second GPU is the real story here, because it changes what this box can do for AI work, while the processor choice creates a split result that buyers need to understand before they sign off.


Configured the way the review unit is built, the T4 lists at about $66,600 on Dell.com. That price is the number to keep in mind through the rest of this piece, because it sits well above the Precision 7875 that serves as the closest comparison, and the T4 wins some tests and loses others.
Where the T4 Sits in Dell's Lineup
Dell positions the T4 as the flagship single-socket tower for organizations that need maximum compute density and expandability. AI development, engineering simulation and scientific computing lead its list of target workloads. It sits above the Dell Pro Precision 9 T2, which uses the same Granite Rapids-WS Xeon 6 workstation platform in a smaller chassis with one GPU. The T4 adds the eight-channel processors at the top of the range, up to 86 cores, and room for up to four graphics cards, depending on the model.
The processor line runs from the 12-core Xeon 634 through the 64-core Xeon 696X to the 86-core Xeon 698X in the review unit. That chip carries 336MB of cache, a 2.0GHz base clock, a 4.8GHz maximum and a 350W rating. Memory goes up to 4TB of DDR5 RDIMM across 16 slots, and the review unit fills eight of them with 32GB modules running at 6400MT/s.
Graphics options start with a single RTX A400. Dell's configurator lists four RTX A1000 cards, three RTX PRO 4000 Blackwell cards, two 300W RTX PRO 6000 Max-Q cards, or one 600W RTX PRO 6000, depending on the model. That spread matters for buyers, because the 600W card and the Max-Q pair represent very different thermal and power envelopes in the same chassis.
Pricing works out roughly like this. The base configuration starts at $5,925 on Dell.com. The review build adds $8,300 for the Xeon 698X, $15,000 for 256GB of memory and $30,820 for the pair of RTX PRO 6000 Max-Q cards. Storage, networking and the 36-month ProSupport upgrade, which Dell adds automatically at this level, come on top of that.
Build, Power and Serviceability
The T4 is a full tower measuring 449.8mm tall, 203.7mm wide and 617.5mm deep. Dell lists its weight at 24.1kg minimum and 33.6kg maximum depending on configuration. The review unit uses the 2400W Platinum power supply, which Dell's configurator selects for the dual RTX PRO 6000 build. Lighter configurations default to a 1500W supply.


Serviceability is one of the T4's stronger points. The side panel opens with a tool-less latch, and the 2400W power supply slides out of the top rear of the chassis by hand. Graphics cards and SSDs need a Phillips screwdriver. Dell marks the parts that come out without tools with blue touch points, and the reviewers found the platform easy to repair in the field. For a machine at this price, that counts for more than it would on a cheaper tower, because downtime on a box that's running inference for a team is expensive in its own right.

Cooling is where I'd give Dell the most credit. Air moves front to back, and large form-fitted cowls direct it down and across the processor and memory. The same cowls keep the PCIe area on a separate channel, so the two 300W graphics cards draw from their own air stream and aren't breathing air already warmed by the 350W Xeon. Two fans sit between the processor zone and the card bay, and the fan assembly lifts out for memory access. It's the best cooling execution the reviewers have seen from Dell in this class of tower.

Expansion, Storage and Connectivity
Expansion runs through six full-height PCIe slots. Two Gen5 x16 slots carry the GPUs, and the rest are a Gen5 x8, a Gen4 x8, and two Gen4 slots with x8 connectors wired at x4. Storage covers three internal M.2 Gen5 slots and one Gen4 slot, two internal 3.5-inch SATA bays and a front FlexBay. The FlexBay holds up to four NVMe drives or two 3.5-inch drives, with an optional slim optical drive. The review unit boots from an internal 2TB Gen5 drive and carries a second 2TB Gen5 drive in an internal M.2 slot.

The front panel has a headset jack, two 10Gbps USB-A ports, a 20Gbps USB-C port with PowerShare, a 10Gbps USB-C port and an optional SD card slot. The rear adds 10GbE and 1GbE RJ-45 ports, four 10Gbps USB-A ports, two 10Gbps USB-C ports, a 20Gbps USB-C port, a serial port, two PS/2 ports and line-out. The review unit adds an Intel E810-XXV dual-port 25GbE SFP28 adapter, a MediaTek Wi-Fi 7 MT7925 card with Bluetooth 5.4, a media card reader and Dell's dual USB-C add-in card. If you plan to attach the T4 to a 25GbE fabric, that add-in adapter is the one to spec.

Security and management features include TPM 2.0, Dell SafeBIOS, chassis intrusion detection, self-encrypting drive support, BIOS secure erase and Intel vPro Enterprise. Dell says the T4 uses up to 85 percent recycled plastic in key components, 90 percent recycled aluminum in the handles and 11 percent recycled steel in the chassis, and the configuration is ENERGY STAR qualified.

Test Setup and Comparison Systems
The testing ran on Linux. All four systems ran Ubuntu 26.04.1 LTS on kernel 7.0.0-43 with NVIDIA driver 595.99.02, CUDA 13, llama.cpp build b11716 and Ollama 0.40.3, using the Linux performance power profile. Results came from three harnesses: sr-client for local AI and client workloads, sr-harness for compile, compression, crypto, web serving and y-cruncher, and sr-suite r5.28 for GROMACS, OpenFOAM, ClickHouse, Valkey and memory bandwidth. Each figure is the median of repeated passes. All four systems ran with the BIOS thermal mode set to Ultra Performance. The T4 ran BIOS 1.2.0, and the harness versions differed slightly across machines, so keep that in mind if you try to reproduce numbers.
Every comparison system is a Dell tower with RTX PRO 6000 Blackwell graphics. The Dell Precision 7875 is the closest match. It pairs the 96-core Threadripper PRO 9995WX with the same two Max-Q cards and 512GB of DDR5-5600. The Dell Pro Precision 9 T2 runs a 24-core Xeon 658X with one Max-Q card and 128GB of DDR5-6400 on four channels. The Dell Pro Max Tower T2 runs a 24-core Core Ultra 9 285K with a single 600W RTX PRO 6000 and 128GB of DDR5-4400 on two channels. That last system shows what the higher power limit adds on a per-GPU basis.
Local AI Serving: Where the Second GPU Pays Off
Every system here carries at least one 96GB RTX PRO 6000 Blackwell, so the single-GPU results mostly measure the card, its power limit and how quickly the platform feeds it. The T4's two cards are what set it apart. The two-GPU tests ran each model three ways where it applied: two independent servers, one per GPU, for throughput; one model split by layer across both cards; and, for MiniMax M2.7, one card with the rest of the model in system memory.
On one GPU, the T4 lands where the card's 300W limit puts it. It generated 166.9 tokens per second for a single user on gpt-oss 120B and peaked at 424.8 tokens per second across 64 concurrent requests. That's within 1 to 9 percent of the Precision 7875's identical Max-Q card on all four models, and level with the Pro Precision 9 T2. The 600W card in the Pro Max Tower T2 runs 13 to 25 percent faster at peak on every model, reaching 520.7 tokens per second on gpt-oss 120B.

Running one server per GPU is where the second card earns its keep. Peak output rose by 1.85 to 2.02 times across the four models. On Ministral 3 3B it went from 582.5 to 1,176.1 tokens per second, and on the 8B model from 383.6 to 765.7. On Qwen3.8 27B it climbed from 158.1 to 310.8. gpt-oss 120B scaled 1.85 times to 787.9 tokens per second on the T4, while the Precision 7875 scaled 1.98 times to 905.8. The Precision 7875 reached its gpt-oss peak at 128 concurrent requests, and the T4 peaked at 64.

Splitting gpt-oss 120B by layer across both cards is the less efficient way to use a second GPU when the model already fits on one. Single-user output was 170.6 tokens per second, within 3 percent of one card. Peak output reached 539.2 tokens per second, 27 percent above one GPU but well short of the 787.9 that two replicas delivered. The split layout earns its place for models that won't fit on one card, which brings us to MiniMax M2.7.
The full sweep on gpt-oss 120B, with 8,192-token prompts and 1,024-token outputs, shows how the second card changes the curve. Total throughput counts prompt and output tokens together. One GPU climbed from 1,502 tokens per second at one request to 2,760 at eight and 3,823 at 64, where each user received 6.6 output tokens per second. With one server per GPU, total throughput reached 4,820 tokens per second at eight concurrent requests and 7,091 at 64. At 64 requests that's 1.85 times the single-card figure, and each user got 12.3 output tokens per second. The split-by-layer configuration delivered 4,853 total tokens per second at 64 requests, but it cut time to first token for a single user from 1,189ms to 719ms.

Pick the layout based on the workload. Replicas win on throughput for a shared endpoint serving many users. A split model wins when you want a faster first token for a single user, or when the model is too large for one card.
MiniMax M2.7: A Model Too Big for One Card
MiniMax M2.7 is a 230-billion-parameter mixture-of-experts model. The 4-bit Q4_K_M quantization used here takes 129.3GB, which is more than the 96GB on one RTX PRO 6000. On one card with the remainder in system memory, the T4 generated 30.3 tokens per second for a single user and took 16.2 seconds to return the first token. Split across both cards, single-user output rose to 83.7 tokens per second, 2.8 times the one-card rate. Time to first token fell from 16.2 seconds to 3.1 seconds, and peak throughput climbed from 35.5 to 127.3 tokens per second.

The one-GPU results show how much the host platform matters once a model spills into system memory. The T4's eight channels of DDR5-6400 put it 34 percent ahead of the four-channel Pro Precision 9 T2 at 22.6 tokens per second. It's 1.9 times the dual-channel Pro Max Tower T2 at 15.8, even though the Pro Max Tower T2's 600W card is faster on every model that fits in VRAM. The Precision 7875 edged the T4 at 32.8 tokens per second, despite having 35 percent less memory bandwidth. Bandwidth alone doesn't explain that result, and the review reports the 7875's lead as measured without isolating its cause.
Model Loading and Cold Starts
Cold model loads favor the Gen5 drives. The T4 loaded gpt-oss:120b from its SK hynix Gen5 SSD in 15.16 seconds, matching the Pro Precision 9 T2 and well ahead of the 25.80 and 28.18 seconds on the Gen4-equipped Precision 7875 and Pro Max Tower T2. Warm loads from the page cache reverse that order. The Precision 7875 managed 5.01 seconds against 7.49 on the T4. Loading a copy of gpt-oss:120b onto each GPU at once took 26.47 seconds on the T4.
For MiniMax M2.7, the T4 took 40.5 seconds to cold-load into one GPU plus system RAM and 27.6 seconds when split across both cards. That's faster than the 65.8 seconds on the Precision 7875 in the same configuration. If your team swaps models often, the cold-load number is worth more to you than the warm one.
Retrieval at Scale and Fine-Tuning
The RAG test hides SQuAD questions among 1,002,067 Wikipedia passages, embeds them with nomic-embed-text, builds an HNSW index and answers with Phi-4-mini. On one GPU, the T4 embedded the corpus at 1,126 passages per second, which took 890 seconds of wall time. The Precision 7875 managed 1,196 passages per second and 838 seconds. The 600W card in the Pro Max Tower T2 was the fastest single GPU here at 722 seconds. Using two GPUs cut the T4's time to 464 seconds, a 1.92 times speedup. The Precision 7875 finished the two-GPU run in 423 seconds.

Index build and vector search run on the CPU, and that's where the T4 leads. It built the HNSW index in 27.3 seconds against 34.8 on the Precision 7875 and 118.6 on the Pro Max Tower T2. At ef 64 it searched 98,503 queries per second, 1.51 times the 7875 and 5.6 times the Core Ultra system. At ef 256 it held 32,149 queries per second with 97.59 percent recall against exact search, and answers came back in a median of 0.123 seconds, although the 7875 answered faster at 0.082 seconds.

Fine-tuning is where the T4 loses ground on a single card. The one-epoch LoRA test trains Ministral 3 8B in BF16 on Dolly 15k across 1,849 steps and scores 200 held-out examples every 100 steps. Every system started at a held-out loss of 1.534 and finished between 1.167 and 1.172, so the comparison comes down to time. One T4 GPU finished the epoch in 31.5 minutes at 1,584 tokens per second, within 3 percent of the Pro Precision 9 T2 on the same Max-Q card. The Precision 7875 took 27.2 minutes, and the 600W card in the Pro Max Tower T2 took 22.3 minutes.

Held-out loss on the T4 fell from 1.534 to 1.219 in the first 100 steps, which took 1.6 minutes, and then eased to 1.169 by the end of the epoch. Adding the second GPU cut the T4's epoch to 19.7 minutes at 2,528 tokens per second, a 1.60 times speedup. That beat the single 600W card's 22.3 minutes. The Precision 7875 completed its two-GPU epoch in 18.1 minutes, a 1.50 times speedup, and the two-GPU runs ended at held-out losses of 1.167 on the T4 and 1.168 on the 7875.

The two cards run data-parallel with two examples per GPU and gradients accumulated to a global batch of eight. The single-GPU run takes four passes and the pair takes two, so both see the same examples at each step and land on the same loss. With only two short examples per GPU in each pass, the second card has little work to absorb, and gradient exchange takes a bigger share of every step. Memory was never the constraint. The epoch peaked at 19.5GiB on one card and 19.7GiB per card on two, a fraction of the 96GB on each RTX PRO 6000.
Coding Agents, Small RAG and Image Generation
The T4 passed every coding agent run at every model tier. That includes all three gpt-oss:20b runs at a 64K context, where the other three systems each passed two of three. The Precision 7875, Pro Precision 9 T2 and Pro Max Tower T2 each had one run flagged by the suite for failing at least one of its 11 tests, and the T4 generated 161.5 tokens per second on gpt-oss:120b. The 600W Pro Max Tower T2 was fastest at 183.2 tokens per second.
On a 20,963-passage RAG ingest, the T4 managed 304.6 passages per second, behind the 433.8 on the Precision 7875 and the 392.1 on the Pro Precision 9 T2. Answer generation with Phi-4-mini ran at 372.5 tokens per second on the T4, against 410.0 on the 7875. Image generation runs on a single GPU, so the T4's 40.96 images per minute on FLUX.2 klein 4B in BF16 sits with the other Max-Q systems and trails the 600W card's 54.25.
CPU Performance: Cores Versus Per-Core Speed
The CPU results set the T4's 86 Xeon cores against the Precision 7875's 96 Threadripper cores, with the two 24-core T2s as reference points. The Threadripper brings 192 threads to the Xeon's 172. The pattern is consistent. The T4 leads where memory bandwidth decides the result, and the 7875 leads where core count and clock speed do.

In compression, rendering and compiling, the T4 scored 619.8 GIPS in 7-Zip and 961.8 samples per minute in Blender Monster. Those figures are 2.2 and 2.4 times the 24-core Pro Precision 9 T2. The Precision 7875 ran 37 percent ahead in 7-Zip and 30 percent ahead in Blender Monster. It also finished the LLVM 21 build in 43.7 seconds against 58.4 on the T4, and the Linux kernel allmodconfig build in 264 seconds against 312. Per-core speed is the T4's weak point. Its one-thread 7-Zip score of 8.85 GIPS was the lowest in this group.
HandBrake's x265 encode is the outlier. The T4 logged 31.4 fps, while the other three systems landed between 40.0 and 41.7 fps despite very different core counts. That suggests the encoder isn't scaling with cores at this preset, and the T4's lower per-core speed on the threads x265 uses explains the gap. During the encode, the T4 averaged about 1GHz across its 172 threads at 51C to 64C, so most of the chip sat idle. SVT-AV1 scaled better, at 159.4 fps on the T4 against 255.8 on the 7875.


Bandwidth-bound work is where the T4 shines. Eight channels of DDR5-6400 gave it 328,329 MB/s in STREAM Triad, 53 percent above the Precision 7875's DDR5-5600 and 2.3 times the four-channel Pro Precision 9 T2. The T4 solved the 3-million-cell OpenFOAM drivAer case in 293 seconds, against 556 on the 7875 and 826 on the Pro Precision 9 T2. The Pro Max Tower T2 didn't finish within the suite's 150-minute limit. The T4 also ran the 22.6-million-cell case in 6,851 seconds.

GROMACS benchPEP, the large all-core molecular dynamics case, ran at 1.243 ns/day on the T4 against 1.034 on the 7875. The OpenBLAS LU solve at N=20000 reached 1,660 GFLOPS against 1,172. y-cruncher computed 10 billion digits of Pi in 53.9 seconds against 83.7 on the 7875, and 25 billion digits in 155.6 seconds against 259.6. Dense compute that stays in cache favors the Threadripper, which led OpenBLAS DGEMM at 4,745 GFLOPS against the T4's 3,664.



Databases and services show the other side of the trade. ClickHouse ran the 43 ClickBench queries in 3.21 seconds hot and 6.45 seconds cold on the T4. That's 0.33 seconds behind the 7875 on the hot run and 0.64 seconds ahead on the cold run. Valkey favors fast cores over many cores. The T4's 666,714 SET requests per second trailed all three comparison systems, with the Core Ultra 9 285K in the Pro Max Tower T2 leading at 1,719,108. The T4's Stockfish run reached 68.21 million nodes per second with all 172 threads loaded, roughly 1.5 times the 48-thread Pro Precision 9 T2 and well under half the 7875's 168.27. The first full run timed out, and the figure above comes from a standalone retest. The review doesn't isolate the cause of the weak scaling, so I'd treat that result as a flag rather than a verdict.
Storage Performance
The T4's 2TB SK hynix Gen5 boot drive delivered sequential reads of 14,152 MB/s and writes of 10,078 MB/s in fio. Its sequential read speed was roughly twice that of the Gen4 drives in the Precision 7875 and Pro Max Tower T2. Random performance peaked at 2,306,346 read IOPS and 1,830,776 write IOPS. The Pro Precision 9 T2 posted 20,914 MB/s on sequential reads, which the suite flagged for run-to-run variance after a rerun. That figure is higher than a single Gen5 x4 link carries, which fits that system's striped array of Samsung PM9E1 drives. The T4's second 2TB drive wasn't part of the storage run, so the single-drive results are what you can compare directly.

Thermals and Clocks Under Load
With the BIOS in Ultra Performance mode, the Xeon 698X peaked at 93C in 7-Zip and 96C in GROMACS. The clock figures are averages across all 172 logical CPUs, which is why they can look low on heavily loaded jobs. Blender and Stockfish loaded every thread and held 2.5GHz to 2.6GHz, above the chip's 2.0GHz base. GROMACS and OpenBLAS load 64 and 86 physical cores and leave the rest of the threads idle near 0.8GHz, which pulls their averages down to about 1.3GHz. GPU-led work kept the processor cooler. During LLM serving, fine-tuning and RAG, the mean CPU temperature stayed in a 61C to 65C range, since those tests lean mostly on the cards.

What the Numbers Add Up To
The T4 is built for local AI work, and the second RTX PRO 6000 is where it shows. Two Max-Q cards raised peak serving throughput 1.85 to 2.02 times across four models and embedding throughput 1.92 times. They cut a LoRA epoch on Ministral 3 8B from 31.5 to 19.7 minutes. They also turned MiniMax M2.7, a model too large for one card, from 30.3 to 83.7 tokens per second for a single user. On gpt-oss 120B, one server per GPU reached 7,091 total tokens per second at 64 requests, while the split layout returned a single user's first token in 719ms against 1,189ms.

The Xeon 698X's eight memory channels gave the T4 a 53 percent bandwidth lead over the Threadripper PRO 9995WX in the Precision 7875. That bandwidth carried the T4 to faster OpenFOAM, GROMACS benchPEP, y-cruncher and HNSW search results than the 7875, and the Gen5 boot drive loaded large models faster on cold starts. The chassis deserves credit too, with the cooling layout and tool-less power supply access.
The trade-offs are real. The 7875 stays ahead on core-bound rendering, compression and compile work, and it leads SVT-AV1 by about 60 percent. The Xeon's per-core speed trails every system in this group, and that shows in Valkey and other lightly threaded services. Single-GPU fine-tuning is also slower on the T4, where the LoRA epoch took 31.5 minutes against 27.2 on the 7875 and 22.3 on the 600W card in the Pro Max Tower T2. Every system finished that run within 0.005 of the same held-out loss, so the difference is purely time.
At about $66,600 as configured, the T4 suits teams that want to serve, embed and fine-tune models on premises, including models that need more than 96GB of GPU memory, while running memory-bound simulation and analytics on the same box. If your work is mostly rendering and compilation, the Threadripper platform will serve you better. If you need only one GPU, the same card fits in the smaller Pro Precision 9 T2. Before you commit, I'd check whether your models actually need the split layout, whether your services are latency-sensitive on single threads, and whether a 25GbE uplink matters to your environment. Those three answers will tell you more about the right machine than any single benchmark.