The Dell Pro Precision 9 T2 is the Xeon variant of the single-GPU tower we reviewed one year ago. It introduces two key changes: Dell’s return of the Precision branding for workstations, and a new platform swapping the Core Ultra 9 285K for Intel Xeon 6 inside the same 32-liter enclosure. Our test unit carried a 24-core Xeon 658X, 128GB DDR5-6400 RDIMM ECC, 300W RTX PRO 6000 Blackwell Max-Q, and two striped 2TB E3.S Gen5 drives in the front FlexBay. This exact configuration lists for roughly $28,300 on Dell.com.

We tested across a matrix: the Pro Precision 9 T2 ran our full benchmark suite with its stock Max-Q GPU and again with the 600W RTX PRO 6000 Blackwell from the Pro Max Tower T2. The Core-based T2 also used both cards. This creates four result columns, isolating either CPU platform or GPU power limit. The benchmark suite matches the set used for our recent HP Omnibook Ultra 16 review, plus several newly added tests.

Xeon prioritizes throughput at the cost of latency. The Pro Precision 9 T2 hit 12,322 in Cinebench 2026 multi-core versus 8,856 on the Core T2, built LLVM in 140.2 seconds against 182.5, and reached 122,233MB/s in STREAM Triad versus 33,003. Its single-core Cinebench score of 454.1 lagged the 285K’s 564.4, while its Speedometer result of 36.8 fell short of 46.3. The 600W GPU delivered 7–20% higher LLM decode performance and 20–39% faster prompt processing and diffusion workloads on both towers, but reduced the Xeon’s all-core CPU performance by 4–5%.

Key Takeaways

Xeon 6 is a throughput-focused single-socket tower platform. It scored 12,322 in Cinebench 2026 multi-core (39% ahead of the Core Ultra 9 285K in the identical chassis), 289.6 GIPS in 7-Zip versus 174.8, 1,638 GFLOPS in DGEMM versus 768, and 3.7× the memory bandwidth over four DDR5-6400 channels. The 285K retains superior per-thread performance: the Xeon trails it by 20% in Cinebench single-core, 21% in Speedometer, and 20–31% in single-threaded Python, with core-to-core latency of 56.2 ns compared to 38.1 ns.

GPU performance dominates AI outcomes. The 600W card justifies its higher power budget for prompt work: MLPerf Llama 3.1 8B ran at 230.9 tokens/s with Max-Q and 252.8 tokens/s with the 600W model; time-to-first-token dropped from 0.082 s to 0.059 s. Qwen-Image 20B rendered a 1024×1024 image in 19.7 seconds versus 26.1 seconds on Max-Q. On the Xeon system, fitting the 600W GPU reduced sustained all-core CPU clocks from 4,274 MHz to 4,178 MHz and cut Cinebench multi-core from 12,322 to 11,705. No such penalty appeared on the Core T2.

The FlexBay E3.S setup uses VROC by Graid for RAID0. This front bay supports two E3.S Gen5 drives, a first for a tower workstation in our lab. Dell’s RAID0 boot option combines them into a 4,097GB volume. The striped array delivered 9,332MB/s sequential read and peaked at 840.1K 4K random read IOPS, while the single Gen5 M.2 inside the Core T2 hit 1.78M IOPS.

From Pro Max to Pro Precision

Dell retired the Precision brand in early 2025, rebranding all workstations as Pro Max, before restoring it. On October 7, Dell renamed its GB10 and GB300 desktop workstations from Pro Max to Pro Precision. The Pro Max line remains for Core-based systems, so our previously reviewed Core Ultra tower stays the Dell Pro Max Tower T2, while this Xeon unit is the Dell Pro Precision 9 T2 (model PW9T2260). The T4 and T6 towers also shifted to Xeon 6 under the Precision badge, differentiated by expansion capability.

The Pro Precision 9 T4 supports the same CPU lineup up to the 86-core Xeon 698X, offers 16 DIMM slots for up to 4TB DDR5-6400 across eight memory channels, accepts one 600W RTX PRO 6000 or dual Max-Q cards on a 1500W supply, with an optional 2400W PSU. The flagship Pro Precision 9 T6 features six expansion slots, a 2400W supply, and space for two 600W RTX PRO 6000 cards, three Max-Q units, or up to five RTX PRO 5000s. The T2 is the single-GPU, four-channel entry point of the family.

Externally, the two T2 towers are identical: same 387 × 187.7 × 438 mm 32L chassis, front bezel, FlexBay, and RTX PRO 6000 Blackwell configuration options. The difference lies on the mainboard. The Pro Max Tower T2 pairs the 24-core Core Ultra 9 285K with four unbuffered DDR5 DIMM slots. The Pro Precision 9 T2 uses Intel’s workstation-focused Xeon 6 platform, supporting registered ECC memory, extra PCIe lanes, and a 250W top-tier processor. Dell lists six Xeon 6 SKUs for the T2, ranging from the 150W 12-core Xeon 634 to our 250W 24-core 658X; upgrading from 634 to 658X adds $1,560.

The 658X runs 24 cores / 48 threads at 3.0 GHz base / 4.9 GHz boost with 144MB L3 cache, triple the 285K’s 36MB, and includes AVX-512 and AMX instructions absent on the Core Ultra chip. It has eight DDR5-6400 channels and 128 Gen5 PCIe lanes, though the T2 only exposes four DIMM slots and five expansion slots; the larger T4 and T6 unlock the remaining lanes.

The Core T2 boots from a single M.2, while our Pro Precision 9 T2 uses Dell’s FlexBay NVMe boot layout: two hot-swappable E3.S Gen5 drives behind the front panel, striped into a single 4,097GB RAID0 volume. E3.S is an enterprise form factor that supports far higher capacities than M.2 and allows front-side drive swaps without opening the chassis. This RAID stack runs VROC, taken over and relaunched by Graid Technology, marking the first workstation in our lab shipping with this implementation.

The 600W GPU requires adjacent slot clearance not needed by the Max-Q card. Our unit shipped with an Intel E810-XXV dual-port 25GbE adapter next to the GPU slot, which we had to remove to install the full-power 600W card. Buyers combining the 600W GPU with high-speed NICs must plan slot layout upfront; the 1500W PSU is the only practical pairing, a $40 upgrade above the standard 750W supply.

Build and Design

The Pro Precision 9 T2 shares the same tool-less tower chassis as the Pro Max model, with a full-height mesh front intake grille, side latch and captive thumbscrew for the side panel, and the FlexBay plus optional slim optical drive behind the lower bezel. Dell lists an empty chassis weight of 11.6kg, rising to 21.8kg fully configured; our 1500W PSU, Xeon cooler, and two-slot GPU put it near the upper end.

Internal airflow is carefully routed. A large cooling shroud cools DIMMs and routes CPU exhaust directly out the chassis, with an integrated shroud fan. The shroud rests on the GPU retention bracket and seals against the side cover. Internally, layout matches the Pro Max: CPU and four DIMM slots sit above the expansion zone, the GPU in the Gen5 x16 slot is secured by an end bracket, and the rear-mounted PSU uses a handle for tool-free removal. All five slots are full-height: slot 2 is the primary Gen5 x16 GPU slot, slot 1 Gen5 x4, the remaining three Gen4 (one x16, two open-ended x8) for NICs, storage adapters, or a second GPU. Three internal M.2 2280 Gen5 slots and two 3.5-inch SATA bays sit behind the E3.S front bay.

Front I/O includes a headset jack, 20Gbps PowerShare USB-C, 10Gbps USB-C, two USB-A ports, and an optional SD reader. Rear I/O provides audio line-out, 20Gbps and 10Gbps USB-C, two SmartPower USB 2.0, two 5Gbps USB-A, 1GbE, an optional module slot supporting a 5GbE NIC or extra USB-A ports, and an optional legacy serial port. Our unit added a MediaTek MT7925 Wi‑Fi 7 card in an internal M.2 2230 slot plus the Intel E810-XXV 25GbE adapter removed for 600W GPU testing.

Dell Pro Precision 9 T2 Specifications

Two pricing caveats apply. Dell’s online configurator currently lists FlexBay drives as 2TB M.2 2280 Gen5 modules at $1,700 each, while our review unit uses E3.S drives; we have asked Dell when the E3.S online option will appear. Also, few enterprise workstations sell at listed web price, so $28,300 represents a ceiling, with volume and contract pricing lower. Memory and GPU account for $19,090 of the total, and the 300W and 600W RTX PRO 6000 options cost the same. Dell quotes our exact review build at $31,700 including the E810 adapter, Wi‑Fi7, optical drive, and three-year ProSupport.

Test Setup and Methodology

All results run via sr-client 1.2, our scripted Windows benchmark suite using Dell factory Windows images, Dell drivers, and NVIDIA 596.71 across all four configurations. Settings were standardized: Windows power mode set to Best performance, sleep/display timers disabled, firewall and UAC prompts off, VBS and Memory Integrity retained as shipped. Each reported value is the median of repeated passes; tests with result spread above 2% rerun before acceptance. Our four configurations combine two towers and two GPUs, isolating CPU and GPU variables. The Core T2 boots from its single 2TB Gen5 M.2, while the Xeon T2 uses its two-drive E3.S stripe, so storage and model-load metrics reflect those native boot layouts.

Several benchmarks debut here. LLM serving uses llama-server across four models, sweeping concurrency and measuring throughput, time-to-first-token, and per-token latency. Image generation renders 1024×1024 outputs on FLUX.2 klein 4B, Z-Image Turbo 6B, and Qwen-Image 20B at Q4_0, Q8_0 and BF16, with PSNR/SSIM comparing quantized outputs to BF16 baselines. The coding agent test runs on Crush; the concurrent agent-plus-browser benchmark uses a revised three-pass Speedometer workflow. Three table markers are used: N/A where a test does not apply to hardware, Partial for accepted runs with one failed subtest. Neither workstation exposed CPU/GPU temperatures to our sensor logging, so we report average CPU clock frequencies instead of thermal tables.

Local AI Performance

With 96GB GDDR7 in every build, all local AI workloads fit fully in VRAM. Results split primarily by GPU power limit: memory-bandwidth-bound workloads see minimal Max-Q/600W separation, while compute-heavy tasks favor the higher-power card. CPU platform impact is limited and directional where visible.

MLPerf Client v2.0 runs Llama 3.1 8B and Phi-4-mini across five prompt categories, reporting geometric-mean time-to-first-token and token rate using llama.cpp CUDA. Agentic workloads run Llama3.1 8B as a data analyst and software engineering agent with multi-turn tool calls.

Identical GPUs deliver near-matching token rates across towers. The 600W card lifts Llama3.1 8B throughput ~9% and cuts its TTFT by 28%. Prompt processing is compute-bound while decode is bandwidth-bound, so TTFT improves far more than sustained generation speed. Agent workflows expose CPU differences: the Core T2 finished tool-call phases roughly 1.5 seconds faster, thanks to the 285K’s higher single-core clock, even with matching model generation speeds.

Model loading tests in Ollama measure cold (file cache flushed) and warm load times to first token, drive read throughput during cold ingest, GPU memory arrival time, and load efficiency. We tested gpt-oss:20b, Qwen3.6-35B-A3B and gpt-oss:120b. The Core T2 loads every model faster because its single Gen5 M.2 sustains higher throughput under Ollama’s memory-mapped loading pattern than the E3.S RAID0 stripe. Warm loads narrow the gap substantially once weights sit in system RAM; GPU power limit shows no consistent loading benefit.

Crush, our terminal coding agent, runs fixed repository test suites with up to three retries to pass 11 unit tests. We track median wall time, pass count, and generation throughput. The concurrent agent + Speedometer test quantifies mutual slowdown. Most Qwen workloads passed all runs, while gpt-oss:20b occasionally failed one of three attempts, a model-level behavior unrelated to hardware. Running the agent and browser benchmark concurrently created almost no slowdown on either tower, though the Xeon’s lower Speedometer baseline persists from weaker single-thread performance.

Our RAG benchmark embeds 20,963 SQuAD v1.1 passages with nomic-embed-text, retrieves top-5 passages for 2,000 queries, and answers 400 questions with Phi-4-mini, scoring exact match and F1. Embedding and retrieval carry CPU-bound steps, favoring the 285K and making the Core T2 faster through the full pipeline. The 600W GPU lifted prompt processing throughput by 25–30% on both platforms, while retrieval accuracy was identical across all four builds.

LLM serving uses llama-server on Ministral 3 3B / 8B, Qwen3.8 27B and gpt-oss 120B, sweeping concurrency within available VRAM and reporting total throughput, output throughput, median TTFT and per-token latency. On 96GB VRAM, gpt-oss 120B supports 64 concurrent slots, Qwen3.8 27B eight slots, and the two Ministral models four slots each. The dense Qwen3.8 27B shows the largest 600W GPU advantage (~24% at eight concurrent users), while the MoE gpt-oss 120B runs faster despite its larger footprint. At the highest concurrency, the Core T2 held a 7–11% lead on gpt-oss 120B due to faster CPU-side batch handling.

Image generation is strongly compute-bound, so the 600W GPU delivers major speedups: Qwen-Image 20B Q8_0 rendered at 26.1 s/image on Max-Q and 19.7 s/image on the 600W card (+33% images per minute). Z-Image Turbo gained 32% and FLUX.2 klein 21%. CPU platform barely impacts diffusion render speed. Quantization fidelity is hardware-independent: Q8_0 closely matches BF16 quality, while Q4_0 degrades visibly. With 96GB VRAM, BF16 weights fit comfortably and often run as fast or faster than Q8_0 by avoiding runtime dequantization overhead.

CPU Performance

The two CPUs split benchmarks cleanly. Xeon 658X dominates fully parallel, high-bandwidth, vector-heavy workloads by 27–150%. The Core Ultra 9 285K wins single-threaded and lightly threaded tests by 13–31% from its higher P-core frequency and lower inter-core latency. The 600W GPU reduces Xeon all-core scores by 4–5% due to lower sustained clocks, with no measurable impact on the Core T2.

In Cinebench 2026, the Xeon hit 12,322 multi-core versus 8,856 (+39%). Single-core reversed: 454.1 vs 564.4. The 658X sustained stable performance across the 30-minute loop, but the 600W card pulled steady-state clocks down from 4,274 MHz to 4,178 MHz.

7-Zip all-thread throughput reached 289.6 GIPS for Xeon against 174.8 GIPS (+66%). Single-threaded, the 285K led 10.82 to 8.74 GIPS.

Blender Cycles renders and the LLVM compile test favored Xeon substantially; LLVM built in 140.2 s versus 182.5 s. The 600W GPU hit its largest CPU penalty on Blender, dropping Xeon render rates 7–10%. HandBrake results varied by encoder: x265 performed nearly equal, while SVT-AV1 ran far faster on the 285K, limited by per-thread encoder parallelism rather than AVX-512 support.

Stockfish chess, primesieve, PyPerformance Python, and Speedometer confirm the single-thread gap. primesieve is unusual among multi-threaded tests in favoring the 285K because its working set fits tightly in cache.

GROMACS molecular dynamics, OpenBLAS BLAS/LU, and STREAM show Xeon’s AVX-512, large L3, and four-channel DDR5-6400 advantage. DGEMM hit 1,638 GFLOPS vs 768 GFLOPS, STREAM Triad 122,233 MB/s vs 33,003 MB/s (3.7× bandwidth). This comes with higher memory and inter-core latency: 148.3 ns memory latency vs 135.2 ns, and 56.2 ns average core-to-core latency vs 38.1 ns.

SuperTuxKart gaming benchmarks, GPU-unconstrained at 1080p Ultra, follow single-thread CPU performance: Core T2 averaged higher FPS, though the Xeon tower maintained a higher steady frame rate with fewer dips.

Average clock data across workloads shows the 600W GPU suppresses Xeon frequencies across most all-core benchmarks. The exact limiting factor—thermal headroom or platform power cap—requires further testing with external logging. For CPU-heavy continuous workloads, Max-Q is the faster configuration on the Xeon tower.

Storage Performance

Our fio storage tests ran against the boot volume at 25GB footprint, measured on file system and raw device. The E3.S RAID0 stripe is not faster than the single SK hynix PCB01 Gen5 M.2. Sequential reads hit 9,332 MB/s vs 11,377 MB/s, while 4K random reads peaked at 840.1K IOPS vs 1.78M IOPS. Tail latency degraded sharply on the RAID array: 15.3 ms p99.99 random read latency at 16 jobs versus only 0.68 ms on the single drive. The VROC RAID layer adds CPU overhead on every I/O and hides individual drive SMART telemetry.

This does not invalidate the FlexBay design. Front-serviceable, hot-pluggable E3.S supports capacities far beyond M.2 and delivers a convenient 4TB boot pair. Users running small-block random I/O, databases, or cached model directories should consider a single E3.S or internal M.2 slot instead of striped RAID0.

Conclusion

The Dell Pro Precision 9 T2 is a Pro Max Tower T2 upgraded to workstation-class Xeon silicon within the identical 32-liter enclosure. The Xeon 658X delivers major gains in multi-core throughput, memory bandwidth, dense floating-point math, and ECC reliability, trading away single-thread responsiveness. That tradeoff matters for interactive viewport and terminal workflows.

GPU choice defines AI performance. Max-Q and 600W RTX PRO 6000 share 96GB VRAM and stay within 7–20% on LLM decode. The 600W card shines on prompt ingestion and diffusion generation, cutting TTFT substantially and speeding image renders by 20–34%. It costs 4–5% Xeon CPU throughput and occupies an adjacent PCIe slot, crowding out high-speed NICs. Max-Q is the balanced pick for mixed CPU render/simulation + local AI work; the 600W card suits dedicated model serving and image rendering, paired with the mandatory 1500W PSU.

E3.S in the FlexBay is a welcome tower workstation feature, and VROC by Graid is a functional RAID stack, albeit with worse small-random latency than a standalone Gen5 M.2. For users needing massive local flash capacity, E3.S unlocks high-capacity options previously limited to external enclosures.

At roughly $28,300 list price before enterprise contract discounts, the tested Pro Precision 9 T2 runs about $4,500 above an equivalently configured Pro Max Tower T2. Part of that gap comes from the second FlexBay SSD; the remainder buys the Xeon workstation platform. It delivers single-socket server-like throughput in a standard desktop tower footprint, justifying Precision’s return to the brand lineup. The Pro Precision 9 T2 is our current Best Overall Desktop Workstation, and its RTX PRO 6000 results also populate our enterprise GPU leaderboard.

Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!