Dell Pro Max Micro Tested: CUDA Speed and a 24GB Ceiling in 2.9 Liters
Dell's 2.9-liter Pro Max Micro leads rival systems on CUDA AI work, but it runs hot, caps at 24GB of GPU memory, and lists at $11,999 as tested.

The Dell Pro Max Micro (FCM2250) is a strange machine to benchmark, because its best results and its worst results come from the same decision. Dell put a 24GB RTX PRO 4000 Blackwell SFF Edition and a 24-core Core Ultra 9 285 into roughly 2.9 liters. For CUDA workloads, that combination beats every system outside Dell's own line in our tests. For anything that needs more graphics memory or sustained multi-core throughput, the same chassis runs into a wall.

We ran it on Ubuntu alongside Dell's Pro Precision 9 T2 tower, the Corsair FlexPrime V20C with a 32GB Intel Arc Pro B70, and two desktops built around 128GB of unified memory: the Dell Pro Max with GB10 and the AMD Ryzen AI Halo. The goal was to see what a discrete RTX PRO card in this footprint actually delivers for local AI, and what it gives up everywhere else.

Where the Micro Fits in Dell's Lineup
Dell pitches the Pro Max Micro as its ultra-compact commercial workstation, sitting below the SFF and tower Pro Max models. It's aimed at organizations that want workstation-class graphics, ISV certification and vPro management in a space too small for a tower. Dell supports mounting it on an all-in-one stand, on a shelf behind a display, on a wall, or under a desk. That's a real differentiator for clinics, trading floors, labs and anywhere else a tower would be a nuisance.
The platform accepts Core Ultra 5, 7 and 9 processors. Dell runs them at 85W with an enhanced thermal solution, even though the chips are rated at 65W. That single decision explains a lot of what follows in the CPU results.
Graphics top out at the 70W RTX PRO 4000 Blackwell SFF Edition in our unit, with 24GB of GDDR7 ECC memory. System memory reaches 96GB of DDR5-6400 on two CSODIMMs, which is what we tested. Dell also offers up to 64GB of ECC memory with its Reliable Memory Technology Pro, which maps out bad memory regions at the BIOS level. Storage runs to 8TB across two M.2 2280 slots, one PCIe Gen5 and one Gen4. Our unit shipped with two 1TB Gen5 self-encrypting drives in a RAID 0 volume.
Two Dell towers sit above it in this comparison. The Pro Precision 9 T2 is the smallest tower in the Xeon-based Pro Precision 9 line. Our test unit paired it with the 24-core Xeon 658X, a 96GB RTX PRO 6000 Blackwell Max-Q and four channels of DDR5-6400. The Pro Precision 9 T4 above it accepts up to four GPUs, depending on the card. Including the T2 shows how results scale from Dell's smallest workstation to a single-GPU tower.
Build, Ports and Serviceability
The chassis is black resin measuring 206 x 79.3 x 178mm (8.11 x 3.12 x 7.00 in). Depending on configuration it weighs between 2.34kg and 2.54kg. Dell says the design carries over from the previous generation, so existing mounts and fleet accessories should still fit. Power comes from an external 280W adapter with a 7.4mm barrel connector. The enclosure uses 46.4 percent post-consumer recycled plastic, and Dell lists it as ENERGY STAR qualified and EPEAT Gold with Climate+.

The front panel has a 20Gbps USB Type-C port, two 10Gbps USB Type-A ports (one with PowerShare) and a headset jack. The rear adds a 10Gbps USB Type-C port, two 10Gbps and two 5Gbps USB Type-A ports, three DisplayPort 1.4a outputs from the integrated graphics, and a 1GbE RJ45 port. The RTX PRO 4000 SFF card adds four mini-DisplayPort outputs, bringing the total to seven display connections. Dell includes four adapters for full-size DisplayPort.

A modular rear port takes one of nine options. Our unit had a 5GbE LAN module. The other choices are VGA, HDMI 2.1, DisplayPort 2.1, serial with PS/2, USB Type-C with DisplayPort Alt Mode, two USB Type-A ports, Thunderbolt 4 with USB Type-C, and 5G optical fiber. Dell also lists 2.5GbE and dual-port 10GbE for the platform. Wireless is Intel's Wi-Fi 7 BE200 with Bluetooth 5.4.
Inside, Dell lists two PCIe Gen4 x8-connector slots, one wired x8 for the graphics card and one wired x1, plus an M.2 2230 slot for the wireless card. Both SSDs sit under a shared heatsink with thermal pads, and the two memory modules go into CSODIMM sockets. Dell describes the chassis as largely tool-less, with memory and storage the owner can replace.

In our hands, the cover came off after one screw. The CPU fan shroud releases with finger clips and exposes the memory slots. The PCIe riser comes out after two more screws, which gives access to both expansion slots. Servicing it is simpler than the footprint suggests.

Security and management follow Dell's commercial desktop template. That means SafeBIOS off-host firmware verification, SafeID credential protection, Secured Component Verification, Opal 2.0 self-encrypting drives, a Kensington slot and a padlock ring. The system carries Intel vPro Enterprise and works with Dell Client Command Suite. It ships with Windows 11 Pro. We replaced that with Ubuntu for testing.
Our unit also came with Dell's Pro 5 Wired Fingerprint ESS Mouse (MS526C), a $44.99 accessory whose reader provides one-touch sign-in through Windows Hello. It connects over USB Type-A on a 5.91ft cable, tracks at 1,600 dpi by default and up to 6,000, and weighs 2.57oz. Dell lists support for Windows 10 and 11, so the fingerprint feature sat outside our Linux testing. For a box that may be mounted out of easy reach, it's a sensible add, though a wireless version would suit those installs better.

Specifications
| Specification | Dell Pro Max Micro |
|---|---|
| Model | Dell Pro Max Micro (FCM2250) |
| Processor | Intel Core Ultra 9 285, 24 cores (8 P-cores, 16 E-cores), 24 threads, up to 5.6GHz 36MB cache, dedicated NPU 65W, run at 85W (Dell) |
| Graphics | NVIDIA RTX PRO 4000 Blackwell SFF Edition, 70W 24GB GDDR7 ECC |
| Memory | 96GB (2x 48GB) DDR5-6400 CSODIMM, non-ECC Up to 64GB DDR5 ECC optional |
| Storage | 2x 1TB M.2 2280 PCIe Gen5 SED, RAID 0 as shipped Sandisk SN8050S as tested Up to 8TB across two M.2 slots (one Gen5, one Gen4) |
| Expansion | 2x PCIe Gen4 x8 slots (wired x8 and x1) 1x M.2 2230 (wireless) |
| Operating System | Windows 11 Pro as shipped Ubuntu 26.04.1 LTS as tested |
| Front Ports | 1x USB Type-C 20Gbps 2x USB Type-A 10Gbps (one with PowerShare) 1x headset jack |
| Rear Ports | 1x USB Type-C 10Gbps 2x USB Type-A 10Gbps 2x USB Type-A 5Gbps 1x RJ45 1GbE 1x optional module port (5GbE as tested) |
| Display Outputs | 4x mini-DisplayPort (RTX PRO 4000 SFF) 3x DisplayPort 1.4a (integrated) |
| Wireless | Intel Wi-Fi 7 BE200 (2x2), Bluetooth 5.4 |
| Audio | Internal speaker, Realtek ALC3204 codec Front headset combo port |
| Security | Dell SafeBIOS, SafeID, Secured Component Verification Opal 2.0 SED, Kensington slot, padlock ring |
| Management | Intel vPro Enterprise Dell Client Command Suite |
| Dimensions (H x W x D) | 206 x 79.3 x 178mm (8.11 x 3.12 x 7.00in), about 2.9L |
| Weight | 2.34kg to 2.54kg (5.15lb to 5.59lb) |
| Power | 280W AC adapter |
| Warranty | 3 years ProSupport with on-site service as configured 1 year base |
| Price | $1,449 base $11,999 as tested (Dell.com) |
How We Tested
Dell ships the Micro with Windows 11 Pro, but we test higher-end workstations on Linux with a current kernel. The lab installed Ubuntu 26.04.1 LTS with the 7.0 kernel on one of the two 1TB SSDs, so the factory RAID 0 volume isn't part of these numbers. The system ran BIOS 1.19.1 on the default power profile with NVIDIA driver 595.99.02. The suite measured 23.9GB of GPU-usable memory on the RTX PRO 4000 SFF.
Results come from three in-house suites. sr-client_linux 0.1 covers local AI, content creation and storage, and each figure is the median of repeated passes. The benchmark harness wraps Phoronix Test Suite workloads and y-cruncher. The server suite runs ClickBench, OpenFOAM, STREAM and Valkey. MLPerf Client's Linux release ships only Intel OpenVINO configurations, so it produced no results on the NVIDIA card. The fine-tuning test was also left out of this run.
The comparison set adds three alternatives to Dell's own hardware. The Corsair FlexPrime V20C is a micro-ATX tower with a 24-core Core Ultra 7 270K Plus, a 32GB Arc Pro B70 and 64GB of DDR5-6400. The Dell Pro Max with GB10 runs NVIDIA's 20-core GB10 with 128GB of unified memory on NVIDIA's Ubuntu 24.04 image. The AMD Ryzen AI Halo runs a 16-core Ryzen AI Max+ 395 with 128GB of unified LPDDR5X on AMD's Linux developer image.
The software path changes the AI numbers, so keep it in mind when reading them. The Micro, T2 and GB10 ran CUDA throughout llama.cpp, Ollama and stable-diffusion.cpp. The Halo used AMD's ROCm for all three. On the V20C, llama.cpp ran on its SYCL backend, while Ollama and stable-diffusion.cpp ran on Vulkan, because Ollama's native path placed none of the model on the Arc card. That gap accounts for much of the Micro's lead in the Ollama and image tests.
| Configuration | CPU | Memory | GPU | Boot Drive | Operating System |
|---|---|---|---|---|---|
| Dell Pro Max Micro (FCM2250) | Core Ultra 9 285, 24 cores | 96GB DDR5-6400 (2x 48GB) | NVIDIA RTX PRO 4000 Blackwell SFF, 23.9GB usable | 1TB Sandisk SN8050S, PCIe Gen5 | Ubuntu 26.04.1 LTS, kernel 7.0 |
| Dell Pro Precision 9 T2 (PW9T2260) | Xeon 658X, 24 cores | 128GB DDR5-6400 (4x 32GB) | NVIDIA RTX PRO 6000 Blackwell Max-Q, 95.6GB usable | Samsung PM9E1 Gen5, 2TB drives in a striped array | Ubuntu 26.04.1 LTS, kernel 7.0 |
| Corsair FlexPrime V20C | Core Ultra 7 270K Plus, 24 cores | 64GB DDR5-6400 (4x 16GB) | Intel Arc Pro B70, 31.9GB usable | 1TB XPG MARS 970 PLUS, PCIe Gen5 | Ubuntu 26.04.1 LTS, kernel 7.0 |
| Dell Pro Max with GB10 (FCM1253) | NVIDIA GB10, 20 cores | 128GB LPDDR5X-8533, unified | NVIDIA GB10, 121.6GB usable | 4TB PCIe Gen4 | Ubuntu 24.04.5 LTS, NVIDIA kernel 7.0 |
| AMD Ryzen AI Halo | Ryzen AI Max+ 395, 16 cores | 128GB LPDDR5X-8000, unified | Radeon 8060S (integrated), 95.5GB usable | 2TB Micron 4600, PCIe Gen4 link | AMD Linux developer image, kernel 7.2 |
Local AI: Where the Micro Wins and Where It Stops
The 24GB card is the whole story for AI work on this machine. The suite measured 23.9GB as usable, which is enough for the 20B agent model and the Ministral serving models in our lineup. The V20C's 32GB card clears the 30GB that the 27B serving model needs. The T2's 96GB card and the 128GB unified-memory systems handle the larger tiers the Micro can't load at all.
LLM Serving
| llama.cpp server (higher is better unless noted) | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| Ministral 3 3B, 1 request (tokens/s) | 80.7 | 243.8 | 84.9 | 53.2 | 43.0 |
| Ministral 3 3B, 4 requests (tokens/s) | 201.5 | 602.6 | 213.3 | 139.8 | 105.1 |
| Ministral 3 3B, 1 request, time to first token (ms, lower is better) | 666 | 197 | 590 | 572 | 1,659 |
| Ministral 3 8B, 1 request (tokens/s) | 52.0 | 167.3 | 63.7 | 37.0 | 30.6 |
| Ministral 3 8B, 4 requests (tokens/s) | 108.9 | 384.9 | 141.8 | 97.2 | 66.6 |
| Ministral 3 8B, 1 request, time to first token (ms, lower is better) | 1,477 | 427 | 1,362 | 1,398 | 4,602 |
| Qwen3.8 27B, 1 request (tokens/s) | Did not fit | 53.8 | 19.8 | 11.2 | 9.1 |
| Qwen3.8 27B, 4 requests (tokens/s) | Did not fit | 113.5 | 33.5 | 27.1 | 18.4 |
| Qwen3.8 27B, 8 requests (tokens/s) | Did not fit | 155.8 | 35.4 | 35.6 | 20.1 |
| Qwen3.8 27B, 1 request, time to first token (ms, lower is better) | Did not fit | 2,771 | 9,005 | 9,207 | 24,024 |
The llama.cpp server ran 4,096-token prompts with 1,024-token outputs on the Ministral 3 models. For one user on the 3B model, the Micro delivered 80.7 tokens per second. That's 5 percent behind the Arc Pro B70 in the V20C and 52 percent ahead of the GB10. On the 8B model it managed 52.0 tokens per second, 18 percent behind the V20C and 41 percent ahead of the GB10. With four concurrent requests it reached 201.5 and 108.9 tokens per second, which put it 44 and 12 percent ahead of the GB10.
The Halo's integrated Radeon 8060S was the slowest system in this test. It served the 3B model at 43.0 tokens per second for one request and 105.1 across four, roughly half the Micro's rate. Its prompt processing lagged too. Median time to first token was 1.7 seconds on the 3B model and 4.6 seconds on the 8B, against 0.7 and 1.5 seconds on the Micro.

The T2 is the clear throughput leader. Its RTX PRO 6000 Max-Q served the 3B and 8B models at 243.8 and 167.3 tokens per second with one request, which is 3.0 and 3.2 times the Micro. At four concurrent requests it reached 602.6 and 384.9 tokens per second. It also ran Qwen3.8 27B at 53.8 tokens per second for one request and 155.8 across eight, with the first token arriving in 2.8 seconds.
Qwen3.8 27B at Q4_K_M needs 30GB of GPU-usable memory for its weights and one request's KV cache. The Micro's 23.9GB couldn't hold it. The V20C produced 19.8 tokens per second on that model, the GB10 managed 11.2, and the Halo 9.1. The same limit ruled out gpt-oss 120B, which needs 90GB, and MiniMax M2.7 at 129.3GB. In practical terms, 24GB covers models up to roughly the 20B class at 4-bit quantization. Anything larger means a bigger card or a unified-memory system.

Model Loading and the Coding Agent
| gpt-oss:20b (Ollama 0.40.3) | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| gpt-oss:20b cold load to first token (s, lower is better) | 5.11 | 5.57 | 6.29 | 6.86 | 4.20 |
| gpt-oss:20b warm load to first token (s, lower is better) | 3.12 | 3.47 | 3.18 | 6.36 | 4.03 |
| Coding agent, successful runs (of 3) | 1* | 2* | 3 | 3 | 2* |
| Coding agent, median wall time (s, lower is better) | 80* | 107* | 368 | 221 | 282* |
| Coding agent, generation rate (tokens/s) | 76.2 | 225.0 | 32.2 | 53.9 | 41.2 |
*Flagged by the suite: the Micro passed one of three agent runs, and the T2 and the Halo passed two of three.
The model-load test times Ollama from launch to first token on gpt-oss:20b, a 12.8GB model, both cold from SSD and warm from the page cache. The Micro reached first token in 5.11 seconds cold and 3.12 seconds warm, the fastest warm load in the group. The Halo had the fastest cold load at 4.20 seconds. The Micro's Gen5 drive read at 6,636MB/s during the cold load, peaking at 8,754MB/s, so storage keeps the runtime fed.
The coding agent runs Crush against gpt-oss:20b with a 64K context through a fixed programming task, three times. The Micro generated 76.2 tokens per second, ahead of the GB10 at 53.9, the Halo at 41.2 and the V20C at 32.2, though only about a third of the T2's 225.0. Its one passing run finished in a median of 80 seconds. The other two runs ended with 0 and 1 of 11 tests passing. The T2 and Halo passed two of three runs, and the V20C and GB10 passed all three. Pass counts on this model shift from run to run, so the generation rate is the more reliable hardware signal here.

RAG at Two Scales
| RAG, 20,963 passages (Ollama 0.40.3) | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| Ingest, passages per second | 268.2 | 392.1 | 87.5 | 211.4 | 95.5 |
| Ingest, embedding rate (tokens/s) | 43,430 | 63,486 | 14,162 | 34,227 | 15,468 |
| Answer, generation rate (tokens/s) | 137.0 | 402.8 | 76.5 | 93.2 | 80.0 |
| Answer, time to first token p50 (s, lower is better) | 0.16 | 0.11 | 0.30 | 0.15 | 0.42 |
| Question to answer p50 (s, lower is better) | 0.22 | 0.14 | 0.39 | 0.22 | 0.50 |
The small RAG test embeds 20,963 SQuAD passages, retrieves the top five for each question, and answers with Phi-4-mini, all through Ollama. The Micro ingested 268.2 passages per second, which is 27 percent ahead of the GB10's 211.4 and roughly three times the V20C's Vulkan rate of 87.5. The T2 reached 392.1. Answers generated at 137.0 tokens per second on the Micro, against 93.2 on the GB10, 80.0 on the Halo and 76.5 on the V20C. The T2 hit 402.8. The Micro's median time from question to answer was 0.22 seconds, level with the GB10.

| RAG at scale, 1,002,067 passages | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|
| Ingest, 1M passages (passages/s) | 313.9 | 1,140.9 | 321.4 | 65.8 |
| Ingest wall time (s, lower is better) | 3,192 | 878 | 3,118 | 15,230 |
| HNSW index build (s, lower is better) | 78.9 | 46.3 | 73.4 | 68.9 |
| HNSW search, ef 64 (queries/s) | 27,359 | 45,085 | 29,643 | 34,012 |
| HNSW search, ef 256 (queries/s) | 7,664 | 13,478 | 8,285 | 9,649 |
| Question to answer p50 (s, lower is better) | 0.20 | 0.12 | 0.36 | 0.46 |
The larger test pushes a million Wikipedia passages through llama.cpp, builds an HNSW index on the CPU and searches it. The Micro ingested 313.9 passages per second and finished in 3,192 seconds, about 53 minutes. That's 2 percent behind the V20C, whose SYCL backend is much stronger than its Vulkan path in the smaller test. Embedding rates on the Micro stayed consistent across both scales at 43,430 tokens per second in the smaller run. The T2 embedded the same million passages in 878 seconds, under 15 minutes, at 3.6 times the Micro's rate.
The CPU-side work splits differently. The Micro built the index in 78.9 seconds, against 73.4 on the V20C and 46.3 on the T2. It searched at 27,359 queries per second at an ef setting of 64, against 29,643 and 45,085, with recall against exact search at 91.4 percent. Answer generation returns to Ollama here, and the Micro's median question-to-answer time of 0.20 seconds beat the V20C's 0.36 by a wide margin. The GB10 has not run the million-passage test yet.
The Halo needed 15,230 seconds, about 4.2 hours, to ingest the million passages at 65.8 per second. That's roughly a fifth of the Micro's rate. Its 16-core CPU did better on the index side, building it in 68.9 seconds and searching at 34,012 queries per second at ef 64, ahead of both the Micro and the V20C.
Image Generation
| stable-diffusion.cpp, 1024×1024 (images/min, higher is better) | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| FLUX.2 klein 4B Q4_0 (images/min) | 11.46 | 40.53 | 4.26 | 9.96 | 3.86 |
| FLUX.2 klein 4B BF16 (images/min) | 10.93 | 41.32 | 2.40 | 10.28 | 4.17 |
| Z-Image Turbo 6B Q4_0 (images/min) | 4.39 | 16.97 | 2.11 | 3.68 | 1.85 |
| Z-Image Turbo 6B BF16 (images/min) | 3.95 | 17.27 | 0.96 | 3.94 | 0.47 |
| Qwen-Image 20B Q4_0 (images/min) | Did not fit | 2.42 | 0.35 | 0.50 | 0.23 |
| Qwen-Image 20B Q8_0 (images/min) | Did not fit | 2.39 | Did not fit | 0.48 | 0.23 |
Image generation is where the CUDA advantage is easiest to see. On stable-diffusion.cpp, the Micro produced 11.46 images per minute on FLUX.2 klein 4B at Q4_0. That's 15 percent ahead of the GB10's 9.96, 2.7 times the V20C's 4.26 on Vulkan and three times the Halo's 3.86 on ROCm. On Z-Image Turbo 6B at Q4_0 it managed 4.39 images per minute, against 3.68, 2.11 and 1.85 on the same three systems. Precision barely moved the numbers. BF16 gave 10.93 on FLUX.2 klein and 3.95 on Z-Image Turbo. The T2's card produced 40.53 FLUX.2 images per minute, 3.5 times the Micro.

Qwen-Image 20B shows the memory ceiling again. It needs 28GB of GPU-usable memory at Q4_0 and 38GB at Q8_0, so it wouldn't load on the 24GB card. The V20C ran the Q4_0 version at 0.35 images per minute. The GB10 ran both quantizations at 0.50 and 0.48, the Halo at 0.23 for both, and the T2 at 2.42 and 2.39. If your team needs that model class locally, plan for more graphics memory than this chassis can take. The RTX PRO 4000 SFF is the top card Dell offers in the Micro.
CPU Performance
Compression, Rendering and Encoding
| Test | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| 7-Zip, all threads, total (GIPS) | 164.0 | 285.1 | 199.6 | 149.5 | 215.8 |
| 7-Zip, one thread, total (GIPS) | 10.90 | 9.27 | 11.26 | 9.68 | 9.66 |
| Blender 4.5 Monster (samples/min) | 243.1 | 400.7 | 330.4 | No arm64 build | 285.3 |
| Blender 4.5 Junkshop (samples/min) | 161.6 | 266.9 | 212.5 | No arm64 build | 173.9 |
| Blender 4.5 Classroom (samples/min) | 114.5 | 187.2 | 152.8 | No arm64 build | 144.0 |
| HandBrake x265 (fps) | 39.7 | 41.5 | 49.6 | No arm64 build | 44.5 |
| HandBrake SVT-AV1 (fps) | 147.9 | 198.5 | 177.4 | No arm64 build | 194.1 |
| LLVM build (s, lower is better) | 173 | 106 | 136 | 175 | 149 |
The Core Ultra 9 285 is a 65W desktop chip that Dell runs at 85W here. It shares its 24-core layout with the 270K Plus in the V20C, and the comparison with that tower is the cleanest read on what the form factor costs. The Micro scored 164.0 GIPS in 7-Zip, 18 percent below the V20C's 199.6. It finished 24 to 26 percent slower across the three Blender scenes, and its HandBrake x265 and SVT-AV1 encodes came in 20 and 17 percent slower. The V20C has a tower chassis and a 360mm liquid cooler, which matters when every core is loaded for minutes at a time.

Against the GB10, the Micro was 10 percent ahead in all-thread 7-Zip and 13 percent ahead with one thread. The LLVM build finished level, at 173 seconds against 175. Blender and HandBrake have no arm64 Linux builds in our suite, so the GB10 has no results there. The Halo's 16-core, 32-thread Ryzen AI Max+ 395 beat the Micro on multi-thread work. It scored 215.8 GIPS in 7-Zip, 32 percent ahead, ran 8 to 26 percent ahead across the Blender scenes, and encoded SVT-AV1 at 194.1 fps against 147.9. It built LLVM in 149 seconds against 173. The Micro kept its 13 percent lead in one-thread 7-Zip.

The T2 scales the way a 24-core, 48-thread Xeon in a tower should. It scored 285.1 GIPS in 7-Zip, 74 percent ahead of the Micro. It delivered 63 to 65 percent more Blender throughput and finished the LLVM build in 106 seconds. Single-thread speed runs the other way. The Core Ultra 9 285 beat the Xeon 658X by 18 percent in one-thread 7-Zip and 14 percent in one-thread Stockfish, and it ran 18 to 20 percent quicker in the single-thread Python tests.
Stockfish, primesieve and Python
| Test | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| Stockfish, one thread (Mnps) | 2.02 | 1.77 | 2.08 | 1.66 | 2.15 |
| Stockfish, all threads (Mnps) | 31.58 | 45.44 | 40.55 | 23.86 | 38.74 |
| primesieve to 10^12, all threads (s, lower is better) | 8.72 | 7.51 | 6.48 | 9.16 | 6.29 |
| primesieve to 10^11, one thread (s, lower is better) | 9.07 | 9.66 | 9.03 | 12.32 | 6.76 |
| PyPerformance go (ms, lower is better) | 57.8 | 72.1 | 56.0 | 85.5 | 69.1 |
| PyPerformance nbody (ms, lower is better) | 38.9 | 47.3 | 36.3 | 55.0 | 43.1 |
All-thread Stockfish reached 31.58 Mnps on the Micro. That trails the T2 at 45.44, the V20C at 40.55 and the Halo at 38.74, but it's 32 percent ahead of the GB10. One-thread results between the two Core Ultra systems were within 3 percent in Stockfish, and level in primesieve at 9.07 and 9.03 seconds. The single-thread Python tests ran 3 to 7 percent behind the V20C and finished about 30 percent faster than on the Arm-based GB10.
GROMACS and OpenBLAS
| Test | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| GROMACS benchMEM, 16 cores (ns/day) | 52.68 | 62.66 | 67.23 | 27.81 | 72.39 |
| GROMACS benchMEM, all cores (ns/day) | 55.52 | 81.56 | 77.77 | 34.61 | 70.95 |
| OpenBLAS DGEMM, all cores (GFLOPS) | 791.6 | 2,318.9 | 1,100.0 | 418.0 | 1,610.3 |
| OpenBLAS LU solve N=20000, all cores (GFLOPS) | 717.0 | 1,615.8 | 938.6 | 393.0 | 1,391.3 |
| Memory latency, 4KiB pages (ns, lower is better) | 129.0 | 143.8 | 105.2 | 123.8 | 162.2 |
GROMACS benchMEM ran at 52.68 ns/day on 16 cores and 55.52 on all 24. That's 22 and 29 percent behind the V20C, and 89 and 60 percent ahead of the GB10. OpenBLAS DGEMM reached 791.6 GFLOPS on all cores, against 1,100.0 on the V20C and 418.0 on the GB10. The T2's Xeon, with AVX-512 and four memory channels, hit 2,318.9. GROMACS was the hottest workload in the suite for this chassis, and the small gain from 16 cores to 24 looks like a processor sitting at its thermal limit. Memory latency on 4KiB pages came in at 129.0ns, against 105.2ns on the V20C and 123.8ns on the GB10.
The Halo posted the fastest 16-core GROMACS result in the group at 72.39 ns/day, 37 percent ahead of the Micro. Across all 32 threads it reached 70.95 ns/day, behind the T2 and the V20C. Its OpenBLAS DGEMM figure of 1,610.3 GFLOPS was about twice the Micro's and second only to the T2.

Server and HPC Workloads
These results come from the benchmark harness and server suite, both Linux-only. They treat the Micro as a general compute node that compiles code, compresses data, serves web and cache traffic, runs analytics queries and solves a CFD case. The Micro, T2 and V20C share Ubuntu 26.04.1, so those three give the most direct comparison. The other two systems run different distributions and compilers.
| Phoronix Test Suite and y-cruncher (higher is better unless noted) | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| Timed Linux kernel compile, allmodconfig (s, lower is better) | 918 | 537 | 639 | 892 | 727 |
| 7-Zip compression rating, all threads (MIPS) | 150,543 | 264,940 | 186,668 | 154,634 | 178,530 |
| 7-Zip decompression rating, all threads (MIPS) | 88,641 | 195,348 | 128,723 | 106,148 | 143,464 |
| 7-Zip compression rating, one thread (MIPS) | 9,351 | 9,100 | 10,260 | 7,474 | 8,019 |
| OpenSSL AES-256-GCM, all threads (GB/s) | 110.0 | 371.2 | 171.4 | 97.1 | 298.6 |
| nginx, 1,000 connections (requests/s) | 155,203 | 323,567 | 225,708 | 114,909 | 223,433 |
| nginx, 4,000 connections (requests/s) | 109,352 | 244,008 | 155,146 | 101,613 | 190,454 |
| y-cruncher Pi, 1B digits (s, lower is better) | 30.10 | 8.51 | 17.54 | Not run | 10.46 |
| y-cruncher Pi, 5B digits (s, lower is better) | 158.24 | 53.40 | 111.17 | Not run | 63.45 |
| y-cruncher Pi, 10B digits (s, lower is better) | 342.87 | 119.21 | 244.44 | Not run | 135.18 |
The Micro compiled the Linux kernel in 918 seconds. That's 3 percent behind the GB10 at 892, and well behind the Halo at 727 and the V20C at 639. All-thread 7-Zip compression landed within 3 percent of the GB10. The Micro led that system by 25 percent with one thread, by 13 percent in OpenSSL AES-256-GCM and by 35 percent in nginx at 1,000 connections. The V20C and the Halo were both faster in the all-thread tests. The T2 led every all-thread row, compiling the kernel in 537 seconds and running OpenSSL at 371.2GB/s against the Micro's 110.0GB/s.

In y-cruncher, the Micro computed five billion digits of Pi in 158.24 seconds, against 111.17 on the V20C and 63.45 on the Halo. Ten billion digits took 342.87 seconds, against 244.44 and 135.18. The test is sensitive to memory bandwidth and vector width. The Halo brings AVX-512 and a wider memory bus, while the Micro runs two channels of DDR5-6400. The T2 was fastest at 53.40 and 119.21 seconds. The GB10 wasn't run on y-cruncher.

| Server suite (higher is better unless noted) | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| ClickBench hot run, 43 queries (s, lower is better) | 8.11 | 4.80 | 7.39 | 6.67 | 5.82 |
| ClickBench cold run, 43 queries (s, lower is better) | 10.80 | 6.92 | 10.41 | 11.01 | 9.30 |
| OpenFOAM drivAer, 3M cells, solve time (s, lower is better) | 2,320 | 826 | 1,636 | 1,222 | 1,148 |
| STREAM Triad (MB/s) | 65,802 | 141,832 | 68,252 | 121,313 | 114,007 |
| Valkey cache, SET (requests/s) | 1,389,084 | 914,832 | 1,852,556 | 810,246 | 1,211,538 |
| Valkey cache, GET (requests/s) | 1,398,409 | 991,576 | 1,866,031 | 884,644 | 1,271,481 |
Memory bandwidth explains most of the server suite. STREAM Triad measured 65.8GB/s on the Micro's two DDR5 channels, close to the V20C's 68.3GB/s and roughly half of the 121.3GB/s on the GB10 and 114.0GB/s on the Halo. The T2 reached 141.8GB/s. The OpenFOAM drivAer case, a 3-million-cell solve, took 2,320 seconds on the Micro. That compares with 826 on the T2, 1,636 on the V20C, 1,222 on the GB10 and 1,148 on the Halo. The ClickBench hot run took 8.11 seconds on the Micro, against 4.80, 7.39, 6.67 and 5.82.

Valkey rewards per-core speed, and the Micro handled 1.39 million SET and 1.40 million GET requests per second in the cache profile. That's 25 percent under the V20C, but 15 and 10 percent ahead of the Halo, 71 and 58 percent ahead of the GB10, and 52 and 41 percent ahead of the T2's Xeon. A workstation that also hosts a cache or a small database next to a model server gets real value from the Core Ultra's per-core strength.

GPU Encoding
For GPU encoding, FFmpeg 8.1 with NVENC transcoded the suite's 4K Big Buck Bunny clip to HEVC at 123 fps on the RTX PRO 4000 SFF. The T2 and the GB10 ran the same encoder at 124 and 134 fps. The V20C reached 154 fps through Intel Quick Sync, and the Halo reached 204 fps through VAAPI. Those two used different encoder settings, so treat them as reference points rather than rankings. A newer FFmpeg build in the suite requires NVIDIA driver 610.00 or later for NVENC, so it didn't run on the 595.99.02 driver installed here.
Thermals Under Load
| Dell Pro Max Micro | CPU Max (C) | CPU Mean (C) | GPU Max (C) | GPU Mean (C) |
|---|---|---|---|---|
| 7-Zip | 101 | 78.7 | 32 | 31.7 |
| Blender | 98 | 82.3 | 35 | 33.7 |
| LLVM build | 100 | 79.4 | 33 | 33.0 |
| OpenBLAS | 104 | 83.5 | 32 | 32.0 |
| GROMACS | 105 | 89.3 | 33 | 32.6 |
| Coding agent | 83 | 70.0 | 69 | 62.3 |
| LLM serving | 76 | 73.5 | 73 | 67.2 |
| RAG at scale | 97 | 79.6 | 73 | 71.7 |
| Image generation | 85 | 72.0 | 72 | 68.5 |
The 285 ran against its thermal ceiling in every all-core test. Package temperature peaked at 98C in Blender, 100C in the LLVM build, 101C in 7-Zip, 104C in OpenBLAS and 105C in GROMACS. Means across those tests ran from 79C to 89C. That's the price of running a 65W desktop chip at 85W in 2.9 liters, and it lines up with the multi-thread gap to the tower. The processor idled at about 40C between tests.

The graphics card ran much cooler. The RTX PRO 4000 SFF idled at 32C and peaked at 73C during LLM serving and the million-passage RAG ingest, averaging 71.7C across a 57-minute run. Image generation held it at 72C and the coding agent at 69C. So the 70W card has thermal headroom even when the CPU doesn't. The SSD stayed at 63C or below through the 74-minute storage test.
Storage Performance
| fio, 25GiB footprint (higher is better) | Dell Pro Max Micro (Ultra 9 285, RTX PRO 4000 SFF 24GB) | Dell Pro Precision 9 T2 (Xeon 658X, RTX PRO 6000 Max-Q 96GB) | Corsair FlexPrime V20C (Ultra 7 270K Plus, Arc Pro B70 32GB) | Dell Pro Max with GB10 (128GB unified) | AMD Ryzen AI Halo (Ryzen AI Max+ 395, 128GB unified) |
|---|---|---|---|---|---|
| Sequential read 128K (MB/s) | 9,433 | 20,914* | 7,311 | 6,947 | 7,232 |
| Sequential write 128K (MB/s) | 6,766 | 12,564 | 6,082 | 6,310 | 1,554* |
| Random read 4K, peak (IOPS) | 2,084,437 | 3,214,066 | 1,184,970 | 932,210 | 1,197,992 |
| Random write 4K, peak (IOPS) | 1,028,079* | 2,353,766 | 184,231 | 997,758 | 121,285* |
| Random read 4K, QD32 x 4 jobs (IOPS) | 1,386,889 | 1,394,262 | 1,184,970 | 754,896 | 1,001,941 |
| Random write 4K, QD32 x 4 jobs (IOPS) | 1,028,079* | 1,175,233 | 178,312 | 946,264 | 121,285* |
*Flagged by the suite for run-to-run variance.
The lab's Ubuntu install sits on one of the two 1TB Sandisk SN8050S drives, which links at PCIe Gen5 x4. These results therefore describe a single drive and leave out Dell's RAID 0 setup. That drive read at 9,433MB/s and wrote at 6,766MB/s. Reads were 29 percent ahead of the V20C's drive and 30 percent ahead of the Halo's, and 36 percent ahead of the Gen4 drive in the GB10. The T2's 20,914MB/s comes from a striped array of Gen5 drives. Peak 4K random read reached 2.08 million IOPS, against 3.21 million on the T2, 1.18 million on the V20C, 1.20 million on the Halo and 932K on the GB10.

Random write peaked at 1.03 million IOPS, level with the GB10's 998K, under half of the T2's 2.35 million and more than five times the V20C's 184K and the Halo's 121K. The suite flagged the Micro's and the Halo's random-write results for run-to-run variance after a rerun. The drive peaked at 63C under its heatsink. A second 1TB drive is already in the box, so a buyer can keep the striped volume for scratch space, or split the drives into a boot disk and a model library. Dell lists up to 8TB across the two slots.

Price, Trade-offs and Who Should Buy This
At $11,999 as configured on Dell.com, our unit costs three times the $3,999.99 price listed for both the Corsair FlexPrime V20C and the AMD Ryzen AI Halo. It's also about 1.6 times the $7,521 Dell lists for a 128GB Pro Max with GB10 and a 1TB SSD. The configuration includes the Core Ultra 9 285, the RTX PRO 4000 SFF, 96GB of memory, two Gen5 SSDs, a 5GbE module, Wi-Fi 7 and three years of ProSupport with on-site service in place of the one-year base warranty. Dell base pricing starts at $1,449. Dell sells most of these systems through account teams and partners, where volume pricing applies.
The premium buys the commercial package: ISV-certified RTX PRO graphics, vPro Enterprise, ECC memory options, Opal self-encrypting drives and onsite support. For a fleet that needs managed, certified hardware in the smallest possible footprint, that's a coherent argument. Teams choosing purely on tokens per dollar will find cheaper routes to the same throughput.
The Micro's design answers the question of where a workstation can go more than it answers how fast one can run. Its best showing is CUDA-backed local inference and image generation at small-to-mid model sizes. Its limits are the 24GB ceiling, which rules out the 27B and larger models, and a 2.9-liter thermal envelope that holds the 285 at 98C to 105C under full load. The V20C's tower cooling and the Halo's larger memory pool both beat it on the workloads they're suited to.
My read is that the next decision point is the one Dell's lineup makes visible. The T2 showed roughly three times the Micro's LLM serving rate, 3.5 times its image generation rate and 74 percent more 7-Zip throughput, and its 96GB card ran the models the Micro couldn't load. If the 24GB limit is what's holding a project back, the Micro isn't the fix. If the workload fits in 24GB and the desk is the constraint, it's the most capable CUDA box that fits there.
