Dell Pro Max with GB10

Dell's Pro Max with GB10, powered by NVIDIA's GB10 chip, delivers performance nearly identical to NVIDIA's reference design but with better sustained throughput before thermal throttling, according to a hands-on review. In llama-bench tests, the Dell GB10 achieved 180.55 tokens per second on a 70B Q8_0 model and 966.48 tokens per second on a 120B MXFP4 MoE model, with the reviewer attributing the improvement to tweaked airflow and higher fan RPMs. The system also worked out of the box with Dell's KM900 wireless keyboard and mouse via Bluetooth in Linux.

Physical Config As I covered in the video, it is almost identical to the NVIDIA design GB10. It comes with a 280w USBc power brick +40w and a few relatively minor tweaks: PXE Network booting, Power LED on the USBc cable and Power LED on the front. It worked great via Bluetooth with the Dell KM900 wireless Keyboard and mouse Right out of the box In Linux Benchmarks Overall, the bottom line is the performance is really similar to the nvidia reference configuration, except that the Dell GB10 has a little better staying power before throttling creeps in. I suspect this is because the front bezel and fan config have been tweaked for a little better airflow. The fan RPMS seem to be a little higher, for sure. This is llama-bench, which is still not as optimized for the spark GB10 platform as, say, TensorRT LLM, but the performance is decent. ggml cuda init: GGML CUDA FORCE MMQ: no ggml cuda init: GGML CUDA FORCE CUBLAS: no ggml cuda init: found 1 CUDA devices: Device 0: NVIDIA GB10, compute capability 12.1, VMM: yes | model | size | params | backend | ngl | n ubatch | main gpu | mmap | test | t/s | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | pp2048 | 180.55 ± 0.32 | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | tg32 | 2.95 ± 0.00 | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | pp2048 @ d4096 | 125.25 ± 0.60 | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | tg32 @ d4096 | 2.85 ± 0.00 | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | pp2048 @ d8192 | 94.92 ± 0.08 | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | tg32 @ d8192 | 2.75 ± 0.00 | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | pp2048 @ d16384 | 63.17 ± 0.32 | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | tg32 @ d16384 | 2.57 ± 0.00 | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | pp2048 @ d32768 | 36.76 ± 1.37 | | llama 70B Q8 0 | 69.82 GiB | 70.55 B | CUDA | 99 | 2048 | 2 | 0 | tg32 @ d32768 | 2.24 ± 0.00 | ./bin/llama-bench -m ../../.cache/huggingface/hub/models–ggml-org–gpt-oss-120b-GGUF/snapshots/0764c156c5379ea3939e051b46d897ca7011a0e7/gpt-oss-120b-mxfp4-00001-of-00003.gguf -ngl 99 -mmp 0 -fa 0 -b 2048 -ub 2048 -p 2048 -n 32 -d 0,4096,8192,16384,32768 ggml cuda init: GGML CUDA FORCE MMQ: no ggml cuda init: GGML CUDA FORCE CUBLAS: no ggml cuda init: found 1 CUDA devices: Device 0: NVIDIA GB10, compute capability 12.1, VMM: yes | model | size | params | backend | ngl | n ubatch | mmap | test | t/s | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | pp2048 | 966.48 ± 2.78 | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | tg32 | 55.92 ± 0.18 | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | pp2048 @ d4096 | 624.44 ± 0.94 | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | tg32 @ d4096 | 51.19 ± 0.21 | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | pp2048 @ d8192 | 466.34 ± 0.53 | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | tg32 @ d8192 | 47.80 ± 0.18 | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | pp2048 @ d16384 | 305.45 ± 0.15 | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | tg32 @ d16384 | 42.31 ± 0.12 | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | pp2048 @ d32768 | 180.88 ± 0.09 | | gpt-oss 120B MXFP4 MoE | 59.02 GiB | 116.83 B | CUDA | 99 | 2048 | 0 | tg32 @ d32768 | 32.93 ± 0.10 | build: 4dca015b7 7070 ./bin/llama-bench -m ../../.cache/huggingface/hub/models–ggml-org–gpt-oss-20b-GGUF/snapshots/e1dc459feff949ff451ce107337a2026daa80df8/gpt-oss-20b-mxfp4.gguf -ngl 99 -mmp 0 -fa 0 -b 2048 -ub 2048 -p 2048 -n 32 -d 0,4096,8192,16384,32768 ggml cuda init: GGML CUDA FORCE MMQ: no ggml cuda init: GGML CUDA FORCE CUBLAS: no ggml cuda init: found 1 CUDA devices: Device 0: NVIDIA GB10, compute capability 12.1, VMM: yes | model | size | params | backend | ngl | n ubatch | mmap | test | t/s | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | pp2048 | 1645.57 ± 3.76 | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | tg32 | 78.89 ± 0.39 | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | pp2048 @ d4096 | 1003.46 ± 2.41 | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | tg32 @ d4096 | 72.42 ± 0.47 | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | pp2048 @ d8192 | 739.27 ± 0.72 | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | tg32 @ d8192 | 67.85 ± 0.27 | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | pp2048 @ d16384 | 474.33 ± 0.56 | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | tg32 @ d16384 | 60.52 ± 0.22 | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | pp2048 @ d32768 | 271.23 ± 0.51 | | gpt-oss 20B MXFP4 MoE | 11.27 GiB | 20.91 B | CUDA | 99 | 2048 | 0 | tg32 @ d32768 | 47.61 ± 0.16 | And TensorRT-LLM can do upwards of 4.5 t/s with the LLama 70b Q8 model on the same hardware, fwiw. vLLM Benchmarks What about Nvidia Nemotron models? Don’t sleep on NVIDIA Nemotron Models. They’re fantastic. These models are great models, and the performance on Nvidia hardware is first-rate. Much of the benchmarks here are to give you some idea of a rough comparison with other platforms – and are provided for the sake of completeness – but what the GB10 offers is not raw performance. Instead, GB10 is meant to give you access to the more or less full-stack ecosystem from Nvidia. Vision Language Models? NVIDIA DGX Spark GB10 is particularly strong for vision-language models because VLM workloads are dominated by GPU compute and accelerator memory locality, not raw memory bandwidth. GB10’s Blackwell GPU brings massive parallelism via CUDA cores, Tensor Cores with FP8/FP4 support , and high SM counts, which efficiently handle convolutional vision backbones, attention blocks, and multimodal fusion layers. While LPDDR5 is way lower bandwidth, VLMs benefit from high arithmetic intensity, aggressive kernel fusion, and reuse within on-package memory, so most hot data stays close to the GPU. In practice, image encoders and cross-attention layers scale with Tensor Core throughput and CUDA scheduling, making GB10 an excellent, power-efficient platform for multimodal inference and fine-tuning despite slower system RAM. The hyperfocus on LLM performance overshadows other capabilities of the platform, imho. This compute advantage translates to image generation as well: diffusion and transformer-based image models are dominated by dense matrix math, attention, and convolution kernels that map cleanly onto CUDA cores and Tensor Cores, so throughput scales primarily with GPU compute rather than host memory speed. With FP8/FP4 Tensor Core acceleration, high SM occupancy, and mature CUDA kernels, the GB10 can keep denoisers, UNets, and vision transformers fed efficiently, allowing image synthesis and editing workloads to run at high utilization even when system memory bandwidth is comparatively modest. NV FP4 Advantages Models quantized to NVFP4 have a broad advantage because NVFP4 is a true 4-bit floating-point format designed to preserve accuracy comparable to 8-bit quantization, rather than behaving like a heavily compressed integer scheme. NVIDIA has shown that NVFP4 maintains model quality close to FP8/INT8 on many transformer workloads while enabling much higher Tensor Core throughput and lower memory traffic, allowing inference to stay compute-bound instead of dequant-bound. That distinction matters for performance: our benchmarks on GPT-OSS 20B are MXFP4—not NVFP4—so there is real headroom left, even for “native” 4-bit models like GPT-OSS, because NVFP4 is a different 4-bit format with hardware-aligned scaling and numerical behavior, not just a smaller representation of the same quantization. Steam / Linux Gaming The Dell Pro Max with GB10 is not designed to be a gaming machine — its strength is as a purpose-built AI workstation. But… the Grace Blackwell Superchip sure can game For an ARM-based Linux enthusiast or game developer, getting Steam + Box64 working and being able to run modern features like DLSS4 and multi-frame generation was an incredibly impressive experience. Box64 was working flawlessly here. Despite occasional glitches, and the realization that this setup isn’t what the GB10 was engineered for, this is the first time I’ve thought that a significant chunk of everyman x86 market share could evaporate overnight. It’s notable for me because it demonstrates the platform is flexible and it can easily pivot beyond its AI-centric raison d’être. … learn more about that here: Nvidia Spark GB10: MSI EdgeXpert running Steam Games ? Cyberpunk 2077, Doom Eternal and More? ? Quickie How-To https://forum.level1techs.com/t/nvidia-spark-gb10-msi-edgexpert-running-steam-games-cyberpunk-2077-doom-eternal-and-more-quickie-how-to/240557