StorageReview

Eight DGX Sparks on one shelf, all reading from the same Solidigm QLC array over 100GbE. We built shared storage for deskside AI so models and datasets live in the data center, and any Spark loads them with one mount line. Eight at once pulled 24.9GiB/s. Full build at the link in bio.
NVIDIA AI Infrastructure Solidigm

The NVIDIA RTX PRO 4500 Blackwell Server Edition delivered 2.7x the L4's output throughput in the same HPE ProLiant DL145 Gen11 edge server, for about 105W more at the wall.

Latency is where the gap is widest. The 4500 generates at 12 to 15ms per token against 35 to 40ms on the L4, and at 256 concurrent requests it kept first-token latency under a second on 256-token prompts while the L4 queued for 9 to 39 seconds. With 32GB and native FP4, it also ran Gemma 4 26B-A4B in NVFP4, a model the L4 can't run natively.

Full review at the link in bio.

Hewlett Packard Enterprise NVIDIA AI Infrastructure

The ASUS Ascent QN10 is the first mini PC built on Qualcomm's Snapdragon X2 Elite: 18 Oryon cores, an 80 TOPS NPU, and three USB4 ports in a 0.7-liter box you can hide behind a monitor. Check the teardown, where a v***r chamber and copper thermal pads on the M.2 drives show how ASUS keeps it near silent under load. Why it matters: this is the first Arm mini PC we've tested that beats x86 outright, 43% to 76% ahead of the NUC 14 Pro on our CPU suites, with local Stable Diffusion and LLM inference running on the NPU while the CPU sits idle. The trade-off is 32GB of soldered memory and a $1,349.99 starting price against a $700 NUC barebones. Full review at the link in bio.

ASUS

The top five slots on the QNAP TS-h966TX take 3.5-inch SATA hard drives for bulk capacity; the four trays underneath hold 2.5-inch U.2 NVMe SSDs that run as a fast pool or a cache. Around back are two Thunderbolt 4 ports for a direct workstation link, plus 10GbE and 2.5GbE for everyone else on the network. Six-core Core i3, DDR5, QuTS hero ZFS, and a tower 36% smaller than QNAP's last Thunderbolt NAS. Pre-order is open at $1,299. Full story at the link in bio.

QNAP

192GB of unified memory with up to 160GB for the GPU, in a 4.2 lb laptop. HP's new ZBook Ultra G3a 16 pairs the AMD Ryzen AI Max+ PRO 495 with the largest GPU memory allocation we've seen on a Ryzen AI Max system, and HP says that's enough for models up to 300B parameters locally. A new dual-fan thermal design holds 100W sustained, up from 55W on the last generation, with a user-replaceable 96Wh battery and a single Gen5 M.2 slot to 8TB. It ships in October, and that's when the benchmarks start. Our preview and teardown are at the link in bio. Z by HP

512GB on one DDR5 RDIMM. Micron has demonstrated the module on multiple server platforms, and AMD and Intel are both validating it for their next-generation servers. Twenty-four of them put 12TB of DRAM in a dual-socket box. The module is rated to 9,200 MT/s, and Micron says a single 512GB part draws 16W where four 128GB modules of the same capacity draw 44.2W. Volume production is planned for the second half of 2027. Full story at the link in bio. Micron Technology

Eaton's HDXL Rack PDU G4 is a zero-U unit rated for 56 to 81kW. Every one of the 24 outlets is Eaton's C39, a combination receptacle that takes either a C14 or a C20 plug, so a 3200W GPU server supply and a 1U switch cord fit the same . The colored labels along the strip mark the branch breakers, 21 of them for 24 outlets, and the module toward the end is the Gigabit Network Module with its LCD and per-outlet status LEDs. Built for racks where a single 4U GPU server can carry eight power cords. Full review at the link in bio.

Switch On to Eaton