We use cookies to make your experience better. To comply with the new e-Privacy directive, we need to ask for your consent to set the cookies. Learn more.
Blog
-
Posted: August 27, 2026Categories: General Topics and NewsRead more »
MGX power rack · H2 2026 No building rewiring 2 MW per rack row from 2027 80+ vendors building to the OCP specNVIDIA's 800 VDC architecture is usually read as a greenfield story. The commercially interesting part is the opposite: the MGX-compatible power rack shipping in the second half of 2026 is designed to drop into halls that are already built, delivering 800 VDC inside the row without touching the building's electrical system. Higher density becomes a row-level equipment decision instead of a multi-year rebuild.
Why it matters
Your existing hall is not a write-offThe limit is not how many watts the building can draw — it is how many conversion stages sit between the grid and the GPU. Fewer stages means more of the power you already pay for reaches the compute.
Inside the guide
Which limit binds firstQuick decision table by situation, the three-stage roadmap with dates, density bands where cooling and power each become the constraint, seven common mistakes, and what to send us for an assessment.
-
Posted: August 19, 2026Categories: General Topics and NewsRead more »
30B total · 3B active per query Up to 4× output speed 1M token context Routing cuts cost to about a thirdNVIDIA released Nemotron 3.5 Lightning and the NeMo Switchyard routing library on 11 August 2026, and Supermicro followed a day later with the hardware side. For infrastructure teams the question is not whether the model is good — it is what to run it on. The honest answer ranges from a single RTX PRO workstation to a rack-scale AI factory, and the variable that decides it is concurrency, not model size.
Why it matters
The sizing trap costs real money3B active parameters describe compute, not memory. All 30B still have to fit on the accelerator, and a 1M-token context makes the KV cache the dominant term. Size from the workflow, not the datasheet.
Inside the guide
From evaluation to AI factoryPlatform ladder by concurrency, seven common sizing mistakes, routing economics with partner benchmarks, air versus liquid thresholds, and exactly what to send us for a configuration.
-
Posted: August 10, 2026Categories: General Topics and NewsRead more »
AF-4 vs AF-6 vs AF-8 OSA vs ESA Usable capacity mathvSAN ReadyNode profiles look like model numbers, but AF-4, AF-6 and AF-8 are performance classes — and picking by name alone is how clusters end up over-bought on compute and under-bought on storage. This guide decodes the profiles, explains when hybrid still makes sense and why ESA changes the hardware question, and walks through the capacity math that turns raw terabytes into what your VMs can actually use.
Why it matters
Raw TB is not usable TBThe storage policy decides the budget: 100 TB raw is 50 TB usable with RAID-1 mirroring, 33 TB at FTT=2 — and up to 80 TB with RAID-5 erasure coding, if the cluster has enough hosts.
What the guide covers
From profile to quoteProfile-by-profile comparison, hybrid vs all-flash, OSA vs ESA host minimums, a full sizing walkthrough and the exact inputs to send us for a Supermicro ReadyNode configuration with confirmed EU lead time.
-
Posted: August 03, 2026Categories: AI InfrastructureRead more »
Native FP64 + AI in one node 1,152 Rubin GPUs · 576 Vera CPUs 3.2 MW Scalable Unit 362 kW racks · DLC-2Converged HPC and AI projects rarely stall on GPU availability. They stall because the data hall cannot power, cool, connect and commission the cluster that was ordered. Supermicro's NVIDIA Vera Rubin NVL4 DCBBS blueprint answers that as one system — 288 nodes, Quantum-X800 InfiniBand, direct liquid cooling, busbar power and L10/L11 validation in a repeatable 3.2 MW block you can replicate as the site grows.
Why it matters
A platform is not yet a deploymentNVIDIA defines the architecture; the blueprint defines how to rack, cool, cable, test and commission it. That is where facility mismatch, cooling shortfalls and slipped go-live dates get designed out — before the order is placed.
What the guide covers
From site survey to first jobScalable Unit anatomy, NVL4 node balance, memory and fabric, 362 kW cooling and power, the five-stage deployment lifecycle, NVL4 vs NVL72 vs HGX Rubin NVL8 vs GB200 NVL4, plus a facility planning checklist for European deployments.
-
Posted: July 29, 2026Categories: General Topics and NewsRead more »
8× 25GbE SFP28 on-SoC Intel vRAN Boost 399 mm · front I/OTelco and edge deployments rarely fail for lack of compute. They fail because the cabinet is too shallow, rear access is impossible, and every added NIC or accelerator becomes another integration problem. The Supermicro SYS-112D-40C-FN8P puts eight 25GbE ports and vRAN FEC acceleration inside the processor — so a 399 mm front-I/O 1U carries a full telecom workload with both PCIe slots still empty.
Why it matters
Three cards fewer per nodeEight 25GbE ports without add-in NICs, vRAN acceleration without a separate FEC card. Fewer part numbers, fewer spares and fewer failure points on every site you roll out.
What the guide covers
Ports, cabinets and rolloutFull port map, rack-fitting options including back-to-back mounting, four configuration examples, a comparison against edge boxes and rack servers, plus a cabinet deployment checklist.
-
Posted: July 14, 2026Categories: AI InfrastructureRead more »
Liquid cooling · PUE toward 1.05 TCO vs TCE Supermicro × Intel webinarData centers consume a growing share of the world's electricity, and AI is accelerating it. This guide — a companion to our Green Computing at Scale webinar with Supermicro and Intel — shows how European teams cut data center energy costs with liquid cooling, Titanium power supplies and resource-saving architecture: which measures fit your site, how to estimate the payback, and what to send us for an energy-efficient Supermicro configuration.
Why it matters
Energy is now a procurement decisionEU electricity prices, CSRD reporting and AI rack density make efficiency measurable money: Supermicro states up to 40% cooling energy savings and PUE toward 1.05 versus the ~1.6 industry average.
The buyer angle
From payback model to EU quoteA 5-number payback model, a decision table for your situation and the exact quote inputs — SERVER SIMPLY returns a configuration, energy comparison and confirmed EU lead time.
-
Posted: July 09, 2026Categories: General Topics and NewsRead more »
Intel Core Ultra Arc Pro B50 / B60 / B70 EU configuration & lead timeSupermicro has expanded its Intel-powered edge AI portfolio with the fanless SYS-E103-14P, the SYS-521AD-LN2 AI mini tower and refreshed short-depth edge platforms. This guide helps buyers compare which system fits industrial inference, local AI development, compact rack deployment and edge-to-core infrastructure planning.
Why it matters
Choose the right edge AI platform before deploymentHelps teams avoid mismatched GPU, power, cooling, form-factor and latency decisions when moving AI inference from the data center to factories, retail sites, vehicles and branch locations.
The buyer angle
Validate system, GPU and lead time in EuropeSERVER SIMPLY helps European teams configure Supermicro edge AI systems, compare Intel Arc Pro and NVIDIA options, confirm EU delivery and plan an upgrade path into central GPU infrastructure.
-
Posted: July 01, 2026Categories: General Topics and NewsRead more »
Liquid-to-Air · 200 kW No facility water loop N+1 redundant pumpsThe Supermicro L2A Sidecar CDU brings up to 200 kW of liquid cooling into existing air-cooled data centers without requiring a primary facility water loop. It is built for AI and HPC teams that need dense GPU racks now, but cannot wait for a full facility liquid-cooling project. The Sidecar sits beside the rack, runs a closed coolant loop to the servers and rejects heat into room air.
Why it matters
The retrofit bridge to liquid-cooled AI racksHelps deploy high-density AI or HPC racks in air-cooled rooms while avoiding cooling towers, primary water loops and facility-wide plumbing work.
The buyer angle
Validate CDU + GPU rack before purchaseSERVER SIMPLY helps European teams size the Sidecar to real rack heat load, configure matching GPU servers and validate EU delivery and lead times.
-
Posted: June 29, 2026Categories: General Topics and NewsRead more »
256 Vera CPUs · 22,528 cores Up to 400 TB LPDDR5X 48U · 100% liquid-cooledThe NVIDIA Vera CPU Rack is a dense, 100% liquid-cooled 48U MGX system that packs 256 Vera CPUs — 22,528 Olympus cores, up to 400 TB of LPDDR5X and 64 BlueField-4 DPUs — to run the CPU-side work behind agentic AI: code execution, tool calls and reinforcement-learning sandboxes across more than 22,500 parallel environments. It is built to run alongside GPU racks, not replace them.
Why it matters
The CPU execution layer for agentic AIStops the CPU side — sandboxes, tool calls and RL environments — from starving expensive GPU capacity, sharing the same liquid-cooled facility as Vera Rubin or GB300 racks.
The buyer angle
GA in 2026 · plan CPU + GPU togetherNeed agentic AI capacity before Vera ships? Deploy current-gen CPU and GPU now and phase in Vera — EU delivery and binding lead times from SERVER SIMPLY.
-
Posted: June 10, 2026Categories: General Topics and NewsRead more »
72× Instinct MI455X HBM4 · ROCm Open UALink + Ultra EthernetAt Computex 2026, Supermicro expanded its rack-scale AI portfolio with the AMD Helios platform — a 72-GPU double-width rack built on AMD Instinct MI455X GPUs, 6th Gen EPYC “Venice” CPUs and AMD Pensando networking, unified by the open ROCm stack. A single rack delivers up to 2.9 FP4 exaFLOPS, 31 TB of HBM4 and 1.4 PB/s of aggregate bandwidth at peak.
Why it matters
An open alternative to NVIDIA NVL72Scales up over Infinity Fabric + UALink and out over Ultra Ethernet, on ROCm — open standards instead of proprietary NVLink, InfiniBand and CUDA.
The buyer angle
432 GB HBM4 per GPU · GA in 2H 2026Per-GPU memory ahead of NVIDIA's B300 (288 GB). Need capacity now? Deploy AMD Instinct today and phase in Helios — EU delivery and lead times from SERVER SIMPLY.