Kentino AI Node V100 HGX-2 — 16× NVIDIA V100 SXM3 32 GB, 512 GB HBM2
Enterprise-grade platform, assembled and tested in the EU.
Kentino AI Node V100 HGX-2
512 GB HBM2, Sixteen GPUs, Full NVSwitch
16x V100 SXM3 32 GB | 124.8 TFLOPS FP64
The DGX-2 class machine. Built to order on a Supermicro host.
A 10U node carrying sixteen NVIDIA V100 SXM3 modules on an HGX-2 baseboard with full NVSwitch any-to-any connectivity: 512 GB of HBM2, 124.8 TFLOPS of FP64, and every GPU talking to every other at full bandwidth. We build these to order on a Supermicro SYS-9029GP-TNVRT host. It is the cheapest route we know to a half-terabyte NVSwitch-pooled VRAM space, and for double-precision HPC it remains a serious tool.
Hardware
| Component | Detail |
|---|---|
| GPUs | 16x NVIDIA V100 SXM3 32 GB HBM2 (Volta, 350 W per module) |
| VRAM pool | 512 GB HBM2, NVSwitch-pooled as one address space |
| Memory bandwidth | 900 GB/s per GPU |
| GPU interconnect | NVLink 2 via NVSwitch - full any-to-any across all 16 GPUs |
| Compute | 2,000 TFLOPS fp16 | 124.8 TFLOPS FP64 |
| Host | Supermicro SYS-9029GP-TNVRT, 10U |
| CPUs | 2x Intel Xeon Scalable, dual LGA3647 |
| System memory | 1.5 TB DDR4-2666 ECC RDIMM (24x 64 GB, all slots populated) |
| Storage | 2x 1.92 TB U.2 NVMe boot + 2x 3.84 TB U.2 NVMe data. Expandable |
| Networking | 2x dual-port 100 GbE |
| Power | ~6.6 kW sustained, redundant PSU |
| Build | Assembled, burned in and memtested by Kentino with a written report |
What it runs
Volta is fp16 and FP64 hardware. These are the workloads it is genuinely good at.
| Component | Detail |
|---|---|
| Double-precision HPC | CFD, FEA, molecular dynamics, climate - 124.8 TFLOPS FP64 per node |
| fp16 training, validated codes | architectures already tuned for Volta |
| Large-pool experiments | 512 GB NVSwitch-coherent VRAM where capacity matters more than speed |
| Teaching and research | real 16-GPU NVSwitch behaviour on a fixed grant budget |
| DGX-2 / HGX-2 estate expansion | node parity with Volta hardware you already run |
Throughput figures are published vendor specifications for this GPU generation, not measurements on this exact configuration. We will give you measured numbers for your workload before you commit.
Power and cooling
Sixteen modules at 350 W is 5.6 kW of GPU - the same wall draw as an eight-GPU Hopper node, for a fraction of the throughput. Budget a dedicated 32 A three-phase feed and roughly one node per 10 kW of rack. The 10U chassis also takes a quarter of a standard rack. It needs genuine datacentre front-to-back airflow and a contained cold aisle below 24 C for 24/7 operation.
Be clear about what this is
No bf16, no fp8, no INT8 tensor cores. Volta has none of them. Most current quantised inference stacks assume bf16 or better, so a large part of today's ecosystem either will not run or will run badly here. 512 GB sounds competitive against a newer node, but the pool is HBM2 at 900 GB/s with fp16-only tensor paths. Do not buy this as an LLM serving node on the capacity number alone - if that is your workload, ask us about the A100 node instead.
Built from sourced baseboards. There is no new-channel supply of HGX-2 baseboards. We source them, and if they are ex-datacentre every module gets fresh thermal pads and paste plus a 72 hour burn-in before it ships - degraded factory pads are the main failure mode on hardware of this age. Lead time depends on sourcing and we will give you a real date, not an optimistic one.
One genuine advantage: V100 predates the Advanced Computing Item thresholds, so there is no KYC form and no End User Certificate. If export paperwork is what is blocking you on newer hardware, this is the node that does not need it.
What you get from Kentino
We build, test and support the complete node: host selection, assembly, thermal work on the baseboard, burn-in with a written report, OS and driver stack pinned to a Volta-compatible CUDA version, rack and airflow guidance, EU-based logistics and warranty.
Priced on request. Tell us your workload and we will come back with a firm quote, a realistic lead time, and an honest view of whether this or a newer node is the better buy.
The questions buyers ask us most often before ordering a server.
How long does it take?
Machines built from components we hold ship quickly; anything requiring a specific GPU generation depends on supply. We give you a date before you pay, and if it moves we tell you rather than letting you find out.
Can the configuration be changed before you build it?
Almost always. GPUs, memory, storage and cooling are chosen per order, and the listed configuration is a starting point rather than a fixed package. If you need more VRAM, faster storage or a different cooling approach, say so before you order and we will quote the change.
Can I collect the server in person?
You can. Our warehouse is in Prague, and collection in person is welcome — most people who come use the visit to go through the machine with our engineer and ask the questions that are awkward over email. For orders within the Czech Republic we also try to deliver personally and walk you through the setup on site.
Can I talk to someone who actually understands the workload?
Yes. We have an engineer who works on AI systems specifically, not a general sales desk. If your question is about batch sizes, quantisation, interconnect or where your bottleneck will be, ask it — that conversation usually changes the configuration for the better.
Which model can I run on this configuration?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
How do you test a server before shipping?
We run it against actual AI workloads rather than synthetic scores: inference throughput, sustained load behaviour and thermals under continuous operation. You get the benchmark results with the machine, so the performance you were promised is the performance you can verify on day one.
Is the server ready to run when it arrives?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
Ships from our EU warehouse. Heavy items may require freight arrangement — contact us for a shipping quote and lead time. 2-year limited warranty with advanced RMA support; extended warranty available.
Not exactly what you need?
Tell us your workload and we'll spec this platform around it — GPUs, memory, storage and cooling matched to what you actually run.