No products in the cart.
256 Cores, 768GB VRAM: Johns Hopkins APL Builds Out AI Research Compute with a VRLA Tech 8-GPU EPYC Server
The Johns Hopkins University Applied Physics Laboratory (APL) is one of the largest university-affiliated research centers in the United States — more than 8,800 staff, 600 active programs, and a mission that spans national security, autonomous systems, space exploration, and health. When APL’s research teams needed dense on-premise GPU compute to run large-scale AI workloads, they turned to VRLA Tech. The build: a VRLA Tech 4U dual-EPYC GPU server with two AMD EPYC 9755 processors, 768GB of ECC DDR5, and eight NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs — 768GB of combined ECC GDDR7 VRAM in a single chassis.
|
768 GB Total ECC GDDR7 VRAM |
256 cores Dual EPYC 9755 / 512 threads |
768 GB DDR5 ECC System RAM |
The Challenge
APL’s research programs span autonomous perception, superhuman decision-making systems, human-machine teaming, and AI safety — all defined in its published AI Technology Roadmap. Many of these workloads share a common infrastructure problem: they require large models, multi-modal inputs, and long-context inference at a scale where cloud GPU instances are either cost-prohibitive, subject to data residency restrictions, or simply not available on demand at the throughput research teams need.
For programs working on national security, space, and autonomous systems, data control and computing availability aren’t nice-to-haves — they’re requirements. On-premise GPU infrastructure with enough VRAM to run frontier-scale models without quantization compression is the baseline.
The requirement was a single chassis that could handle the full stack: large-scale model inference across multiple researchers concurrently, training runs on custom datasets, and enough headroom to accommodate model sizes that will continue to grow over the system’s useful life. With 768GB of VRAM across eight GPUs, the system needed to serve all of that without becoming obsolete in two years.
The Build
VRLA Tech configured an 8-GPU 4U AMD EPYC GPU server on the ASUS ESC8000A-E13P chassis — a dual-socket 4U platform designed for high-density GPU compute. The configuration was spec’d to maximize VRAM headroom, memory bandwidth, and storage throughput simultaneously.
| Component | Specification |
|---|---|
| Chassis | ASUS ESC8000A-E13P — 4U, 8-bay 2.5″, dual-socket SP5, 8× PCIe Gen5 GPU slots |
| CPU | 2× AMD EPYC 9755 — 128 cores / 256 threads each (256C / 512T total), 2.7GHz base / 4.1GHz boost, Zen 5c, SP5 socket, 500W TDP, 512MB L3 cache per socket |
| Memory | 768GB (24× 32GB) DDR5 ECC RDIMM — 12-channel per socket |
| Boot / OS NVMe | 2× 1.92TB PCIe 4.0 M.2 NVMe (mirrored or RAID1 for OS resilience) |
| Data NVMe | 4× 7.68TB PCIe 4.0 NVMe U.2 SSD — 30.72TB raw dataset and checkpoint storage |
| GPU | 8× NVIDIA RTX PRO 6000 Blackwell Server Edition — 96GB ECC GDDR7 each, 768GB total, 1,597 GB/s memory bandwidth per card, up to 600W, blower-exhaust cooling |
| Networking | Dual 200GbE QSFP112 PCIe — high-bandwidth interconnect for distributed workloads and fast data ingestion |
| Power | 3+1 × 3200W PSU (240V required) — N+1 redundancy |
| OS | Ubuntu 24.04 LTS — preinstalled and configured |
| Warranty | 3-year parts warranty · Lifetime US-based engineer support |
“The RTX PRO 6000 Server Edition is the right card for this density. At 8 GPUs in a 4U chassis, blower-exhaust cooling is non-negotiable — it’s the only configuration that maintains thermal margins at sustained load.”
What the System Runs
Large-Scale Model Inference
With 768GB of pooled GDDR7 VRAM, the system can serve frontier-size open-weight models — models in the 400B+ parameter range in quantized formats, or models up to roughly 400B parameters in FP8 — without splitting across multiple systems. Research teams can query a single running model instance across concurrent sessions, or run multiple smaller models simultaneously for comparative evaluation and ensemble approaches.
Training on Custom Datasets
APL’s research requires training on domain-specific datasets — sensor data, multi-modal inputs, and classified corpora that cannot be sent to a third-party compute provider. The dual-EPYC platform provides the memory bandwidth and PCIe throughput to keep all eight GPUs saturated during training runs, and the 30TB+ NVMe storage array ensures dataset loading isn’t the bottleneck.
Multi-Modal and Autonomous Systems Research
APL’s AI roadmap specifically names autonomous perception, human-machine teaming, and neuro-inspired computing as research vectors. Multi-modal models — which process images, video, sensor data, and text together — are substantially more VRAM-intensive than text-only LLMs at equivalent parameter counts. The 8-GPU configuration provides enough headroom to run these workloads at scale without quantization compromises that would degrade research fidelity.
Why On-Premise Made Sense Here
APL operates across national security, space, and health programs — many of which work with data that cannot be processed on commercial cloud AI services. Even for research programs without classified data requirements, on-premise compute offers predictable capacity, no per-token or per-GPU-hour cost at inference time, and full control over the software stack.
Cloud GPU instances for 8× H100 or equivalent workloads run $20–40/hour at list price. A fully configured VRLA Tech 8-GPU EPYC server pays for itself in compute-hours within months at sustained research utilization — and every future workload runs on infrastructure the institution owns outright.
The dual 200GbE network fabric matters here too. Research environments that ingest large datasets, share storage across a cluster, or need high-throughput connections between GPU nodes benefit from fabric-grade networking built into the server itself — rather than a 10GbE afterthought. VRLA Tech’s full GPU server lineup covers configurations from 2U 4-GPU deployments up to the 4U 8-GPU platform APL is running.
Building AI infrastructure for a university lab or research center?
VRLA Tech is SAM.gov registered and accepts government and institutional purchase orders. We’ve built AI systems for national laboratories, defense research programs, and university applied science departments. Configurations range from single-GPU workstations for individual researchers to multi-node 8-GPU server clusters.
See our university and national laboratory AI infrastructure page →
Configure a GPU Server for Your Research Program
Dual-EPYC 8-GPU servers, single-socket 4-GPU configurations, and custom multi-node builds — spec’d and built in Los Angeles.
Frequently Asked Questions
Frequently Asked Questions
What GPU server configuration does Johns Hopkins APL use for AI research?
Johns Hopkins University Applied Physics Laboratory deployed a VRLA Tech 4U dual-EPYC GPU server built on the ASUS ESC8000A-E13P chassis with two AMD EPYC 9755 processors (128 cores each, 256 total), 768GB DDR5 ECC RDIMM, and eight NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs — 768GB of combined ECC GDDR7 VRAM. VRLA Tech builds and configures these systems in Los Angeles with a 3-year parts warranty and lifetime US-based engineer support. The system supports large-scale AI training, multi-modal inference, and long-context language model workloads at the research scale APL programs require.
What is the difference between the RTX PRO 6000 Server Edition and the Workstation Edition for a multi-GPU server?
The NVIDIA RTX PRO 6000 Blackwell Server Edition uses a blower-style cooler that exhausts heat directly out the rear of the server chassis, preventing thermal recirculation in dense 8-GPU configurations. The Workstation Edition uses an open-air cooler that exhausts into the chassis — a design that creates thermal management problems at high GPU density. For 8-GPU builds like the VRLA Tech 4U EPYC server, the Server Edition is the required card. VRLA Tech, based in Los Angeles since 2016, validates thermal profiles for each chassis and GPU combination before shipping, backed by a 3-year parts warranty and lifetime US-based engineer support.
Can a university research lab or federally funded R&D center buy directly from VRLA Tech?
Yes. VRLA Tech is SAM.gov registered, accepts government and institutional purchase orders, and has built AI infrastructure for Johns Hopkins University Applied Physics Laboratory, Los Alamos National Laboratory, George Washington University, General Dynamics, and Miami University. VRLA Tech is based in Los Angeles and has been building custom AI workstations and GPU servers since 2016. Every system ships with a 3-year parts warranty and lifetime US-based engineer support. Contact vrlatech.com/contact-us/ for a quote.
How much VRAM is available in an 8-GPU RTX PRO 6000 Blackwell server?
An 8-GPU VRLA Tech EPYC server equipped with NVIDIA RTX PRO 6000 Blackwell Server Edition cards provides 768GB of combined ECC GDDR7 VRAM — 96GB per card across eight cards, at 1,597 GB/s memory bandwidth per GPU. This is sufficient to run frontier-scale open-weight models in full precision, train large multi-modal models, or serve multiple concurrent AI workloads across research teams. VRLA Tech builds these systems in Los Angeles with a 3-year parts warranty and lifetime US-based engineer support.
VRLA Tech builds custom AI GPU servers and workstations in Los Angeles. Founded in 2016, VRLA Tech serves enterprise clients, national laboratories, defense contractors, and university research programs. SAM.gov registered. 3-year parts warranty, lifetime US-based engineer support.
