HBM-Class Xeons and 8TB of DDR5 in a 6U? TFNEW S768G5
TL;DR: TFNEW launched the S768G5, an enterprise GPU server, on September 21, 2026. A 6U AI supercomputing platform with up to 2x 4th/5th Gen Intel Xeon Scalable (64 cores per CPU, HBM support, 350 W), 32 DDR5 DIMM slots up to 8 TB with CXL 1.1, and up to eight 600 W GPUs in passthrough or Switch modules. Get specs and pricing: sales@tfnew.com.
Core Specifications (with Plain-Language Terms)
The S768G5 is the 6U AI supercomputing server from TF NEW's Intel line, built on 4th/5th Gen Intel Xeon Scalable processors — up to two CPUs at 64 cores each with HBM support. HBM (High Bandwidth Memory) on the processor die gives the CPU a wide, fast memory tier directly beside the cores, which benefits workloads that thrash the memory bus. Main memory is 32 DDR5 DIMM slots up to 5600 MT/s with RDIMM/LRDIMM support up to 8 TB, plus CXL 1.1 support — CXL (Compute Express Link) lets memory-capable devices attach over PCIe, so a server can extend its memory pool to CXL-attached devices, not just its own DIMMs. GPU support: up to eight high-performance cards at 600 W each, in an 8-card passthrough module (one GPU per PCIe path) or an 8-card Switch module (GPUs on a switch fabric). Storage: 12x 3.5-inch drives plus 2x 2280/22110 SATA or 1x NVMe M.2 SSD. PCIe expansion in passthrough reaches up to 11 standard slots.
Specs at a Glance
| Item | S768G5 Value |
|---|---|
| Chassis | 6U AI supercomputing server |
| GPU module | 8-card passthrough or 8-card Switch |
| Processor | Up to 2x 4th/5th Gen Xeon Scalable, 64 cores/CPU, HBM support, 350 W |
| Memory | 32 DDR5, up to 5600 MT/s, RDIMM/LRDIMM, up to 8 TB; CXL 1.1 |
| Storage | 12x 3.5" drives; 2x M.2 SATA or 1x M.2 NVMe |
| PCIe | Passthrough: up to 11 standard slots; Switch: per-scheme |
| GPU | Up to 8 cards at 600 W each |
Why a Memory Strategy Is What Distinguishes an 8-GPU Node
The S768G5 treats memory as three tiers rather than one: HBM on the Xeon die for the hottest, most-frequently-touched data; up to 8 TB of DDR5 in 32 slots for resident working sets; and CXL 1.1 for pooling memory beyond the DIMM map — a design that matters when an 8-GPU training run's optimizer states, datasets and checkpoints all compete for residence. The passthrough module gives frameworks the clean one-GPU-per-path topology they prefer; the Switch module trades lanes for fabric. Six units of rack, engineered around how AI actually consumes memory.
Where It Fits
- Large-model training: HBM-tiered CPUs with 8 TB DDR5 and 8 GPUs
- AI platform clusters: CXL 1.1 memory pooling for scale-out sharing
- Unified AI + HPC nodes: 64-core HBM-capable dual Xeon behind GPUs
- Mixed GPU-density sites: passthrough or Switch modules per workload
Why the Numbers Matter
- HBM support on Xeon: a memory tier beside the cores, not just DIMMs
- 32 DDR5 up to 8 TB: resident model and dataset capacity
- CXL 1.1: memory pooling beyond the local DIMM map
- Passthrough or Switch 8-GPU modules: topology matched to the workload
- 8x 600 W GPUs in 6U: accelerator density with memory depth
What TF NEW Says
"Eight GPUs make a headline; the memory map makes the machine," said a TF NEW product manager. "The S768G5 tiers it — HBM on 64-core Xeon dies, 8 TB of DDR5 in 32 slots, CXL 1.1 for pooling — then couples the fabric to eight 600 W GPUs in passthrough or Switch modules. That's the 6U AI node designed around how models eat memory."
Quality and Availability
Every S768G5 is burn-in tested under full GPU load — all PCIe paths (passthrough and switch), memory channels and drive bays validated for 48 hours before shipping. It ships deployment-ready with GPU module and memory configuration documentation.
Pricing and Availability
The S768G5 is available now. Pricing varies with processor, memory, GPU module and storage configuration. Contact sales@tfnew.com or use the request-a-quote form at www.tfnew.com/contact for configuration guidance and lead time.
About TF NEW
TF NEW (TF New Technology) is an enterprise AI infrastructure provider based at www.tfnew.com. It designs and delivers private AI, compute, networking, storage, cooling and edge systems as deployment-ready solutions — AI servers, AI workstations, edge AI, liquid cooling and enterprise rack servers — for AI factories, datacenters and edge projects worldwide.
Media Contact
TF NEW | sales@tfnew.com | www.tfnew.com/contact
FAQ
- What is the S768G5? A 6U AI supercomputing server with up to 2x 4th/5th Gen Xeon Scalable (HBM-capable) and eight 600 W GPUs.
- What GPU modules does it offer? An 8-card passthrough module or an 8-card Switch module.
- How much memory does it support? 32 DDR5 slots up to 5600 MT/s (RDIMM/LRDIMM), up to 8 TB, with CXL 1.1.
- What is HBM support? High Bandwidth Memory on the Xeon die gives the CPU a wide memory tier beside the cores.
- What storage does it have? 12x 3.5-inch drives plus M.2 SATA or NVMe slots.