UK’s trusted IT infrastructure partner since 2003
sales@servnetuk.com
0800 987 4111
Servnet
FinanceToolsConfiguratorGet in Touch
GPU servers · GX2550 M8s · GX2580 M8s · launched July 2026

Fujitsu PRIMERGY GPU servers
Eight PCIe GPUs, or eight HGX B300

The PRIMERGY GX family is Fsas's answer to AI infrastructure: a 4U server for up to eight PCIe GPUs for inference, rendering, visualisation and HPC, and an 8U HGX system for training at scale. Both are assembled in Japan and managed with the same ISM tooling as the rest of PRIMERGY.
DOUBLE-WIDTH GPUS, GX2550
8 × B300[src]
HGX GPUS, GX2580
2.3 TB[src]
HBM3E PER GX2580
CORES PER XEON 6900P
800 Gb/s[src]
NETWORKING, GX2580
2 Jul 2026[src]
PLANNED AVAILABILITY
Overview

The two PRIMERGY GPU servers

Fujitsu PRIMERGY GPU servers are the GX2550 M8s and GX2580 M8s, launched by Fsas Technologies in Munich on 2 July 2026 with general availability planned the same day through Fsas and certified partners[src].

Of the two PRIMERGY GPU servers, the GX2550 M8s pairs two Xeon 6900P processors, up to 128 cores each, with up to eight double-width PCIe GPUs on a dual-root PCIe 5.0 switch[src]. The GX2580 M8s pairs two Xeon 6700P processors with an NVIDIA HGX B300 board of eight GPUs, NVLink and NVSwitch[src]. Both are assembled at Fujitsu's Kasashima factory as part of what Fsas calls its Made in Japan initiative[src][src].

PRIMERGY GX2550 M8s against GX2580 M8sBoth list a 3-year onsite warranty and a service lifecycle of at least 5 years.GX2550 M8sGX2580 M8sHeight4U8UProcessors2 × Xeon 6900P2 × Xeon 6700PCores per CPU, max12886Memory, max6 TB, 24 DIMMs8 TB, 32 DIMMsGPUsUp to 8 PCIe, double-width8 × HGX B300GPU interconnectDual-root PCIe 5.0 switchNVLink 5 + NVSwitchPCIe 5.0 x16 slots13 full height2 full heightDrives8 × 2.5" NVMe, E3.S or E3.LUp to 8 × E1.S NVMePower supplies4 × 3,200 W, 3+16 × 6,600 W, 3+3Weight29.7 kg119 kgBoth list a 3-year onsite warranty and a service lifecycle of at least 5 years.
From the two GX datasheets. The split is the usual one in GPU servers: PCIe cards for flexibility and inference, an HGX board with NVLink for training large models.
GPU memory

How much model fits in GPU memory

GPU memory decides which models a server can hold without splitting them across machines. Fsas gives the GX2580 M8s total and the per-card memory of the GX2550 M8s options[src][src].
GPU memory per card and per system on PRIMERGY GXModel size is only part of it; context, batch size and precision change what fits.GPU memory, GBRTX PRO 6000 BSEper card, GDDR796 GBH200 NVLper card, HBM3e141 GB8 × RTX PRO 6000GX2550 M8s · our arithmetic768 GB8 × HGX B300GX2580 M8s · Fsas figure2.3 TBModel size is only part of it; context, batch size and precision change what fits.
Per-card memory from the GX2550 M8s datasheet's GPU list and the GX2580 M8s system total from its datasheet. The eight-card GX2550 figure assumes eight RTX PRO 6000 cards, which the datasheet supports; we confirm H200 NVL counts per configuration.

Our AI GPU calculator works out how much GPU memory a model needs at a given precision and context, so you can see whether a GX2550 is enough or a GX2580 is needed.

Facilities

Power and rack planning

The GX2580 M8s is the heaviest PRIMERGY model: 119 kg, against 105 kg for the eight-socket RX8770 M7, with six 6,600 W supplies[src][src].
PRIMERGY GX power supply capacityTwo GX2580 M8s can need more power than many existing racks are fed with.Kilowatts of supply with redundancy held backGX2550 M8s4 × 3,200 W, 3+19.6 kWGX2580 M8s6 × 6,600 W, 3+319.8 kWTwo GX2580 M8s can need more power than many existing racks are fed with.
Supply capacity with redundancy held back (three of four supplies on the GX2550, three of six on the GX2580), from the datasheets. This is supply rating, not measured draw, which depends on GPUs and load.
Example GX rack layoutsCheck floor loading as well as power: two GX2580 M8s alone weigh 238 kg.Inference: 4 × GX2550 M8s16U · up to 32 PCIe GPUs26U free to grow intoGX2550 M8s #1GX2550 M8s #2GX2550 M8s #3GX2550 M8s #416U of 42UTraining: 2 × GX2580 M8s16U · 16 HGX B300 GPUs26U free to grow intoGX2580 M8s #18 × HGX B300 · 119 kgGX2580 M8s #28 × HGX B300 · 119 kg16U of 42UCheck floor loading as well as power: two GX2580 M8s alone weigh 238 kg.
Two example layouts drawn to scale: four GX2550 M8s for PCIe-GPU inference, and two GX2580 M8s for training, each in 16U of a 42U rack. Power and cooling, not rack space, usually set the real limit.
Specifications

PRIMERGY GX specifications

SpecificationGX2550 M8sGX2580 M8s
Processors2 × Xeon 6900P, 7 models, 72 to 128 cores[src]2 × Xeon 6700P, 6 models, 48 to 86 cores[src]
Memory24 DIMMs, 512 GB to 6 TB; RDIMM 6,400 or MRDIMM 8,800 MT/s32 DIMMs; 4 TB at 6,400 MT/s (1DPC) or 8 TB at 6,000 MT/s (2DPC)
GPUsUp to 8 double-width PCIe: RTX PRO 6000 Blackwell Server Edition, H200 NVL8 × NVIDIA HGX B300 with NVLink and NVSwitch; up to 2.3 TB HBM3e
Expansion13 × PCIe 5.0 x16 full height2 × PCIe 5.0 x16 full height
Networking1 × 1 GbE management port; no onboard data LAN listed, so networking goes in the PCIe slots (our reading)[src]8 × NVIDIA ConnectX-8 SuperNIC, up to 800 Gb/s
Drives8 × 2.5-inch NVMe, E3.S or E3.LUp to 8 × E1.S NVMe
Power4 × 3,200 W Titanium, 3+16 × 6,600 W Titanium (96%), 3+3
Size and weight4U, 437 × 737 × 178 mm, 29.7 kg8U, 449 × 950 × 356 mm, 119 kg
Cooling10 fans12 fans
ManagementISM integration; remote, update and asset managementISM integration; remote, update and asset management
Warranty3 years onsite3 years onsite

From the GX2550 M8s and GX2580 M8s datasheets. The GX datasheets do not list the iRMC S6 controller found on other PRIMERGY models; they describe management through ISM.

Context

PRIMERGY GX and the alternatives

Smaller GPU builds

RX2540 M8

Up to two double-width or six single-width GPUs in a 2U general server. RX2540 M8 →
Supermicro

Fsas’s extended portfolio

Fsas also resells Supermicro edge, storage and HGX systems in Europe; these are Supermicro-built[src]. Supermicro →
Pooling

Share GPUs instead

If utilisation is the problem, pooling GPUs across hosts can beat buying more. GPU pooling calculator →
Supply & quote

Get a PRIMERGY GX server sized and quoted

  • We size GPU memory and count against your models and data before choosing GX2550 or GX2580.
  • We plan power, cooling and floor loading with you first.
  • We specify networking for multi-node clusters.
  • We quote the systems with GPUs and Fsas support.

Get a PRIMERGY or ETERNUS quote

Tell us what it will run and what it replaces. We configure the system and quote it with Fsas support.

Business enquiries only — please use your company email. Gmail, Hotmail and other personal addresses are not accepted.

We only ship hardware to the countries in the dropdown. Software and maintenance services are global.

We use these details to reply to your enquiry and prepare any quote. No marketing emails without your consent. Privacy Policy

No obligation · response within 4 business hours

FAQs

PRIMERGY GPU servers: common questions

What are the PRIMERGY GX servers?

Fsas Technologies’ two GPU servers for AI, HPC and data-intensive work, launched in Europe in July 2026. The 4U GX2550 M8s takes up to eight double-width PCIe GPUs; the 8U GX2580 M8s carries eight NVIDIA HGX B300 GPUs linked by NVLink and NVSwitch.

Which GPUs does the GX2550 M8s support?

Up to eight double-width PCIe GPU cards through a dual-root PCIe 5.0 switch design. Its datasheet names the NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB) and lists the H200 NVL (141 GB) as a GPU option. Confirm the card count for H200 NVL builds per configuration.

How much GPU memory does the GX2580 M8s have?

Up to 2.3 TB of HBM3e per system across its eight HGX B300 GPUs, which connect through fifth-generation NVLink at 1.8 TB/s with NVSwitch. Eight ConnectX-8 SuperNICs give up to 800 Gb/s of network bandwidth for clustering.

Where are the PRIMERGY GX servers made?

Fsas markets them as “Made in Japan”: its datasheets say they are assembled at the Fujitsu Kasashima factory in Japan, which Fsas presents as supply chain transparency for regulated industries.

How much power does a GX server need?

The GX2550 M8s has four 3,200 W Titanium supplies in a 3+1 arrangement and the GX2580 M8s six 6,600 W supplies in 3+3. That is 9.6 kW and 19.8 kW of supply capacity with redundancy held back, so rack power and cooling need planning before anything else.

Can Servnet supply PRIMERGY GX servers?

Yes. We can supply and quote both GX models with GPUs, networking and Fsas support, and help size them against your model and data before you commit. Fsas also resells Supermicro GPU systems in Europe; those are Supermicro products, not PRIMERGY.

The Fsas Technologies range

More PRIMERGY and ETERNUS pages

Keep exploring

Related Servnet guides, tools and hardware

Sources

Every specification on this page comes from Fsas Technologies' own datasheets, web pages and support portal, listed below with their dates. Where a datasheet and other Fsas material disagree, we use the datasheet. Figures are confirmed against the current revision when we quote.

  1. Data sheet: PRIMERGY GX2550 M8s · PDF dated 2 Oct 2026
  2. Data sheet: PRIMERGY GX2580 M8s · PDF dated 2 Oct 2026
  3. Fsas press release: Fsas Technologies launches "Made in Japan" PRIMERGY GX servers · Munich, 2 Jul 2026
  4. Positioning Card: PRIMERGY Server Family · PDF dated 2 Oct 2026

Fujitsu, PRIMERGY, ETERNUS and iRMC are trademarks of Fujitsu; Fsas Technologies is a Fujitsu Group company. Intel, Xeon, AMD, EPYC and NVIDIA are trademarks of their owners. Servnet is an independent UK IT reseller; this page is our own guide to the range.

Talk to a UK specialist

Get expert advice or a no-obligation quote — servers, storage, networking, maintenance, finance and cloud. We reply the same working day.

or call 0800 987 4111