What is the Rack-GPU Pro 64GB?

A 4U rack GPU server from Radius. Two AMD Radeon AI PRO R9700 cards provide 64 GB of VRAM for on-prem LLM inference; the host is a Ryzen 9 7900 with 32 GB of DDR5-5600.

Which models fit in VRAM?

By weight size, Llama 3.1 70B (about 40 GB) and Nemotron Super 49B (about 26 GB) fit in VRAM. Qwen3.5-122B-A10B int4 (71.5 GB) and gpt-oss-120B (66 GB) also use the 32 GB of host RAM. The figures leave out KV cache and working memory.

What is the hardware configuration?

An AMD Ryzen 9 7900 (12 cores / 24 threads), 32 GB of DDR5-5600, a B850 ATX motherboard, a 1 TB NVMe PCIe Gen4 SSD and a 1,600 W Platinum power supply. Three 140 mm fans cool the system in the Radius Rack-GPU 4U chassis.

How much memory bandwidth does it have?

The two cards' GDDR6 provides 1,280 GB/s in total (2 × 640 GB/s). That figure sets the ceiling for decode speed.

Which operating system does it ship with?

Radius delivers the system with Linux.

What does it cost and how fast is delivery?

The price is $7,950 + VAT, $9,540 with VAT. Radius assembles the system and delivers it within one week of the order; you order on RackDepo.

Which warranty applies?

Radius delivers the system with a 2-year Hardcare Basic warranty.

Loading 3D model

Rack-GPU Pro 64GB / 4U GPU server · AI workstation

Rack-GPU Pro 64GB

A 4U AI workstation for on-prem LLM inference: 2 × Radeon AI PRO R9700, 64 GB VRAM, 1,280 GB/s of memory bandwidth.

Price$7,950+ VAT$9,540 incl. VAT

System architecture · 4U

One host, two Radeon AI PRO R9700.

The host platform pairs a Ryzen 9 7900 on AM5 with 32 GB of DDR5-5600. The two R9700 provide 64 GB of VRAM in total, and a single 1,600 W Platinum PSU supplies the whole system.

32 GB GDDR6 per card: 2 × R9700
Total VRAM: 64 GB
Platinum · single PSU: 1,600 W
186 × 445 × 455 mm · rack rails: 4U

Accelerators

2 × Radeon AI PRO R9700

GDDR6 · 2 × 32 GB · 256-bit per card: 64 GB
Total bandwidth · 2 × 640: 1,280 GB/s
RDNA4 · 2 × 64: 128 CU
INT4 theoretical · 2 × 191: 382 TFLOPS

Host

Ryzen 9 7900 and 32 GB DDR5.

Ryzen 9 7900: 12 cores / 24 threads, 3.7–5.4 GHz, 64 MB L3 cache. The processor runs on a B850 ATX workstation motherboard with 2 × 16 GB of DDR5-5600.

VRAM · 2 × 32 GB GDDR6: 64 GB
Host RAM · DDR5-5600: 32 GB
Cores / threads: 12 / 24

Power

1,600 W Platinum, one power supply.

Each R9700 has a total board power (TBP) of 300 W. A single 1,600 W Platinum PSU supplies both cards and the host platform. The cards' combined 600 W load leaves 1,000 W of power headroom for the processor and the platform.

Card TBP / PSU capacity600 / 1,600 W

Two cards (TBP) · 600 W
Headroom for processor and platform · 1,000 W

Platinum · single PSU: 1,600 W
Card TBP: 2 × 300 W

Cooling

Front-to-back airflow

Three 140 mm fans take in cool air through the honeycomb front panel. The R9700 blower fans exhaust the heated air through the slot brackets, out of the chassis.

Air intake · front panel: 3 × 140 mm
Blower exhaust · slot bracket: 2 × 300 W

Components

Configuration

Processor: AMD Ryzen 9 790012 cores / 24 threads · 3.7–5.4 GHz · 64 MB L3
Memory: 32 GB DDR5-56002 × 16 GB
Accelerators: 2 × AMD Radeon AI PRO R970032 GB GDDR6 per card · 256-bit · 300 W
Motherboard: B850 · AM5 · ATXAMD AM5 workstation motherboard
Storage: 1 TB NVMe PCIe Gen4NVMe SSD
Power supply: 1600 W PlatinumSingle power supply
Chassis: Radius Rack-GPU 4U3 × 140 mm fans · 8 PCIe slot design · rack rails

Assembly

Assembly and delivery

Radius assembles the system and delivers it within a week of the order.

Rack-GPU Pro 64GB

2 × R9700
64 GB GDDR6
32 GB DDR5
1,600 W Platinum
4U

Price$7,950+ VAT$9,540 incl. VAT

Buy on RackDepo Specifications

Questions: 0 (216) 344 54 64

Inference · decode

Memory bandwidth sets the upper bound on decode speed.

At every decode step an LLM reads all of its active weights from VRAM, so the read rate sets the step time. In batched operation, concurrent requests share the same read.

1,280 GB/s of parallel GDDR6

Each R9700 accesses its own GDDR6 in parallel: 2 × 640 GB/s. The ASUS Ascent GX10 uses a single LPDDR5x pool shared between CPU and GPU.

tok/s≈bandwidth÷bytes / token

Rack-GPU Pro 64GB · 2 × 640 GB/s GDDR6, in parallel1,280 GB/s

ASUS Ascent GX10 · LPDDR5x, unified273 GB/s

Tensor parallel · TP=2

In tensor parallel, each card holds and computes half of every weight matrix. An all-reduce joins the partial results at the end of each layer, so the model uses both cards' VRAM and bandwidth together.

GPU0W / 2GPU1W / 2

all-reducePCIe

Model compatibility

Which models you can run

The Pro 64GB runs 70B-class dense models entirely in VRAM, and 120B-class MoE models with the support of 32 GB of host RAM. The chart sets reference model weights against 64 GB of VRAM and the 96 GB total with host RAM.

Weight size and memory

Weights only; KV cache and working memory excluded.

Runs in VRAMRuns with host RAM

64 GB · VRAM

Qwen3.8-Flash-Next W4A16169 GB

200B+ MoE>128 GB

Qwen3.5-122B-A10B int471.5 GB

gpt-oss-120B MXFP466 GB

Llama 3.1 70B (dense)~40 GB

Nemotron Super 49B (dense)~26 GB

96 GB · VRAM + 32 GB host RAM

Model	Weight size	Rack-GPU Pro 64GB
Qwen3.8-Flash-Next W4A16	169 GB	Choose the Pro 128GB
Qwen3.5-122B-A10B int4	71.5 GB	Runs with host RAM
gpt-oss-120B MXFP4	66 GB	Runs with host RAM
Llama 3.1 70B (dense)	~40 GB	Runs in VRAM
Nemotron Super 49B (dense)	~26 GB	Runs in VRAM
200B+ MoE	>128 GB	Choose the Pro 128GB

Compare the chassis, Pro 64GB and Pro 128GB

Datasheet

Specifications

Processor: AMD Ryzen 9 7900 · AM5 · 12 cores / 24 threads · 3.7–5.4 GHz · 64 MB L3
Memory: 32 GB DDR5-5600 (2 × 16 GB)
Accelerators: 2 × AMD Radeon AI PRO R9700 · 32 GB GDDR6 · 256-bit · 300 W
Total VRAM: 64 GB
Memory bandwidth: 1,280 GB/s (2 × 640 GB/s, in parallel)
Compute: 2 × 64 = 128 CU · theoretical INT4 382 TFLOPS
Motherboard: B850 · AMD AM5 · ATX workstation motherboard
Storage: 1 × 1 TB NVMe PCIe Gen4 SSD
Operating system: Linux
Power supply: 1,600 W Platinum · single power supply
Chassis: Radius Rack-GPU 4U · 3 × 140 mm high-performance quiet fans · 8 PCIe slot design · rack rails
Dimensions: 186 × 445 × 455 mm (H × W × D)
Warranty: 2 years Hardcare Basic
Delivery: Within a week of the order

FAQ

Frequently asked questions

What is the Rack-GPU Pro 64GB?: A 4U rack GPU server from Radius. Two AMD Radeon AI PRO R9700 cards provide 64 GB of VRAM for on-prem LLM inference; the host is a Ryzen 9 7900 with 32 GB of DDR5-5600.
Which models fit in VRAM?: By weight size, Llama 3.1 70B (about 40 GB) and Nemotron Super 49B (about 26 GB) fit in VRAM. Qwen3.5-122B-A10B int4 (71.5 GB) and gpt-oss-120B (66 GB) also use the 32 GB of host RAM. The figures leave out KV cache and working memory.
What is the hardware configuration?: An AMD Ryzen 9 7900 (12 cores / 24 threads), 32 GB of DDR5-5600, a B850 ATX motherboard, a 1 TB NVMe PCIe Gen4 SSD and a 1,600 W Platinum power supply. Three 140 mm fans cool the system in the Radius Rack-GPU 4U chassis.
How much memory bandwidth does it have?: The two cards' GDDR6 provides 1,280 GB/s in total (2 × 640 GB/s). That figure sets the ceiling for decode speed.
Which operating system does it ship with?: Radius delivers the system with Linux.
What does it cost and how fast is delivery?: The price is $7,950 + VAT, $9,540 with VAT. Radius assembles the system and delivers it within one week of the order; you order on RackDepo.
Which warranty applies?: Radius delivers the system with a 2-year Hardcare Basic warranty.

Buy

Rack-GPU Pro 64GB

Radius assembles and delivers the system. Place your order on RackDepo, and contact our sales team for configuration requests.

Price$7,950+ VAT$9,540 incl. VAT

Buy on RackDepo Contact sales

Questions: 0 (216) 344 54 64 · satis@radius.com.tr

Rack-GPU Pro 64GB AI workstation in its 4U rack chassis

Complete, assembled system
Delivery within a week
2 years Hardcare Basic warranty
Linux operating system

For the chassis alone, see the Rack-GPU 4U chassis page.

Before you decide, see the 4U GPU server comparison.