ChillStar IT Solution

NVIDIA GPUs for AI servers and workstations

We supply new NVIDIA RTX PRO 6000 Blackwell and H200 GPUs, as cards or built into complete systems, and help you choose the edition that suits your chassis, power and cooling. Prices and stock move quickly, so every order is quoted.

Request a GPU quote WeChat or email

At a glance

All three RTX PRO 6000 editions use the same GPU with 24,064 CUDA cores and 96 GB of memory. They differ in cooling, power and size, which decides the machine each one fits.

GPUMemoryBandwidthMax powerCoolingFits in
RTX PRO 6000 Blackwell Server Edition96 GB GDDR71.6 TB/s600 WPassiveRack servers
RTX PRO 6000 Blackwell Workstation Edition96 GB GDDR71.8 TB/s600 WFlow-through fansTower workstations
RTX PRO 6000 Blackwell Max-Q Workstation Edition96 GB GDDR71.8 TB/s300 WBlower fanMulti-GPU workstations
H200 NVL141 GB HBM3e4.8 TB/s600 WPassiveRack servers
H200 SXM (HGX H200 servers)141 GB HBM3e4.8 TB/s700 WServer-cooled moduleHGX 4- or 8-GPU servers

RTX PRO 6000 Blackwell

The diagrams show each card side-on, as installed, with the back of the case on the left.

RTX PRO 6000 Blackwell Server Edition

The data centre version of the RTX PRO 6000. A passive dual-slot card for rack servers with strong front-to-back airflow. Power is configurable up to 600 W to match what the chassis can cool.

No fans of its own: the server's fans push air through the heatsink, front to back.

Good fit: Multi-GPU servers for AI inference, virtual workstations, rendering and video processing.

Request a quote

Architecture
NVIDIA Blackwell
CUDA cores
24,064
GPU memory
96 GB GDDR7 with ECC
Memory bandwidth
1,597 GB/s
FP32
120 TFLOPS
AI (FP4, peak)
4 PFLOPS
Max power
Up to 600 W, configurable
Cooling
Passive
Size
Dual slot, 4.4 × 10.5 in
Interface
PCIe 5.0 x16
Multi-Instance GPU
Up to 4 × 24 GB

RTX PRO 6000 Blackwell Workstation Edition

The full 600 W desktop card, and the fastest of the three editions. Its double flow-through cooler suits a well-ventilated tower case. At 5.4 inches tall it is too tall for standard server GPU slots.

Two fans pull air up through the card and out the other side, so the case needs good top or rear exhaust.

Good fit: Tower workstations with one or two GPUs for AI development, fine-tuning, 3D and simulation.

Request a quote

Architecture
NVIDIA Blackwell
CUDA cores
24,064
GPU memory
96 GB GDDR7 with ECC
Memory bandwidth
1,792 GB/s
FP32
125 TFLOPS
AI (FP4, sparse)
4,000 TOPS
Max power
600 W
Cooling
Double flow-through fans
Size
Dual slot, 5.4 × 12 in
Interface
PCIe 5.0 x16
Multi-Instance GPU
Up to 4 × 24 GB, 2 × 48 GB
Display
4 × DisplayPort 2.1b

RTX PRO 6000 Blackwell Max-Q Workstation Edition

The same 96 GB GPU limited to 300 W, with a blower cooler that sends heat straight out of the case. Because each card exhausts its own air, four can sit side by side in one workstation.

A blower fan pushes hot air along the card and out through the bracket at the back of the case.

Good fit: Multi-GPU workstations, up to four cards, where total power and heat need to stay manageable.

Request a quote

Architecture
NVIDIA Blackwell
CUDA cores
24,064
GPU memory
96 GB GDDR7 with ECC
Memory bandwidth
1,792 GB/s
FP32
110 TFLOPS
AI (FP4, sparse)
3,511 TOPS
Max power
300 W
Cooling
Blower fan, rear exhaust
Size
Dual slot, 4.4 × 10.5 in
Interface
PCIe 5.0 x16
Multi-Instance GPU
Up to 4 × 24 GB, 2 × 48 GB
Display
4 × DisplayPort 2.1b

H200

Both versions have the same 141 GB of HBM3e memory. NVL is a PCIe card. SXM is a module built into HGX servers.

H200 NVL

A Hopper data centre GPU with 141 GB of HBM3e memory at 4.8 TB/s, the most memory per card on this page. That headroom matters when serving large language models. It is a passive dual-slot PCIe card for rack servers.

Passive, like the Server Edition. NVLink bridges join two or four cards so they can share work at high speed.

Good fit: LLM inference and fine-tuning on models that do not fit in 96 GB, and HPC work that needs FP64.

Request a quote

Architecture
NVIDIA Hopper
GPU memory
141 GB HBM3e
Memory bandwidth
4.8 TB/s
FP64
30 TFLOPS
FP32
60 TFLOPS
FP8 Tensor (sparse)
3,341 TFLOPS
Max power
Up to 600 W, configurable
Cooling
Passive
Interface
PCIe 5.0 x16, dual slot
NVLink
2- or 4-way bridge, 900 GB/s per GPU
Multi-Instance GPU
Up to 7 × 16.5 GB

H200 SXM (HGX H200 servers)

The module version of H200. It mounts on an NVIDIA HGX baseboard in 4- or 8-GPU servers, with full NVLink between every GPU and up to 700 W per GPU. It is supplied as part of a complete server, not as a separate card.

SXM modules sit flat on an HGX baseboard inside the server and are cooled by the server's own fans.

Good fit: Training and large-scale inference where every GPU needs fast links to every other GPU.

Request a quote

Architecture
NVIDIA Hopper
GPU memory
141 GB HBM3e per GPU
Memory bandwidth
4.8 TB/s per GPU
FP64
34 TFLOPS
FP32
67 TFLOPS
FP8 Tensor (sparse)
3,958 TFLOPS
Max power
Up to 700 W per GPU
NVLink
900 GB/s per GPU
Multi-Instance GPU
Up to 7 × 18 GB
Systems
HGX H200 with 4 or 8 GPUs

Complete GPU servers

We also supply complete systems built around these GPUs, with the CPU, memory, storage and power supplies matched to the cards. Typical builds:

  • Workstation with up to four RTX PRO 6000 Max-Q cards
  • 4U rack server with eight RTX PRO 6000 Server Edition cards
  • Rack server with four or eight H200 NVL cards and NVLink bridges
  • HGX H200 server with eight SXM GPUs

Eight GPUs at 600 W each draw 4.8 kW before counting the rest of the server, so we check your power circuits, rack and cooling before anything is ordered.

Before you order

Every quote names the exact part number, condition (new) and warranty terms, and includes GST. You get a tax invoice with every sale.

High-end GPUs can be subject to Australian and US export controls. For goods leaving Australia we confirm the destination and end use, and may ask for a signed end-use statement.

Specifications are taken from NVIDIA's published datasheets and may change. Performance figures are NVIDIA's peak numbers; FP4 and FP8 figures marked sparse assume sparsity. NVIDIA, RTX, Blackwell, Hopper, NVLink and HGX are trademarks of NVIDIA Corporation. ChillStar IT Solution is an independent business and is not affiliated with NVIDIA.

Get in touch

Tell us what you need and where you are. Email or WeChat, whichever suits you.

Email

oliver.feng@chillstar.org

WeChat 微信

LRRMANAGER

Request a quote

Address

26 Foundation Blvd
Burwood East VIC 3151

Languages

English and Mandarin. 欢迎用中文联系我们。

ABN

47 124 930 677