When comparing GPUs, check dedicated video memory, or VRAM, together with bandwidth and installation requirements. A100 PCIe and H100 SXM do not offer the same bandwidth despite both having 80GB configurations. RTX PRO 6000 workstation and server variants also need separate power and cooling checks.
Below, we compare NVIDIA products by dividing them into workstation, graphics/inference server, and HBM computation server . We’ve put together memory and bandwidth, high-bandwidth memory (HBM), virtual GPU (vGPU) and hardware partitioning (MIG) support, and public selling prices. Specifications are based on manufacturer documentation, and usage judgments are interpretations based thereon. We do not include direct benchmarks or user reviews.
Table of Contents
Before reading the table: The bandwidth is the manufacturer’s declared value of GPU local memory and is not an actual value. In this article, 1TB/s is converted to 1,000GB/s. Except for A16, capacity and power are based on one GPU.
Workstation RTX is NVIDIA’s professional product line, which doesn’t mean it’s all data center-specific GPUs. This is not an exhaustive list of region-specific product configuration codes (SKUs), mobile GPUs, GeForce, or all older Tesla products.
GPU product line and memory/connection method
Design objectives for workstations and servers
Whether you’re editing CAD drawings or working with 3D scenes and local AI models, look at monitor output, app compatibility, and in-case cooling. RTX A6000, RTX 6000 Ada, and RTX PRO Blackwell Workstation are in this category. This is usually an active cooling method that uses a fan on the card itself.
For servers, they are further divided depending on the tasks required. L4·L40S·RTX PRO Server handles a wide range of image processing, rendering, VDI, and inference and mainly uses GDDR. A100·H100·H200·B200·B300 are designed around AI calculation and HPC and use HBM.
Passive server cards do not have fans on the card, but instead rely on strong front-to-back airflow from the chassis. You need to check whether it can be installed in a desktop slot and whether it can be sufficiently cooled.
Material: NVIDIA RTX A6000 Datasheet · NVIDIA RTX PRO 6000 Blackwell Server Edition · NVIDIA L40S Official Specifications · NVIDIA A100 · PCIe / SXM Datasheet
Difference between using HBM and memory capacity
GDDR6·GDDR7 is memory used in workstations and general-purpose graphics servers. HBM2·HBM2e·HBM3·HBM3e·HBM4 secures high bandwidth with a wide interface. The RTX PRO 6000’s 96GB GDDR7 has a larger capacity than the H100’s 80GB HBM, but depending on the configuration, the H100’s bandwidth is higher. First, we check whether the model is included and consider the proportion of memory access.
MIG support cannot be determined based on memory type. Blackwell RTX PRO supports MIG while using GDDR7. On the other hand, RTX A6000 and RTX 6000 Ada do not support MIG even if they use ECC memory. Please check the required features separately in the support table for the exact product.
Resources: NVIDIA RTX PRO 6000 Blackwell Workstation Edition · NVIDIA RTX PRO 5000 Blackwell · 48GB / 72GB · NVIDIA MIG Supported GPU List · Updated 2026-10-13
Classification of VRAM bandwidth, PCIe, and NVLink
The value in the bandwidth table is , the rate at which the GPU accesses its VRAM. NVLink, which is used for communication between PCIe and GPU between CPU, system memory, and devices, is a different route. The H200’s VRAM 4.8TB/s and NVLink 900GB/s are not compared in the same channel. Also check if it is a two-way sum notation.
Even if you insert two 48GB cards, not all apps can use them like 96GB cards. It requires model partitioning or multi-GPU support in the renderer, and also incurs communication costs between devices. The sum of capacity and bandwidth across multiple cards is not read as single card performance.
Resources: NVIDIA H200 · SXM / NVL Specification · NVIDIA Blackwell Datasheet · HGX B200 Specification Page 8
Selection order: Required VRAM capacity → Allowable latency/throughput → Multi-GPU communication → Virtualization method → Cooling, power, total cost
Workstation GPU Comparison
For a workstation, first determine whether the model or scene fits in memory. Moving from 48GB to 72GB or 96GB can make larger workloads possible. The execution speed of a workload that already fits does not increase in proportion to memory capacity.
| product | Memory · HBM usage | Bandwidth GB/s | board power |
|---|---|---|---|
| RTX A6000 | 48GB GDDR6 HBM: No | 768 | 300W |
| RTX 6000 Ada | 48GB GDDR6 HBM: No | 960 | 300W |
| RTX PRO 4500 Blackwell Workstation | 32GB GDDR7 HBM: No | 896 | 200W |
| RTX PRO 5000 Blackwell 48GB | 48GB GDDR7 HBM: No | 1,344 | 300W |
| RTX PRO 5000 Blackwell 72GB | 72GB GDDR7 HBM: No | 1,344 | 300W |
| RTX PRO 6000 Blackwell Workstation | 96GB GDDR7 HBM: No | 1,792 | 600W |
| RTX PRO 6000 Blackwell Max-Q | 96GB GDDR7 HBM: No | 1,792 | 300W |
| product | generation | Check vGPU | MIG |
|---|---|---|---|
| RTX A6000 | Ampere | graphics | Not supported |
| RTX 6000 Ada | Ada | graphics | Not supported |
| RTX PRO 4500 Blackwell Workstation | Blackwell | Not listed | Up to 2 |
| RTX PRO 5000 Blackwell 48GB | Blackwell | Not listed | Up to 2 |
| RTX PRO 5000 Blackwell 72GB | Blackwell | Graphics · From 20.2 | Up to 2 |
| RTX PRO 6000 Blackwell Workstation | Blackwell | Not listed | Up to 4 |
| RTX PRO 6000 Blackwell Max-Q | Blackwell | Not listed | Up to 4 |
‘Graphics’ in the table refers to products on the NVIDIA vGPU Software list, and ‘Not Listed’ refers to cases where the corresponding SKU is not in the list at the confirmation date. Not listed does not mean no PCIe passthrough. MIG is written based on hardware support, and host conditions are checked in Chapter 4.
Resources: NVIDIA RTX A6000 Datasheet · NVIDIA Ada Professional GPU Architecture White Paper · NVIDIA RTX PRO 4500 Blackwell Workstation Datasheet · NVIDIA RTX PRO 5000 Blackwell · 48GB / 72GB · NVIDIA RTX PRO 6000 Blackwell Workstation Edition · NVIDIA RTX PRO 6000 Blackwell Max-Q Datasheet · NVIDIA vGPU Software Supported GPU List · Updated 2026-10-02 · NVIDIA MIG Supported GPU List · Updated 2026-10-13
Generational differences between RTX A6000 and RTX 6000 Ada
A6000 is Ampere’s 48GB GDDR6·768GB/s product, and RTX 6000 Ada is the same 48GB and 960GB/s. Tensor·RT core generations also change. If you are considering whether to secure capacity with a used A6000, check the price and the generation gap in rendering and AI kernels.
A6000 supports NVLink between two GPUs. However, even if there is a bridge, separate software support for memory integration is required. If you’re switching to an RTX 6000 Ada, also make sure your existing workload uses NVLink. It is difficult to decide on replacement just by looking at the difference in product name.
Resources: NVIDIA RTX A6000 Datasheet · NVIDIA Ada Professional GPU Architecture White Paper
RTX PRO 5000 48GB·72GB capacity and bandwidth
Both configurations have the same official bandwidth of 1,344GB/s and board power of 300W. The advantage of 72GB is that you can afford to keep large models and scenes on the GPU. If the task is already being processed within 48GB, performance is not calculated as 1.5 times because the capacity is 1.5 times larger.
Virtualization requires tighter separation of configurations. In the vGPU support list confirmed in October 2026, RTX PRO 5000 Blackwell Workstation Edition 72GB is listed as from 20.2. This should not extend to the 48GB model or all other RTX PRO Workstation models. The datasheet states that MIG offers configurations of up to 2×24GB on the 48GB model and up to 2×36GB on the 72GB model.
Resources: NVIDIA RTX PRO 5000 Blackwell · 48GB / 72GB · NVIDIA vGPU Software Supported GPU List · Updated 2026-10-02
Power conditions of RTX PRO 6000 Workstation·Max-Q
Workstation and Max-Q share 96GB GDDR7·1,792GB/s. The power conditions are 600W and 300W, respectively. In environments with multiple GPUs or power limitations, you can consider Max-Q. Sustained computation performance varies depending on app load, power limitations, and cooling, so it cannot be determined to be equal based on memory specifications alone.
96GB is useful for local large-scale language models (LLMs), rendering with large textures, and large-scale data exploration. On the other hand, in large-scale distributed learning, the inter-GPU connection structure becomes a more important condition in addition to single card capacity. It should also be distinguished from the RTX PRO 6000 for servers. The bandwidth based on the official page is Workstation·Max-Q 1,792GB/s, Server 1,597GB/s.
Resources: NVIDIA RTX PRO 6000 Blackwell Workstation Edition · NVIDIA RTX PRO 6000 Blackwell Max-Q Datasheet · NVIDIA RTX PRO 6000 Blackwell Server Edition
If you plan to use MIG on your desktop: RTX PRO 6000 Workstation·Max-Q requires the procedure of changing the display mode to compute before activating MIG. At this time, the physical screen output of the card may be turned off. The configuration must include a separate GPU for the monitor, remote access means, and driver/vBIOS conditions.
Resources: NVIDIA MIG setup prerequisites · Workstation display mode
Server GPU vs. HBM
GDDR server GPU: A product line that processes VDI, video, rendering, and inference together
Not all servers that use HBM use HBM. A2·A10·A16·A40, L4·L40·L40S, and RTX PRO Server are GDDR based. It can be reviewed first in servers that require graphics and media capabilities, desktops for multiple users, and environments that run inference and rendering together.
| product | Memory · HBM usage | Bandwidth GB/s | board power |
|---|---|---|---|
| A2 | 16GB GDDR6 HBM: No | 200 | 40–60W |
| A10 | 24GB GDDR6 HBM: No | 600 | 150W |
| A16 board | 4 × 16GB GDDR6 HBM: No | 232 per GPU | 250W overall on board |
| A40 | 48GB GDDR6 HBM: No | 696 | 300W |
| L4 | 24GB GDDR6 HBM: No | 300 | 72W |
| L40 | 48GB GDDR6 HBM: No | 864 | 300W |
| L40S | 48GB GDDR6 HBM: No | 864 | 350W |
| RTX PRO 4500 Blackwell Server | 32GB GDDR7 HBM: No | 800 | 165W |
| RTX PRO 6000 Blackwell Server | 96GB GDDR7 HBM: No | 1,597 | Up to 600W |
| product | generation | Check vGPU | MIG |
|---|---|---|---|
| A2 | Ampere | graphics | Not supported |
| A10 | Ampere | graphics | Not supported |
| A16 board | Ampere | graphics | Not supported |
| A40 | Ampere | graphics | Not supported |
| L4 | Ada | graphics | Not supported |
| L40 | Ada | graphics | Not supported |
| L40S | Ada | graphics | Not supported |
| RTX PRO 4500 Blackwell Server | Blackwell | Graphics/Operation | Up to 2 |
| RTX PRO 6000 Blackwell Server | Blackwell | Graphics/Operation | Up to 4 |
Resources: NVIDIA A2 official specifications·NVIDIA A10 official specifications·NVIDIA Ampere product comparison table·A16 included· NVIDIA A40 Datasheet · Lenovo Press · NVIDIA L4 Product Guide · NVIDIA L40 Datasheet · NVIDIA L40S Official Specifications · NVIDIA RTX PRO 4500 Blackwell Server Edition · NVIDIA RTX PRO 6000 Blackwell Server Edition · NVIDIA vGPU Software Supported GPU List · 2026-10-02 Update
A16’s 64GB is the combined capacity of four 16GB GPUs. bandwidth of 232GB/s is also based on one GPU. We don’t calculate that you can load a single 64GB model as is. Distinguish between configurations intended to accommodate multiple VDI users and loading of a single large model.
L4 is a combination of 24GB·300GB/s·72W. It is suitable for inference and image processing servers where low power and low slot occupancy are important. L40S is close to a general-purpose accelerator that handles larger tasks at 48GB·864GB/s·350W.
Neither product supports MIG. If you plan to split across multiple virtual machines (VMs), you should review supported vGPU configurations.
Although L40 and L40S have the same memory capacity and bandwidth, their operation specifications and power conditions are different. You should not treat them as the same product just by looking at the memory table. The RTX PRO 4500 Server is a single-slot passive product with 32GB·800GB/s·165W, which is different from the Workstation version’s 896GB/s·200W. The RTX PRO 6000 Server is a top choice with 96GB and MIG, but it is not an HBM product.
Resources: NVIDIA Ampere Product Comparison Table · Includes A16 · Lenovo Press · NVIDIA L4 Product Guide · NVIDIA L40S Official Specifications · NVIDIA L40 Datasheet · NVIDIA RTX PRO 4500 Blackwell Server Edition · NVIDIA RTX PRO 4500 Blackwell Workstation Datasheet · NVIDIA RTX PRO 6000 Blackwell Server Edition
HBM Server GPU: Specifications for PCIe, SXM, and HGX
PCIe and SXM·HGX differ in the system to be installed. PCIe is placed in an expansion slot, and SXM/HGX modules require a dedicated baseboard and cooling/power design. Even in the quotation, individual modules, PCIe cards, and complete servers are not listed at the same price.
| product | Memory · HBM usage | Bandwidth GB/s | board power |
|---|---|---|---|
| A30 | 24GB HBM2 HBM: Yes | 933 | 165W |
| A100 40GB | 40GB HBM2 HBM: Yes | 1,555 | 250W |
| A100 40GB | 40GB HBM2 HBM: Yes | 1,555 | 400W |
| A100 80GB | 80GB HBM2e HBM: Yes | 1,935 | 300W |
| A100 80GB | 80GB HBM2e HBM: Yes | 2,039 | 400W |
| H100 80GB | 80GB HBM2e HBM: Yes | 2,000 | Up to 350W |
| H100 80GB | 80GB HBM3 HBM: Yes | 3,350 | Up to 700W |
| H100 NVL | 94GB HBM3 HBM: Yes | About 3,900 | 350–400W |
| H200 SXM | 141GB HBM3e HBM: Yes | 4,800 | Up to 700W |
| H200 NVL | 141GB HBM3e HBM: Yes | 4,800 | Up to 600W |
| GPU in HGX B200 | 180GB HBM3e HBM: Yes | 7,700 | Up to 1,000W |
| GPU in HGX B300 | 270GB HBM3e HBM: Yes | 7,700 | Up to 1,100W |
| product | form | Check vGPU | MIG |
|---|---|---|---|
| A30 | PCIe | operation | Up to 4 |
| A100 40GB | PCIe | operation | Up to 7 |
| A100 40GB | SXM4 | operation | Up to 7 |
| A100 80GB | PCIe | operation | Up to 7 |
| A100 80GB | SXM4 | operation | Up to 7 |
| H100 80GB | PCIe | operation | Up to 7 |
| H100 80GB | SXM5 | operation | Up to 7 |
| H100 NVL | Based on 1 PCIe GPU | operation | Up to 7 |
| H200 SXM | SXM | operation | Up to 7 |
| H200 NVL | PCIe | operation | Up to 7 |
| GPU in HGX B200 | Module for HGX | operation | Up to 7 |
| GPU in HGX B300 | Module for HGX | operation | Up to 7 |
Materials: NVIDIA A30 Datasheet · NVIDIA A100 · PCIe / SXM Datasheet · NVIDIA Hopper Architecture · PCIe H100 HBM2e Description·NVIDIA H100·SXM/NVL Specification·NVIDIA H200·SXM/NVL Specification·NVIDIA Blackwell Datasheet · HGX B200 Specifications 8 pages · NVIDIA Blackwell Ultra Datasheet · HGX B300 Specifications 5 pages · NVIDIA AI Enterprise 8 · Compute vGPU hardware-specific documents
H100 PCIe 80GB is HBM2e, H100 SXM 80GB is HBM3, the bandwidth is about 2TB/s and 3.35TB/s respectively. To compare memory specifications accurately, you must read the form after the name H100.
The H100 NVL is 94GB HBM3 per GPU, and the 188GB notation is the sum of both GPU configurations. Approximately 3.9TB/s in the table is based on a single GPU.
The A100 80GB also has 1,935GB/s PCIe and 2,039GB/s SXM. H200 provides 141GB HBM3e·4.8TB/s for both SXM and NVL, but the maximum power and connection method are different. The PCIe format of the H200 NVL is meaningful in consideration of utilizing existing enterprise servers, but the power, cooling, and compatibility of the 600W GPU must be met separately.
Resources: NVIDIA Hopper Architecture · PCIe H100 HBM2e Description · NVIDIA H100 · SXM / NVL Specification · NVIDIA A100 · PCIe / SXM Datasheet · NVIDIA H200 · SXM / NVL Specification
B200·B300 Standard for writing down numbers: This table applies the HGX individual GPU items from the current NVIDIA public data sheet. The B200 is 180GB, 7.7TB/s, and up to 1,000W, and the B300 is 270GB, 7.7TB/s, and up to 1,100W. In the same data, the GPU for GB200 NVL72 is different at 186GB·8TB/s, and the GPU for GB300 NVL72 is at 279GB·8TB/s.
We did not arbitrarily mix the seller’s 192GB/288GB indications or other system figures. In actual ordering, the shipping SKU and applicable data sheet revision must be confirmed together.
Resources: NVIDIA Blackwell Datasheet · HGX B200 Specifications Page 8 · NVIDIA Blackwell Ultra Datasheet · HGX B300 Specifications Page 5
What you see if you put memory bandwidth in the same unit
If you divide the H200’s 4,800GB/s by the L40S’s 864GB/s, it is approximately 5.56 times. This does not mean that inference will be faster by the same percentage. The proportion of bottlenecks varies between token generation and input processing in small batches and in large batches. After matching the model, precision, context, and number of concurrent requests, latency and throughput are measured.
Scientific calculations using double-precision floating point (FP64) should also be considered separately. Products with high AI TOPS are not necessarily best suited for all HPC codes. The HGX B300 discrete GPU FP64/FP64 Tensor value in the Blackwell Ultra datasheet is listed at 1.2 TFLOPS, while the HGX B200 in the Blackwell datasheet is listed at 37 TFLOPS. Ignoring this difference, it doesn’t make sense to replace FP64 workloads just because the generation is new.
Resources: NVIDIA Blackwell Datasheet · HGX B200 Specifications Page 8 · NVIDIA Blackwell Ultra Datasheet · HGX B300 Specifications Page 5
Latest generation 2026 See also: Rubin
NVIDIA’s current HGX page lists the Rubin discrete GPU as 288GB HBM4·22TB/s. However, Rubin’s commercial virtualization configuration has not been confirmed based solely on the MIG support list and vGPU hardware documentation confirmed in this article. Therefore, it was not included in the operation comparison table above, but was classified as a reference item for the latest generation. Without confirmation of public pricing and hypervisor support, we avoided giving it the same score as a purchasable alternative.
Resources: NVIDIA HGX · Rubin Discrete GPU Public Specification · NVIDIA MIG Supported GPU List · Updated 2026-10-13 · NVIDIA AI Enterprise 8 · Compute vGPU By Hardware Document
vGPU·MIG support conditions
vGPU is a way to provide GPUs to VMs, MIG is a hardware partitioning feature.
vGPU is a software/driver-based function that provides physical GPU resources to VMs. MIG is a hardware function that creates instances with separate compute and memory-related resources. They can be used together, but you must check each for support and operating conditions. Just because you support one doesn’t mean you can use the other features or all VM combinations.
| way | resource allocation | suitable situation | Check before purchasing |
|---|---|---|---|
| GPU passthrough | Assign physical GPUs directly to specific VMs | When one VM uses the entire GPU | IOMMU, server, hypervisor, driver compatibility |
| Time-sliced vGPU | Memory profile and GPU execution time distribution for each VM | VDI, virtual workstations, flexible sharing | vGPU families, licenses, and supported hypervisors |
| Bare Metal / Container MIG | Separation of hardware resources per GPU instance | Separation of inference services, resource allocation per researcher | Configuration of profiles and media/computing resources for each GPU |
| MIG-backed vGPU | Providing GPUs to VMs based on MIG instances | When VM-level isolation and fixed resources are required | GPU, host, guest OS, vGPU release combination |
Resources: NVIDIA AI Enterprise 8 · Compute vGPU hardware-specific documentation · NVIDIA AI Enterprise 8 · MIG-backed vGPU limitations
Graphics vGPU·AI Enterprise calculation support document
L4·L40·L40S·A2·A10·A16·A40·RTX A6000·RTX 6000 Ada can be found in the NVIDIA vGPU Software support list. Also includes RTX PRO 6000 Server and RTX PRO 4500 Server. On the other hand, compute vGPU support for A30·A100·H100·H200·B200·B300 must be checked in the NVIDIA AI Enterprise document. You should not conclude that compute vGPUs are not possible just because they are not on the list of general graphics vGPUs.
In the previous table, ‘graphics’ and ‘operation’ refer to the document path used for verification. This does not mean that the GPUs on the graphics list cannot use CUDA, NVIDIA’s GPU parallel computing platform, nor does it mean that all graphics products support all combinations of virtual PCs (vPCs), virtual apps (vApps), and virtual workstations (RTX vWS). Server certification, vGPU release, host driver, guest driver, hypervisor and OS must be verified in one configuration.
Resources: NVIDIA vGPU Software Supported GPU list · Updated 2026-10-02 · NVIDIA AI Enterprise 8 · Computation vGPU hardware-specific documents
MIG support depends on the model as well as its generation. In Ampere, A30 supports up to 4 instances and A100 up to 7; A2, A10, A16 and A40 do not support MIG. Ada L4, L40, L40S and RTX 6000 Ada also lack MIG support. In Blackwell RTX PRO, the 6000 series supports up to 4 instances, while the 5000 and 4500 series support up to 2.
Resources: NVIDIA MIG Supported GPU List · Updated 2026-10-13 · NVIDIA RTX PRO 4500 Blackwell Server Edition · NVIDIA Blackwell Ultra Datasheet · HGX B300 Specifications Page 5
Things to keep in mind when interpreting MIG instance counts
The explanation of dividing the A100 into up to 7 does not mean dividing 80GB into seven arbitrary capacities. You need to check the profile and placement combination for each GPU. Even if the memory size is the same, the number of SMs and media engine allocation may be different. Since system memory, storage, and network are not separated, service delay times are also checked separately.
If you’re planning a Windows VM, you should read the limitations one more time. According to the AI Enterprise 8 document, MIG-backed vGPU of A30·A100·H100·H200·B200·B300 does not support Windows guests. This document guides time-sliced vGPU to Windows guests of these GPUs. RTX PRO 6000 Server and RTX PRO 4500 Server are separately listed as supporting MIG-backed vGPU for Windows guests.
Resources: NVIDIA AI Enterprise 8 · MIG-backed vGPU limitations
Using MIG on bare metal Linux or containers does not automatically require a vGPU license for the VM. When providing vGPUs to a VM, you must check the licensing and operating terms of the applicable product family. Live migration is also a feature only available in supported combinations, and you should not assume that you can freely move between different MIG profiles.
Resources: NVIDIA MIG Settings Prerequisites · Workstation Display Mode · NVIDIA AI Enterprise 8 · MIG-backed vGPU Limitations
Include in your quote the SKU of your GPU, server model, hypervisor version, driver release, and guest OS. Also check the required vGPU profile and license type and period. In the case of MIG, it is easy to compare under the same conditions by leaving down memory, computing resources, and media engine requirements for each instance.
Public pricing and system-wide costs
Approximate prices based on public sales examples
The amounts below are rounded reference prices from public overseas sales pages checked on October 18, 2026. They are not market-wide minimums, averages or a uniform official NVIDIA price list. Some listings require stock confirmation or quote prices by order quantity. Where an older search snippet differed from the product page, the opened product page took precedence.
Prices are in US dollars and exclude Korean VAT, shipping and import costs. Won conversions use an assumed comparison rate of USD 1 = KRW 1,400. This is not a claim about the actual exchange rate on the check date. Use the applicable payment exchange rate and supplier quotation when ordering.
| Product/Configuration | Approximate price USD · Assumption exchange rate KRW | Terms of Sale / Interpretation | Source |
|---|---|---|---|
| RTX A6000 | From about $5,500 From about 7.7 million won | New product revealed starting price $5,460, check final quote | Supplier Marked Price |
| RTX 6000 Ada | Individual quote — | Numbers omitted due to differences between past prices in the index and current sales list. | Sale List |
| RTX PRO 4500 Workstation | Approximately $5,600 Approximately 7.8 million won | New product list price $5,550, receipt confirmation required | Product page |
| RTX PRO 5000 48GB | Approximately $9,500 Approximately 13.3 million won | PNY New List Price $9,450 | Product page |
| RTX PRO 5000 72GB | Approximately $11,600 Approximately 16.2 million won | New product release price for NVIDIA SKUs | Sale List |
| RTX PRO 6000 Workstation | Approximately $16,000 Approximately 22.4 million won | NVIDIA Marketplace display price, out of stock display | official sales page |
| RTX PRO 6000 Max-Q | Approximately $24,400 Approximately 34.2 million won | NVIDIA SKU new products individually marked, need to confirm receipt | Product page |
| A2 / A10 / A16 | From about $1,500 / $3,400 / $4,200 each From about 2.1 million / 4.8 million / 5.9 million won each | Supplier disclosure starting price, individual SKU/status confirmation | Supplier List |
| A40 / L40 | Starting from approximately $5,700 each / $7,800 Starting from approximately 8 million won each / 10.9 million won | A40 supplier starting price / L40 refurbishment case | A40 · L40 |
| L4 | Approximately $4,300 Approximately 6 million won | NVIDIA SKU new product display price | Product page |
| L40S | Approximately $9,000 / $12,000 Approximately 12.6 million / 16.8 million won | NVIDIA SKU Refurbished $8,995 / New $11,995 Case | Sale List |
| RTX PRO 4500 Server | Individual quote — | Lack of confirmation of public unit price for same SKU/condition | Official Product Information |
| RTX PRO 6000 Server | Approximately $19,500–20,500 / GPU Approximately KRW 27.3–28.7 million / GPU | PNY kit, minimum order of 8 units, unit price per quantity section | Unit price by quantity |
| A30 / A100 40GB PCIe / A100 80GB PCIe | From about $5,700 / $9,400 / $15,000 each about 8 million / 13.2 million / from 21 million won each | Supplier disclosure starting price, separate from SXM price | Supplier List |
| H100 80GB PCIe | From about $34,500 / $65,000 Case From about 48.3 million won / 91 million won Case | Check different supplier markings, inventory, warranty, and configuration | Starting price · Individual display price |
| H200 NVL | Approximately $33,000–34,500 / GPU Approximately KRW 46.2–48.3 million / GPU | Minimum order of 4, unit price ranges from 4 to 49. | Unit price by quantity |
| HGX B200 8-GPU Server | Approximately $419,000–433,000 / Server Approximately KRW 587–606 million / Server | Exxact public configuration example including CPU, memory, etc. | Server configuration price |
| HGX B300 8-GPU Server | From about $497,000 / Server From about 696 million won / Server | BIZON public starting price, final specifications and delivery quotation required | Server Sales List |
Not only whether it is new or refurbished, but also the OEM kit, warranty, software period, quantity and delivery also affect the price. The two H100 examples in the table also do not represent a general price range. This is reference material showing that it is difficult to determine the appropriate purchase price based solely on the seller’s publicly listed price. This does not imply endorsement of any particular seller or purchase.
A100 SXM·H100 SXM·H100 NVL·H200 SXM has not been converted to the above individual price. This is to avoid mixed quotes for GPU modules, bridges, HGX baseboards, and complete servers. For the B200 and B300, the server price divided by 8 was not used as the ‘selling price of one GPU’. Rubin is not offering a dollar amount because he has not been able to identify a comparable public purchase price.
GPU selection criteria for each application
| work | Products to compare first | the core of judgment |
|---|---|---|
| CAD·DCC·3D·Video editing | RTX A6000 / RTX 6000 Ada / RTX PRO 4500·5000·6000 Workstation | Project VRAM, ISV certification, RT/media functions, noise/power |
| Local LLM/Individual Study | RTX PRO 5000 48·72GB / RTX PRO 6000 96GB | Is the model and KV cache included, and is the target response speed? |
| Low-power inference and video analysis | A2 / L4 / RTX PRO 4500 Server | Power budget, media engine, model size, concurrent streams |
| Office VDI/Remote Graphics | A16 / L4 / RTX PRO 4500 Server / L40S / RTX PRO 6000 Server | VRAM, screen resolution, encoding, vGPU license per user |
| LLM for services·Multi-user reasoning | L40S / RTX PRO 6000 Server / H100 / H200 / B200 / B300 | Actual throughput, p95 latency, cache capacity, isolation conditions |
| Large-scale multi-GPU training | H100·H200 SXM / HGX B200·B300 | Inter-GPU communication, inter-node network, learning precision/kernel support |
| FP64-centric scientific computing | Comparison by code from A100 / H100 / H200 / B200 | FP64·FP64 Tensor specifications, libraries, memory and communication bottlenecks |
The order of review below is based on published specifications and is not a ranking of measured performance. If a smaller model runs at the speeds you need at L4, there may be little reason to spend more on the H200’s bandwidth. Conversely, if the model and cache exceed 48GB, in addition to the unit cost of the card, multiple partitioning, communication, and operational complexities must be calculated.
VRAM calculation with 70B model
With a simple weight calculation, if 70 billion parameters are stored as 2 bytes of 16-bit floating point (FP16), it is about 140 GB, and if they are stored as 4-bit 0.5 bytes, it is about 35 GB. This adds quantization metadata, runtime workspace and KV cache. Therefore, you cannot conclude that ‘48GB is always sufficient because it is 70B 4-bit’ or ‘one 141GB H200 is sufficient for 70B FP16 service.’
Actual memory requirements depend on model structure and quantization, context, and number of concurrent requests. Check whether the weights are loaded and whether the long text and multi-user services operate stably. KV cache is attention key/value data that stores the previous token so that it is not recalculated. If your request is longer, you should also consider this space.
Final quote including GPU, software, and facilities
Once you have decided on a GPU, you need to add up not only the server’s slots, auxiliary power, power supply unit (PSU), and cooling, but also the CPU, RAM, NVMe storage, network adapter (NIC), switch, optical module/cable, rack power, and support contract. In virtualized environments, also check how vGPU or AI Enterprise licenses are applied and renewal costs. For products that are advertised as including a software subscription, such as the H200 NVL, it is best to check the terms and rights included in the actual purchased SKU in the contract.
Source: NVIDIA H200 · SXM / NVL Specification
You can also calculate the power difference. Assuming that the additional power consumption actually averages 300 W and is maintained 24 hours a day, 365 days a year, this adds up to 0.3 kW × 24 × 365 = 2,628 kWh per year. This is not a measurement or electricity bill estimate. Rather than using the GPU’s nameplate power as average power consumption, you should calculate the total cost of ownership by factoring in actual utilization and cooling costs.
Before testing a candidate, write down the model or project you will use, its precision, and its context, deployment, and concurrency. Allowable latency, virtualization method, usage time and power limit, and warranty conditions are also adjusted. When these conditions are the same, it becomes easier to compare differences in actual tasks than GPU ratings.
Data sources and verification criteria
Official specifications and support documentation were available on October 18, 2026. Since the system configuration, SKU, and memory notation may vary depending on the document, the standards for each table are specified in this article. In particular, the HGX capacity in the Blackwell datasheet and the board/framebuffer notation in the AI Enterprise vGPU document are not the same, so it is not interpreted that the VRAM in Table 3 can be allocated to the VM as is. Support availability and pricing may change in the future.
- NVIDIA RTX PRO 6000 Blackwell Workstation Edition
- NVIDIA RTX PRO 6000 Blackwell Max-Q Datasheet
- NVIDIA RTX PRO 5000 Blackwell · 48GB / 72GB
- NVIDIA RTX PRO 4500 Blackwell Workstation Datasheet
- NVIDIA Ada Professional GPU Architecture White Paper
- NVIDIA RTX A6000 Datasheet
- NVIDIA RTX PRO 6000 Blackwell Server Edition
- NVIDIA RTX PRO 4500 Blackwell Server Edition
- Lenovo Press · NVIDIA L4 Product Guide
- NVIDIA L40 Datasheet
- NVIDIA L40S Official Specifications
- NVIDIA A2 Official Specifications
- NVIDIA A10 Official Specifications
- NVIDIA Ampere product comparison table · A16 included
- NVIDIA A40 Datasheet
- NVIDIA A30 Datasheet
- NVIDIA A100 PCIe / SXM Datasheet
- NVIDIA H100 · SXM / NVL Specification
- NVIDIA Hopper Architecture PCIe H100 HBM2e Description
- NVIDIA H200 · SXM / NVL Specification
- NVIDIA Blackwell Datasheet · HGX B200 Specifications Page 8
- NVIDIA Blackwell Ultra Datasheet · HGX B300 Specifications Page 5
- NVIDIA HGX · Rubin Discrete GPU Public Specification
- NVIDIA vGPU Software Supported GPU List · Updated 2026-10-02
- NVIDIA AI Enterprise 8 · Computation vGPU hardware-specific documentation
- NVIDIA AI Enterprise 8 · Blackwell vGPU Profile
- NVIDIA MIG supported GPU list · Updated 2026-10-13
- NVIDIA MIG setup prerequisites · Workstation display mode
- NVIDIA AI Enterprise 8 · MIG-backed vGPU limitations
Price sources are linked directly to each row in Table 5. For specifications, the manufacturer’s technical documentation took precedence over the vendor’s description. The judgment of use in this article is an editorial analysis based on public specifications, and is not a guarantee of performance in individual systems.
