GeekRadar / The comparison desk

COMPARE.

GPU

NVIDIA

NVIDIA B200

Graphics processor

Video memory: 180 GB^*
Memory bus: 8,192 bit^*
Rated memory bandwidth: 8,190.00 GB/s^*

Benchmark results

LLM Inference: llama.cpp Bench: llama 7B Q4_0 (Q4_0) · version unreported · score315.85^* tokens/sPlatform unreported Compare results

Recorded results

Only the same benchmark, version, workload, scenario and unit share a row. Missing results are not zero. System results include the full submitted hardware and software configuration.

Blender Cycles / classroomCUDA / Linux · 4.5.0 · samples/min

NVIDIA B200

1,833.80 samples/min

3 submissions · median result

Latest submission: Dec 28, 2025

Rendering backend: CUDA
Operating system: Linux
Statistic: Median
Evaluation type: community rendering submissions

Blender Cycles / junkshopCUDA / Linux · 4.5.0 · samples/min

NVIDIA B200

2,228.61 samples/min

3 submissions · median result

Latest submission: Dec 28, 2025

Rendering backend: CUDA
Operating system: Linux
Statistic: Median
Evaluation type: community rendering submissions

Blender Cycles / monsterCUDA / Linux · 4.5.0 · samples/min

NVIDIA B200

4,124.48 samples/min

3 submissions · median result

Latest submission: Dec 28, 2025

Rendering backend: CUDA
Operating system: Linux
Statistic: Median
Evaluation type: community rendering submissions

Processing & platform

Architecture: Blackwell
Process node: 5 nm^*
GPU cores: 18,944^*
Graphics interface: PCIe 6.0 x16^*
Reported chip die area: 1,628 mm²^*
Transistors: 208,000 million^*
GPU L2 cache: 126 MB^*
Graphics processor: GB100^*
Texture mapping units: 592^*
Raster output units: 24^*
Streaming multiprocessors: 148^*
Tensor cores: 592^*
Chip foundry: TSMC^*
Card or module format: SXM Module^*
Declared GPU TDP: 1,000 W^*
Suggested system PSU: 1,400 W^*
OpenCL: 3.0^*
CUDA compute capability: 10.0^*

Memory & storage

Memory options: 90^*
Video memory: 180 GB^*
Video memory type: HBM3e^*
Memory bus: 8,192 bit^*
Rated memory bandwidth: 8,190.00 GB/s^*

Power specifications

Thermal design power: 1,000 W^*

System configuration

Detailed specifications

Specifications 1 · 28 fields

GPU architecture — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

Blackwell^*

Calculated theoretical peak: BF16 tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

2,250^* TFLOPS

Scope and conditions

Calculation: 4.5 / 2^*

Theoretical peak: BF16 tensor with sparsity per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

4,500^* TFLOPS

Calculated theoretical peak: FP16 tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

2,250^* TFLOPS

Scope and conditions

Calculation: 4.5 / 2^*

Theoretical peak: FP16 tensor with sparsity per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

4,500^* TFLOPS

Theoretical peak: FP32 non-tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

75^* TFLOPS

Calculated theoretical peak: FP4 tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

9,000^* TFLOPS

Scope and conditions

Calculation: 18 / 2^*

Theoretical peak: FP4 tensor with sparsity per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

18,000^* TFLOPS

Theoretical peak: FP64 non-tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

37^* TFLOPS

Theoretical peak: FP64 tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

37^* TFLOPS

Calculated theoretical peak: FP6 tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

4,500^* TFLOPS

Scope and conditions

Calculation: 9 / 2^*

Theoretical peak: FP6 tensor with sparsity per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

9,000^* TFLOPS

Calculated theoretical peak: FP8 tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

4,500^* TFLOPS

Scope and conditions

Calculation: 9 / 2^*

Theoretical peak: FP8 tensor with sparsity per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

9,000^* TFLOPS

Calculated theoretical peak: INT8 tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

4,500^* TOPS

Scope and conditions

Calculation: 9 / 2^*

Theoretical peak: INT8 tensor with sparsity per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

9,000^* TOPS

Calculated theoretical peak: TF32 tensor dense per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

1,100^* TFLOPS

Scope and conditions

Calculation: 2.2 / 2^*

Theoretical peak: TF32 tensor with sparsity per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

2,200^* TFLOPS

Physical configuration — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

SXM (HGX 8-GPU baseboard)^*

GPU interconnect bandwidth (NVLink 5) bidirectional — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

1,800^* GB/s

GPU interconnect bandwidth per direction — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

900^* GB/s

PCIe interface declaration — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

PCIe Gen5: 128 GB/s^*

Maximum board power — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

1,000^* W

Peak memory bandwidth per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

7,700^* GB/s

Memory capacity per GPU — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

180^* GB

Memory reported in a documented driver sample — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

183,359^* MiB

Scope and conditions

Configuration: Documented driver sample; not nominal VRAM capacity.^*

Memory technology — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

HBM3E^*

GPU model — NVIDIA Blackwell GPU (HGX B200 / DGX B200) — 180 GB HBM3E

NVIDIA Blackwell GPU (HGX B200 / DGX B200)^*

Specifications 2 · 6 fields

Form factor declaration: HGX / SXM module^*
Graphics model: B200^*
Full graphics designation: NVIDIA B200^*
Manufacturer: NVIDIA^*
Memory technology: HBM3E^*
Thermal design power declaration: 1,000^* W

Specifications 3 · 31 fields

proc transistors

104,000 million^*

cores

18,944^*

tensor cores

592^*

sm count

148^*

base clock

700^*

boost clock

1,965^*

memory clock

2,000^*

ram

90^*

name

NVIDIA B200^*

test name Graphics benchmark result

LLM Inference: llama.cpp Bench: llama 7B Q4_0 (Q4_0)^*

Scope and conditions

benchmark_context

suite: Per-row third-party test name^*
test_configuration_verified: No^*
comparable: No^*

ram type

HBM3e^*

bus width

4,096^*

bus interface

PCIe 5.0 x16^*

ram bandwidth

4.10 TB/s^*

fp16 performance

1.1912^*

fp32 performance

74.45^*

fp64 performance

37.22^*

score Graphics benchmark result

315.85^*

Scope and conditions

benchmark_context

suite: Per-row third-party test name^*
test_configuration_verified: No^*
comparable: No^*

tdp

1,000^* W

suggested psu

1,400^* W

slot width

SXM Module^*

api cuda

10^*

is retail board

No^*

gpi value

65.81^*

unit Graphics benchmark result

tokens/s^*

Scope and conditions

benchmark_context

suite: Per-row third-party test name^*
test_configuration_verified: No^*
comparable: No^*

vendor

nvd^*

manufacturer

NVIDIA^*

arch name

Blackwell^*

card release date

2024-03-13^*

proc foundry

TSMC^*

proc process size

5^* nm

Specifications 4 · 29 fields

CUDA compute capability

10.0^*

OpenCL

3.0^*

Graphics interface

PCIe 6.0 x16^*

Memory bus

8,192^* bit

Reported shader processors

18,944^*

Reported theoretical FP16 peak

1.1094^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Reported theoretical FP32 peak

69.34^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Reported theoretical FP64 peak

34.67^* TFLOPS

Scope and conditions

Arithmetic density: Not stated^*

Declared L1 cache configuration

256 KB (per SM)^*

Declared GPU L2 cache

126 MB^*

Reported display outputs

No outputs^*

Theoretical pixel fill rate

43.92 GPixel/s^*

Reported chip die area

1628 mm²^*

Chip foundry

TSMC^*

Process node declaration

5 nm^*

Transistors

208,000 million^*

Graphics processor

GB100^*

Reported video memory

180^* GB

Rated memory bandwidth

8.19 TB/s^*

Video memory type

HBM3e^*

Raster output units

24^*

Reported shading units

18,944^*

Card or module format

SXM Module^*

Streaming multiprocessors

148^*

Suggested system PSU

1400 W^*

Declared GPU TDP

1000 W^*

Tensor cores

592^*

Theoretical texture fill rate

1,083.4 GTexel/s^*

Texture mapping units

592^*

Specifications 5 · 5 fields

GPU Bandwidth

up to 8^* TB/s

Scope and conditions

native_context

capacity unit base: not_established^*
configuration: HGX 8-GPU baseboard^*
memory role: production_HGX_configuration^*
resource scope: per_GPU^*

Memory per GPU

180^* GB · HBM3e^*

Scope and conditions

native_context

capacity unit base: not_established^*
configuration: HGX 8-GPU baseboard^*
memory role: production_HGX_configuration^*
resource scope: per_GPU^*

Nominal GPU Memory

192^* GB · HBM3e^*

Scope and conditions

native_context

Memory scope: Nominal per-GPU memory; not the total CPU + GPU module memory or runtime-available capacity^*
memory role: nominal_release_inventory^*
not dedicated VRAM for shared memory systems: Yes^*
not runtime available or HGX usable capacity: Yes^*

GPU Aggregate Bandwidth per 8-GPU Node

up to 64^* TB/s

Scope and conditions

native_context

capacity unit base: not_established^*
configuration: HGX 8-GPU baseboard^*
memory role: production_HGX_configuration^*
resource scope: eight_GPU_aggregate^*

Memory per 8-GPU Node

1.44^* TB · HBM3e^*

Scope and conditions

native_context

capacity unit base: not_established^*
configuration: HGX 8-GPU baseboard^*
memory role: production_HGX_configuration^*
resource scope: eight_GPU_aggregate^*

* Provisional data