Intel Gaudi 3 vs NVIDIA H200 SXM
Intel Gaudi 3
NVIDIA H200 SXM
|
16384 Shaders
128GB HBM2e
438MHz
|
16896 Shaders
141GB HBM3e
1980MHz
|
|
Peak AI Performance
1.84 PFLOPS
FP8 Tensor (FP32 Accumulate)
|
Peak AI Performance
3.96 POPS
INT8 Tensor Sparse
|
|
FP32
14.35 TFLOPS
|
FP32
66.91 TFLOPS
|
|
FP16
28.7 TFLOPS
|
FP16
133.82 TFLOPS
|
|
Form Factor
OAM Module
-
|
Form Factor
SXM5
-
|
|
TDP
900W
|
TDP
700W
|
|
-
-
-
-
-
|
-
-
-
-
-
|
|
GB6 OpenCL N/A
0%
|
GB6 OpenCL N/A
0%
|
|
GB6 Metal N/A
0%
|
GB6 Metal N/A
0%
|
|
GB6 Vulkan N/A
0%
|
GB6 Vulkan N/A
0%
|
|
GB5 OpenCL N/A
0%
|
GB5 OpenCL N/A
0%
|
|
GB5 CUDA N/A
0%
|
GB5 CUDA N/A
0%
|
|
GB5 Metal N/A
0%
|
GB5 Metal N/A
0%
|
|
GB5 Vulkan N/A
0%
|
GB5 Vulkan N/A
0%
|
|
OCT 2020.1 N/A
0%
|
OCT 2020.1 N/A
0%
|
|
OCT Metal N/A
0%
|
OCT Metal N/A
0%
|
|
Peak AI
Performance
1.84 PFLOPS
FP8 Tensor (FP32 Accumulate)
|
Peak AI
Performance
3.96 POPS
INT8 Tensor Sparse
|
|
-
-
-
|
-
-
-
|
|
FP8
-
-
-
1.84 PFLOPS
Tensor (FP32 Accumulate)
-
|
FP8
-
1.98 PFLOPS
Tensor (FP16 Accumulate)
3.96 PFLOPS
Tensor (FP16 Accumulate) Sparse
1.98 PFLOPS
Tensor (FP32 Accumulate)
3.96 PFLOPS
Tensor (FP32 Accumulate) Sparse
|
|
FP16
28.7 TFLOPS
-
-
458.75 TFLOPS
Tensor (FP32 Accumulate)
-
|
FP16
133.82 TFLOPS
989.43 TFLOPS
Tensor (FP16 Accumulate)
1.98 PFLOPS
Tensor (FP16 Accumulate) Sparse
989.43 TFLOPS
Tensor (FP32 Accumulate)
1.98 PFLOPS
Tensor (FP32 Accumulate) Sparse
|
|
FP32
14.35 TFLOPS
229.38 TFLOPS
Tensor
-
|
FP32
66.91 TFLOPS
-
-
|
|
-
-
-
|
FP64
33.45 TFLOPS
61.84 TFLOPS
Tensor
|
|
BF16
28.7 TFLOPS
1.84 PFLOPS
Tensor
-
|
BF16
133.82 TFLOPS
989.43 TFLOPS
Tensor
1.98 PFLOPS
Tensor Sparse
|
|
TF32
458.75 TFLOPS
Tensor
-
|
TF32
494.72 TFLOPS
Tensor
989.43 TFLOPS
Tensor Sparse
|
|
-
-
-
|
-
-
-
|
|
-
-
-
-
|
INT8
-
1.98 POPS
Tensor
3.96 POPS
Tensor Sparse
|
|
-
-
|
INT32
33.45 TOPS
|
|
-
-
|
-
-
|
|
Pixel
Fillrate
-
|
Pixel
Fillrate
47.52 GPixel/s
|
|
-
-
|
-
-
|
|
Texture
Fillrate
-
|
Texture
Fillrate
1045.44 GTexel/s
|
|
Foundry
TSMC
-
|
Foundry
TSMC
-
|
|
Fabrication Node
N5
-
|
Fabrication Node
4N
-
|
|
-
-
-
|
Die Size
814 mm²
-
|
|
-
-
|
Transistor Count
80 Billion
-
|
|
-
-
-
|
Transistor Density
98.28M/mm²
-
|
|
Form
OAM Module
|
Form
SXM5
|
|
Shading Units
16384 Shaders
-
|
Shading Units
16896 Shaders
-
|
|
Texture Mapping Units
-
|
Texture Mapping Units
528 TMUs
|
|
Render Output Units
-
|
Render Output Units
24 ROPs
|
|
Tensor Cores
64 T-Cores
|
Tensor Cores
528 T-Cores
|
|
-
-
|
-
-
|
|
-
-
|
Streaming Multiprocessors
132 SMs
|
|
Compute Units
8 CUs
|
-
-
|
|
-
-
|
-
-
|
|
-
-
|
-
-
|
|
-
-
-
438MHz
|
1830MHz Tensor
-
1065MHz Base
1980MHz
|
|
-
-
|
-
-
|
|
L1
-
-
-
Unknown
|
L1
64KB/SM Tex
256KB/SM
-
-
|
|
L2
Unknown
|
L2
50MB Shared
|
|
-
-
-
|
-
-
-
|
|
128GB
HBM2e
ECC
|
141GB
HBM3e
-
|
|
Bus Width
8192Bit
|
Bus Width
6144Bit
|
|
Clock
1808MHz
Transfer Rate
3.6GT/s
Bandwidth
3702.8GB/s
|
Clock
1566MHz
Transfer Rate
6.3GT/s
Bandwidth
4810.8GB/s
|
|
-
-
-
-
eSRAM
96MB
12800GB/s
-
-
|
-
-
-
-
-
-
-
-
-
|
|
TDP
900W
|
TDP
700W
|
|
-
-
|
-
-
|
|
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
No Ports
|
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
No Ports
|
|
Max Resolution
Unknown
|
Max Resolution
Unknown
|
|
Max Resolution Refresh Rate
-
|
Max Resolution Refresh Rate
-
|
|
Variable Refresh Rate
-
-
-
|
Variable Refresh Rate
G-Sync
FreeSync
-
|
|
Display Stream Compression (DSC)
Not Supported
|
Display Stream Compression (DSC)
Not Supported
|
|
Multi Monitor Support
Unknown
|
Multi Monitor Support
Unknown
|
|
-
-
|
-
-
|
|
No Encoders
-
|
No Encoders
-
|
|
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
|
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
|
|
No Decoders
|
Model
7x NVDEC 5
|
|
-
-
-
-
-
-
-
-
-
-
-
-
-
-
|
Codec
MPEG-1
MPEG-2
MPEG-4
-
VC-1
VP8
VP9
-
AVC (H.264)
HEVC (H.265)
-
AV1
-
-
|
|
-
-
-
-
|
-
-
-
-
|
|
-
-
OpenCL
3.0
-
-
|
-
-
OpenCL
3.0
-
-
|
|
-
-
-
-
-
-
-
-
-
-
|
-
-
CUDA
9.0
-
-
PureVideo HD
VP11
VDPAU
Feature Set K
|
|
Not a Card
-
-
-
|
Not a Card
-
-
-
|
|
-
-
-
-
-
-
-
-
|
-
-
-
-
-
-
-
-
|
|
-
-
PCIe Version
5.0
PCIe Lanes
16
|
-
-
PCIe Version
5.0
PCIe Lanes
16
|
|
Multi GPU Support
Supported
Type
RoCE
|
Multi GPU Support
Supported
Type
NVLink
|
|
-
-
-
-
-
-
|
-
-
-
-
-
-
|
Copy Link