GPU Nodes

Introduction Euler contains several different GPU models. Since GPU nodes are purchased to meet shareholders' explicit needs, the specifications (CPU, RAM, disk, networks) vary considerably, even for nodes with the same GPU model.
GPU nodes are only accessible to shareholders who invested in this type of resource.

Operational:

Nodes	GPUs/Node	GPU	GPU-RAM	processing units
4	8	Nvidia Tesla V100	16 GB	5,120
93	8	Nvidia RTX 2080 Ti	11 GB	4,352
14	8	Nvidia TITAN RTX	24 GB	4,608
16	8	Nvidia Quadro RTX 6000	24 GB	4,608
33	8	Nvidia RTX 3090	24 GB	10,496
3	8	Nvidia Tesla A100	40 GB	8,192
3	10		80 GB
2	8
80	8	Nvidia RTX 4090	24 GB	16,384

Nodes	APUs/Node	CPU-Cores/APU	APU	RAM	processing units
6	4	24	AMD MI300A	128 GB	14,592

APUs combine CPU cores and GPU cores on one die, sharing common HBM3 memory.
(List is based on 'sinfo -o "%10D %80G %80N')

Decommissioned:

Nodes	GPUs/Node	GPU	GPU-RAM	Decommissioned in
10	8	Nvidia GTX 1080	8 GB	2023
50	8	Nvidia GTX 1080 Ti	11 GB	2024

Specification overview

The capabilities of different GPU models differ quite substantially. Therefore we provide this overview. For more details, check the sources of these numbers by clicking the links in the table's header.
Question marks (?) mean that we don't know this value. Dashes (-) mean that this operation is not provided natively.

Theoretical peak compute

Precision	AMD MI300A			nVidia A100
Raw	Tensor Core	Sparse	Raw	Tensor Core
FP64 https://en.wikipedia.org/wiki/Floating_point_operations_per_second TFLOP/s	61.3	122.6	-	9.7
FP32 https://en.wikipedia.org/wiki/Floating_point_operations_per_second TFLOP/s	122.6	122.6	-	19.5
FP16 https://en.wikipedia.org/wiki/Floating_point_operations_per_second TFLOP/s	-	980.6	1'961.2	78.0
FP8 https://en.wikipedia.org/wiki/Floating_point_operations_per_second TFLOP/s	-	1'961.2	3'922.3	-
TF32 https://en.wikipedia.org/wiki/Floating_point_operations_per_second TFLOP/s	-	490.3	980.6	-
BF16 https://en.wikipedia.org/wiki/Floating_point_operations_per_second TFLOP/s	-	980.6	1'961.2	-
INT32 [TOP/s]	19.5	-	-	19.5
INT8 [TOP/s]	?	1'961.2	3'922.3	?
INT4 [TOP/s]	?	-	-	?
BOOL [TOP/s]	?	-	-	?

Theoretical peak compute

Theoretical peak bandwidth

	AMD MI300A	nVidia A100	nVidia RTX 4090	nVidia RTX 3090
RAM to GPU [GB/s]	5’300	1’555	1’008	936
GPU to GPU [GB/s]	256	6.25	?	?

Theoretical peak bandwidth

Benchmarks

To evaluate the performance of our GPUs, we ran several benchmarks. Here's what we measured.

Application	Hardware	Units	RTX 4090	A100	MI300A
Stream	1 GPU	[GB/s]	| 927	| 1'371	| 2'923
HPL	1 GPU	https://en.wikipedia.org/wiki/Floating_point_operations_per_second GFLOP/s	| 1'198	| 14'270	| 42'800
	4 GPUs	https://en.wikipedia.org/wiki/Floating_point_operations_per_second GFLOP/s	| 4'741	| 36'850	| ?
HPCG	1 GPU	https://en.wikipedia.org/wiki/Floating_point_operations_per_second GFLOP/s	| 188	| 249	| 548
	4 GPUs	https://en.wikipedia.org/wiki/Floating_point_operations_per_second GFLOP/s	| 688	| 891	| 1'915
Stream	1 GPU	[GB/s]	| 927	| 1'371	| 2'923
Hashcat MD5	4 GPUs	[GH/s]	| 259	| 184	| 472

STREAM is the de facto industry standard benchmark for measuring sustained memory bandwidth.
HPL solves a (random) dense linear system in double precision (64 bits) arithmetic. It is compute bound.
HPCG uses the conjugate gradient method to solve a large sparce linear system benchmark It is I/O bound.
Hashcat is a password recovery tool and MD5 is a widely used hash function producing a 128-bt hash value.

AlexNet (FP16) is a neural network for image classification and won ILSVRC in 2012.

ResNet50 (FP16) is a neural network for image classification and won ILSVRC in 2015.

DenseNet-121 (FP16) is a deep learning architecture designed for image classification and other tasks like segmentation.