This post compares the Total Cost of Ownership (TCO) for Lambda servers and clusters VS cloud instances with NVIDIA A100 GPUs. We first calculate the TCO for individual Lambda HyperplaneA100 and Scalar servers and then compare the cost of renting a similarly equipped AWS EC2 p4d.24xlarge instance. We then walk through the cost of building and operating server clusters using NVIDIA A100 GPUs.
Lambda A100 servers v.s. AWS p4d.24xlarge
Two models of Lambda servers are available with NVIDIA Tesla A100 GPUs: HyperplaneA100 and ScalarA100. The main difference is how A100 GPUs are interconnected in each server. The A100 GPUs in the HyperplaneA100 are connected via NVLink + NVSwitch. In the Scalar server, the GPUs are connected with PCIe. For this reason, the HyperplaneA100 server scales better when multiGPU training becomes communicationbottlenecked. More importantly, HyperplaneA100 servers support InfiniBand (8x IB ports per server) and GPUDirect Remote Direct Memory Access (RDMA), enabling them to scale efficiently for multinode distributed training. In addition, the A100PCIe GPU model used in the Scalar server consumes less power than the A100SXM4 GPU used in the Hyperplane servers (250W v.s. 400W), so they may experience a 10% throughput drop while performing machine learning tasks.
On the cloud side, the AWS EC2 p4d.24xlarge instance is similar to the eight GPU Lambda Hyperplane 8A100. The main difference is the choice of CPU – the HyperplaneA100 uses the latest AMD EPYC CPUs over the Intel Xeon CPUs due to their higher number of cores and PCIe 4.0 support.
Test Configurations
HyperplaneA100  ScalarA100  p4d.24xlarge  

GPU  8x A100SXM440GB  8x A100PCIe40GB  8x A100SXM440GB 
CPU  2x AMD EPYC 7763 (64Core)  2x AMD EPYC 7763 (64Core)  Intel(R) Xeon(R) Platinum 8275CL (24Core) 
System Memory  1.0 TB ECC  1.0 TB ECC  1.1 TB ECC 
Storage  15.36 TB NVMe  15.36 TB NVMe  8 TB NVMe 
Takeaway
We will give a walkthrough of the performance and cost analysis in later sections, but first, we need to discuss the key metric for measuring costeffectiveness: flops/$
or flops per dollar. Flops/$
is the amount of computation you can buy with one dollar. It is calculated by dividing the total petaflops generated over a time period by the total cost of purchasing and operating the system over that same time period.
Higher flops/$
is more costeffective. As the following two tables show, onprem servers give significantly better flops/$
compared to cloud instances, as long as one does not scrap the value of the servers to zero after a year (100%
annual depreciation).
Note that these two tables assume 50%
occupancy for the onprem and reserved cloud instances. We will have a breakeven occupancy
analysis later in this blog.
Flops/$ (1yr analysis with 50% occupancy)
HyperplaneA100 (25% annual depreciation)  HyperplaneA100 (100% annual depreciation)  ScalarA100 (25% annual depreciation)  ScalarA100 (100% annual depreciation)  p4d (qll upfront)  p4d (partial upfront)  p4d (no upfront)  p4d (on demand)  

Total Cost 1yr  $55,534  $160,534  $47,279  $138,779  $164,955  $168,321  $176,737  $143,544 
Total Petaflops 1yr  2,459,808  2,459,808  2,459,808  2,459,808  2,459,808  2,459,808  2,459,808  2,459,808 
Petaflops/$  44.3  15.3  52.0  17.7  14.9  14.6  13.9  17.1 
Flops/$ (3yr analysis with 50% occupancy)
HyperplaneA100 (25% annual depreciation)  HyperplaneA100 (100% annual depreciation)  ScalarA100 (25% annual depreciation)  ScalarA100 (100% annual depreciation)  p4d (all upfront)  p4d (partial upfront)  p4d (no upfront)  p4d (on demand)  

Total Cost 3yr  $166,602  $201,602  $141,837  $172,337  $285,893  $304,135  $328,473  $430,629 
Total Petaflops 3yr  7,379,424  7,379,424  7,379,424  7,379,424  7,379,424  7,379,424  7,379,424  7,379,424 
Petaflops/$  44.3  36.6  52.0  42.8  25.8  24.3  22.5  17.1 
Breakeven Occupancy Analysis
Since an ondemand instance is billed hourly, an onprem server solution only becomes cheaper if the time occupancy reaches a certain amount. For example, increasing occupancy from 50%
to 100%
will increase 1yr petaflops/$
from 44.3
to 74.5
for a Lambda HyperplaneA100.
We use the formula below to compute the breakeven occupancy x
for Lambda severs to become cheaper than the ondemand instance over a y
number of years:
Annual Colocation Cost
* x
* y
+ Admin Cost
* y
+ Upfront
 Salvage Value
= Annual Ondemand Cost
* x
* y
The exact costs in the above equation can be found in the later TCO analysis. The following table shows the breakeven occupancy for different servers and annual depreciation models.
Breakeven Occupancy
HyperplaneA100 (25% annual depreciation)  HyperplaneA100 (100% annual depreciation)  ScalarA100 (25% annual depreciation)  ScalarA100 (100% annual depreciation)  

1yr breakeven ratio  17%  56%  15%  48% 
3yr breakeven ratio  17%  21%  15%  19% 
For 1yr TCO and assuming 25%
annual depreciation, a HyperplaneA100 server is more costeffective as long as it is occupied for 17%
for the time. Even if the value of the server goes to zero after one year, a HyperplaneA100 server is more costeffective as long as it is utilized 56%
of the time. With the ScalarA100 server, the numbers further drop down to 15%
and 48%
. The breakeven occupancies are the same for 1yr and 3yr TCO when 25%
annual depreciation is used since the salvage value of the server linearly decreases with the number of years. With 100%
annual depreciation, the breakeven occupancies for 3yr TCO are significantly lower than 1yr TCO, since most of the cost is booked the first year.
TCO Analysis
We use the following formula to compute the y
yr TCO for both Lambda A100 servers:
TCO(y
) = (Annual_CoLocation_Cost
+ Annual_Admin_Cost
) * y
+ min(Upfront
, Server_Value_Depreciation_per_year
* y
)
Notice the TCO for AWS instances has zero annual colocation cost, administration cost and value depreciation. We use the reference rates for a p4d.24xlarge instance to compute the upfront cost and annual rental cost for different AWS plans.
1yr TCO with 50% occupancy
HyperplaneA100 (25% annual depreciation)  HyperplaneA100 (100% annual depreciation)  ScalarA100 (25% annual depreciation)  ScalarA100 (100% annual depreciation)  p4d (all upfront)  p4d (partial upfront)  p4d (no upfront)  p4d (on demand)  

Upfront  $140,000  $140,000  $122,000  $122,000  $164,955  $84,161  $0  $0 
Annual rental  $0  $0  $0  $0  $0  $84,160  $176,737  $143,544 
Annual CoLocation Cost  $10,534  $10,534  $6,779  $6,779  $0  $0  $0  $0 
Annual Admin Cost  $10,000  $10,000  $10,000  $10,000  $0  $0  $0  $0 
Server Value Depreciation per year  $35,000  $140,000  $30,500  $122,000  $0  $0  $0  $0 
Total Cost Over 1yr  $55,534  $160,534  $47,279  $138,779  $164,955  $168,321  $176,737  $143,544 
3yr TCO with 50% occupancy
HyperplaneA100 (25% annual depreciation)  HyperplaneA100 (100% annual depreciation)  ScalarA100 (25% annual depreciation)  ScalarA100 (100% annual depreciation)  p4d (all upfront)  p4d (partial upfront)  p4d (no upfront)  p4d (on demand)  

Upfront  $140,000  $140,000  $122,000  $122,000  $285,893  $152,071  $0  $0 
Annual rental  $0  $0  $0  $0  $0  $50,688  $109,491  $143,544 
Annual CoLocation Cost  $10,534  $10,534  $6,779  $6,779  $0  $0  $0  $0 
Annual Admin Cost  $10,000  $10,000  $10,000  $10,000  $0  $0  $0  $0 
Server Value Depreciation per year  $35,000  $140,000  $30,500  $122,000  $0  $0  $0  $0 
Total Cost Over 3yr  $166,602  $201,602  $141,837  $172,337  $285,893  $304,135  $328,473  $430,629 
Assuming 50%
occupancy and 25%
annual value depreciation, owning a onprem A100 server is much cheaper than renting it on AWS: Lambda HyperplaneA100 saves 41.7%
($285,893  $166,602 = $119,291
) over a 3yr TCO in respect to the cheapest AWS p4d.24xlarge option. Lambda ScalarA100 enlarges the savings to 50.4%
($285,893  $141,837 = $144,056
). The gap is even larger for a 1year TCO due to the significant amount of premium AWS charges for its 1year reserved instances. Lambda HyperplaneA100 saves 66.3%
($164,955  $55,534 = $109,421
) over a 1yr TCO in respect to the cheapest reserved AWS p4d.24xlarge options, and Lambda ScalarA100 enlarges the savings to 71.3%
($164,955  $47,279 = $117,676
).
Interestingly, the ondemand AWS instance is more costeffective than reserved instances at the 50%
occupancy rate. However, onprem Lambda servers are still much cheaper (see the breakeven occupancy analysis in the previous section).
The annual colocation cost is calculated without InfiniBand networking, storage server, management server, and rack. It assumes 50%
for Power Utilization Ratio, and Colocation Cost / kWh / Month
is estimated to be $350
(supported by our study of colocation cost). The main difference in colocation cost between HyperplaneA100 and ScalarA100 is the A100PCIe GPUs consuming less power than the A100SXM4 GPUs. We budget the annual administration cost for a single server to be $10,000
.
Lambda HyperplaneA100 cluster
So far, we have discussed the TCO for a single A100 server. This assumed there is an existing cluster/cloud infrastructure to host that server. This section will examine the cost of building an onprem A100 cluster from scratch, as well as its operation cost. The key variables in designing an onprem GPU cluster include computing nodes, networking, storage, management node, and the design of the rack itself. Our engineers implement best practices based on Lambda's experience in building machine learning infrastructures for research institutes in both industry and academia.
The following table shows the amortized cost/year/node
of InfiniBand networked HyperplaneA100 clusters, in particular:
 InfiniBand networking
 Racks & PDUs
 System administration cost of $10,000 / year / four servers
HyperplaneA100 Cluster with InfiniBand Networking
1x Cluster 
8x Cluster 
16x Cluster 


Upfront  $233,422  $1,610,948  $3,137,649 
Annual Total Operating Costs  $23,894  $111,804  $209,723 
Annual System Administration Cost  $10,000  $20,000  $40,000 
Annual Colocation Cost  $13,893  $101,804  $199,722 
Total Cost at Year 1  $257,316  $1,722,752  $3,347,372 
Total Cost at Year 3  $305,104  $1,946,360  $3,766,818 
Total Cost at Year 5  $352,892  $2,169,968  $4,186,264 
Number of nodes  1  8  16 
Amortized Cost / Year / Node (5 Years of Use)  $70,578  $54,249  $52,328 
Similarly, the following table shows the TCO of a ScalarA100 clusters (without InfiniBand networking and assuming distributed training is not the use case)
ScalarA100 Cluster
1x Cluster 
8x Cluster 
16x Cluster 


Upfront  $161,757  $1,028,964  $2,025,382 
Annual Total Operating Costs  $16,779  $59,084  $107,432 
Annual System Administration Cost  $10,000  $20,000  $40,000 
Annual Colocation Cost  $6,779  $49,084  $97,432 
Total Cost at Year 1  $178,536  $1,088,048  $2,132,814 
Total Cost at Year 3  $212,094  $1,206,216  $2,347,678 
Total Cost at Year 5  $245,652  $1,324,384  $2,562,542 
Number of nodes  1  8  16 
Amortized Cost / Year / Node (5 Years of Use)  $49,130  $33,109  $32,031 
The above tables compare the HyperplaneA100 TCO and the ScalarA100 TCO. Although ScalarA100 clusters come at a lower upfront and operation cost, which type of A100 server should be used depends on the use cases. We recommend HyperplaneA100 for clusters that run distributed training across multiple nodes. This is due to the superior internode communication of the 8 InfiniBand ports in the Hyperplane server. On the other hand, if the main use case of the cluster is inference or singlenode training, the ScalarA100 server can be a more costeffective choice.
Performance
Last but not least, we benchmark these servers using various deep learning models and compare their training throughput in the figures below. All benchmarks used the NGC PyTorch container and can be reproduced using this repo.
As expected, the Lambda HyperplaneA100 server and AWS p4d.24xlarge instance deliver similar performance (the averaged difference is less than 1%) since they use similar hardware configurations. ScalarA100 is about 7.2% and 8.2% slower in TF32 and AMP (Automatic Mixed Precision), respectively, due to its slower GPU interconnection and less power consumption.
Footnotes
1. We think 50% for Power Utilization Ratio is a fair estimation considering most of the time the system didn't draw full power. This is due to factors such as job scheduling, I/O or devicetodevice communication bottlenecks, and suboptimized code.
2. $10,000 per server is generious considering the average salary for a full time data center system administrator is $64,892
.
3. We assume a simple linear model and depreciate the value of the server by 25%
of its purchase price per year.
4. The cheapest 1yr reserved plan (all upfront) on AWS is more expensive than the TCO of a Lambda HyperplaneA100 with 100%
annual depreciation rate. Which means you can use less money to buy and operate a Lambda HyperplaneA100 server for a year and keep the server for free afterwards.
5. Increasing occupancy rate from 50%
to 100%
will double the total petaflops over a year, but only increase the 1yr TCO by $10,534
(doubling the electric bill). In consequence the petaflops/$
will increase to 74.5