The Future of Compute
Q2 2025
Raleigh
Bangalore
Ho Chi Minh City
Shanghai
Taipei
Santa Clara
Portland Warsaw
Pune
Ampere is a Global Company
9 Worldwide Locations and Global Design and Manufacturing Capability
2
THE TIME IS NOW
TO CHANGE THE TRAJECTORY OF AI COMPUTING
3
Ampere Processors
AI Compute is Everywhere
Ampere: AI Compute Processors
Scalability
Efficiency Larger Low Latency Private Caches
Single-Threaded Cloud Core
Consistent Operating Frequency
Maximum Core Counts
Power and Area-Efficient
Ampere Architecture
Advanced Architectural Features
Fine Grain Power Management
Performance
Right Sized AI Computing
5
Ampere® Processors for Edge AI and Telco
Consistent Throughput
Ampere Altra 128 Cores
Intel 64 Cores with Hyper-Threading
Throughput
(OPS/Sec)
Noisy Neighbor
Entry and Exit
Time
Linear Scaling
Am
pereAltra
128
Cores
Intel 64 Cores with
Hyper-Threading
Performance
% Utilization
0 50 100
p.99
Latency
Runs
Ampere Altra vs Intel latency
in milliseconds
Predictable Low Latency
Deliver high throughput, low latency, low jitter, deterministic at 90%+ load
Via many single-threaded, fixed frequency, efficient cores with large private caches
6
AmpereCloud Native Processors Families
7
Scaling from 32 to 192 cores, designed for AI Computing performance, efficiency and density
Ampere® Altra® Family AmpereOne®
32 to 128 Cores
1MB Private L2 Cache per Core
16 and 32MB System Level Cache
8 channel DDR4 – up to 4TB
128 lanes PCIe Gen4
40W to178W Usage Power
96 to 192 Cores
2MB Private L2 Cache per Core
64MB System Level Cache
8 channel DDR5 – up to 4TB
128 lanes PCIe Gen5
185W to 292W Usage Power
AmpereOne® M
96-192 Cores
2MB Private L2 Cache per Core
64MB System Level Cache
12 channel DDR5 – up to 3TB
96 lanes PCIe Gen5
239W to 348W Usage Power
General purpose, low power AI Compute
for Edge and Telco applications with power
constraints. Most efficient AI Compute for
embedded & Edge.
Efficient, flexible compute for Cloud Native
workloads and traditional AI Inference
applications such as DLRM, CV, NLP.
Advanced compute for enterprise AI,
optimized for LLMs and Agentic AI, with
cutting edge security and VM/Container
density for parallel AI execution.
8
AI Compute from Edge to Cloud
Web Service Stacks
Database Stacks
Telecom
Video & CDN Stacks
Artificial Intelligence
Networking
Cloud Native Processors Designed for AI Compute
AI Inference, Cloud Native Applications & Workloads Run Best on Ampere
AmpereOne®: The Next Generation of Sustainable Computing
See end notes on comparative w Genoa, Sierra Forest & Emerald Rapids.
729 673 579 458
283 380 330 350
2.57 1.77 1.75 1.31
0
125
250
375
500
625
750
AmpereOne A192-32X AMD EPYC 9654 Intel Xeon 6780E Intel Xeon 8592+
Performance Usage Power (W)
SPEC CPU®2017 Socket-level Performance & Usage Power
Perf/Wà
Up to 90% more Efficient than the latest X86 Processors
9
1.28 1.25
1.15
1.27
1.02
0.91
1.22
1.39
1.70
1.79 1.82
1.41 1.39
1.86
0.00
0.20
0.40
0.60
0.80
1.00
1.20
1.40
1.60
1.80
2.00
NGINX Redis Memcached MySQL Elasticsearch PostgreSQL Cassandra
Socket-level Performance & Efficiency
Am pereOne A192-32X Performanc e Am pereOne A192-32X Performanc e/Watt AM D EPY C 9654 (base)
AmpereOne®: 40-90% More Efficient on Real Cloud Native Apps
* See End Notes
10
AmpereOne®: Up to 2X More Efficient on AI Inference Apps
* See End Notes
ü Recommender Engines ü Vision Processing ü Language Processing
11
AmpereOne® Platform – Sustainable AI Compute Performance
AmpereOne delivers up to 2X better Perf/Rack than legacy X86 Processors
12
Ampere Confidential
Cloud Service Provider: Uber
Advantage
Background: Uber aims to achieve HW and capacity diversity to provide
flexibility for engineering to choose the optimal infrastructure for Uber
applications. Uber also strives to be a zero-emissions platform company.
Opportunity: Data center space and energy savings for OCI translates to better
price-performance and cost optimizations for Uber, while reducing overall CO2
footprint.
Solution: Uber used a four-phase approach to evaluate four different classes of
workloads and were able to troubleshoot key differences in x86 and Arm64 to
improve performance.
Result: Uber has successfully converted a large part of their compute in OCI
from E4 (AMD) to Ampere A1(Ampere® Altra®) and A2(AmpereOne®) shapes
and are currently qualifying the most critical workloads for large scale
deployment.
Uber
https://amperecomputing.com/blogs/how-uber-transitioned-part-1
https://www.oracle.com/customers/uber/
https://www.cio.com/article/3513933/uber-embraces-the-cloud-with-customized-cpus.html *All trademarks, logos and brand names are the property of their respective owners.
1
3
Ampere® Nested Virtualization
https://amperecomputing.com/blogs/unlocking-layers
• Nested Virtualization provides:
• Hardware-enforced isolation and performance gains
• Seamless On-Prem to Cloud migration
• Simplification of test/dev environments across industries
• Operating Principles
• Nested virtualization reduces performance bottlenecks caused
by VM entry/exit operations.
• Enablement:
• AmpereOne Platforms support NV2(Arm v8.4+), and is enabled by
default in Linux kernel
Nested Virtualization offers Security, Isolation and Performance gains
14
Technical enablement of Arm64 virtualization is coming along, and fast
15
The most popular open source virtualization management platforms are
commercially supported on Arm64, and support is improving all the time
16
App Layers: Start WithAn Inventory of Your Software Stacks
17
Compilers and Runtimes Support Arm64 ISA as a Tier 1 Platform
*Other product names used in this publication are for
identification purposes only and may be trademarks of their
respective companies.
1 Operating Systems
2 Compilers/Runtimes
Compiled Runtime and Interpreted
The Ecosystem is Ready
18
*Other product names used in this publication are for
identification purposes only and may be trademarks of their
respective companies.
Applications Database Infra Tools Networking
& Storage
Language &
Runtimes
Orchestration,
Virtualization &
Containers
Operating
Systems
Alma Linux
Ubuntu
For a non-exhaustive list of Ampere ready software, visit
https://amperecomputing.com/developers/ampere-ready-software
Ampere Offers Energy-Efficient Future For AI And Cloud

Ampere Offers Energy-Efficient Future For AI And Cloud

  • 1.
    The Future ofCompute Q2 2025
  • 2.
    Raleigh Bangalore Ho Chi MinhCity Shanghai Taipei Santa Clara Portland Warsaw Pune Ampere is a Global Company 9 Worldwide Locations and Global Design and Manufacturing Capability 2
  • 3.
    THE TIME ISNOW TO CHANGE THE TRAJECTORY OF AI COMPUTING 3
  • 4.
  • 5.
    Ampere: AI ComputeProcessors Scalability Efficiency Larger Low Latency Private Caches Single-Threaded Cloud Core Consistent Operating Frequency Maximum Core Counts Power and Area-Efficient Ampere Architecture Advanced Architectural Features Fine Grain Power Management Performance Right Sized AI Computing 5
  • 6.
    Ampere® Processors forEdge AI and Telco Consistent Throughput Ampere Altra 128 Cores Intel 64 Cores with Hyper-Threading Throughput (OPS/Sec) Noisy Neighbor Entry and Exit Time Linear Scaling Am pereAltra 128 Cores Intel 64 Cores with Hyper-Threading Performance % Utilization 0 50 100 p.99 Latency Runs Ampere Altra vs Intel latency in milliseconds Predictable Low Latency Deliver high throughput, low latency, low jitter, deterministic at 90%+ load Via many single-threaded, fixed frequency, efficient cores with large private caches 6
  • 7.
    AmpereCloud Native ProcessorsFamilies 7 Scaling from 32 to 192 cores, designed for AI Computing performance, efficiency and density Ampere® Altra® Family AmpereOne® 32 to 128 Cores 1MB Private L2 Cache per Core 16 and 32MB System Level Cache 8 channel DDR4 – up to 4TB 128 lanes PCIe Gen4 40W to178W Usage Power 96 to 192 Cores 2MB Private L2 Cache per Core 64MB System Level Cache 8 channel DDR5 – up to 4TB 128 lanes PCIe Gen5 185W to 292W Usage Power AmpereOne® M 96-192 Cores 2MB Private L2 Cache per Core 64MB System Level Cache 12 channel DDR5 – up to 3TB 96 lanes PCIe Gen5 239W to 348W Usage Power General purpose, low power AI Compute for Edge and Telco applications with power constraints. Most efficient AI Compute for embedded & Edge. Efficient, flexible compute for Cloud Native workloads and traditional AI Inference applications such as DLRM, CV, NLP. Advanced compute for enterprise AI, optimized for LLMs and Agentic AI, with cutting edge security and VM/Container density for parallel AI execution.
  • 8.
    8 AI Compute fromEdge to Cloud Web Service Stacks Database Stacks Telecom Video & CDN Stacks Artificial Intelligence Networking Cloud Native Processors Designed for AI Compute AI Inference, Cloud Native Applications & Workloads Run Best on Ampere
  • 9.
    AmpereOne®: The NextGeneration of Sustainable Computing See end notes on comparative w Genoa, Sierra Forest & Emerald Rapids. 729 673 579 458 283 380 330 350 2.57 1.77 1.75 1.31 0 125 250 375 500 625 750 AmpereOne A192-32X AMD EPYC 9654 Intel Xeon 6780E Intel Xeon 8592+ Performance Usage Power (W) SPEC CPU®2017 Socket-level Performance & Usage Power Perf/Wà Up to 90% more Efficient than the latest X86 Processors 9
  • 10.
    1.28 1.25 1.15 1.27 1.02 0.91 1.22 1.39 1.70 1.79 1.82 1.411.39 1.86 0.00 0.20 0.40 0.60 0.80 1.00 1.20 1.40 1.60 1.80 2.00 NGINX Redis Memcached MySQL Elasticsearch PostgreSQL Cassandra Socket-level Performance & Efficiency Am pereOne A192-32X Performanc e Am pereOne A192-32X Performanc e/Watt AM D EPY C 9654 (base) AmpereOne®: 40-90% More Efficient on Real Cloud Native Apps * See End Notes 10
  • 11.
    AmpereOne®: Up to2X More Efficient on AI Inference Apps * See End Notes ü Recommender Engines ü Vision Processing ü Language Processing 11
  • 12.
    AmpereOne® Platform –Sustainable AI Compute Performance AmpereOne delivers up to 2X better Perf/Rack than legacy X86 Processors 12
  • 13.
    Ampere Confidential Cloud ServiceProvider: Uber Advantage Background: Uber aims to achieve HW and capacity diversity to provide flexibility for engineering to choose the optimal infrastructure for Uber applications. Uber also strives to be a zero-emissions platform company. Opportunity: Data center space and energy savings for OCI translates to better price-performance and cost optimizations for Uber, while reducing overall CO2 footprint. Solution: Uber used a four-phase approach to evaluate four different classes of workloads and were able to troubleshoot key differences in x86 and Arm64 to improve performance. Result: Uber has successfully converted a large part of their compute in OCI from E4 (AMD) to Ampere A1(Ampere® Altra®) and A2(AmpereOne®) shapes and are currently qualifying the most critical workloads for large scale deployment. Uber https://amperecomputing.com/blogs/how-uber-transitioned-part-1 https://www.oracle.com/customers/uber/ https://www.cio.com/article/3513933/uber-embraces-the-cloud-with-customized-cpus.html *All trademarks, logos and brand names are the property of their respective owners. 1 3
  • 14.
    Ampere® Nested Virtualization https://amperecomputing.com/blogs/unlocking-layers •Nested Virtualization provides: • Hardware-enforced isolation and performance gains • Seamless On-Prem to Cloud migration • Simplification of test/dev environments across industries • Operating Principles • Nested virtualization reduces performance bottlenecks caused by VM entry/exit operations. • Enablement: • AmpereOne Platforms support NV2(Arm v8.4+), and is enabled by default in Linux kernel Nested Virtualization offers Security, Isolation and Performance gains 14
  • 15.
    Technical enablement ofArm64 virtualization is coming along, and fast 15
  • 16.
    The most popularopen source virtualization management platforms are commercially supported on Arm64, and support is improving all the time 16
  • 17.
    App Layers: StartWithAn Inventory of Your Software Stacks 17 Compilers and Runtimes Support Arm64 ISA as a Tier 1 Platform *Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies. 1 Operating Systems 2 Compilers/Runtimes Compiled Runtime and Interpreted
  • 18.
    The Ecosystem isReady 18 *Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies. Applications Database Infra Tools Networking & Storage Language & Runtimes Orchestration, Virtualization & Containers Operating Systems Alma Linux Ubuntu For a non-exhaustive list of Ampere ready software, visit https://amperecomputing.com/developers/ampere-ready-software