Meta has updated Andromeda, the system that selects ads for further evaluation on Facebook and Instagram. Its complete pipeline now processes more than one billion ads per second on a single GPU server. The company puts this at roughly 100 times the throughput of its 2024 system. The improvement comes from a new neural network architecture, fused GPU operations and less data moving between servers.

For advertisers, these changes mean the system can make better matches between people and ads across a larger candidate pool while keeping the same latency budget. Meta attributes a 2.4% increase in Facebook conversions to Andromeda’s development since 2025. That is the company’s reported overall system impact, rather than a promise for every campaign or an effect attributable solely to the 2026 update.

Why Ad Selection Starts with Candidate Retrieval

Ad recommendation involves several stages. First, retrieval reduces a vast pool of eligible ads to a small shortlist within milliseconds. More computationally expensive ranking models then evaluate the surviving candidates in detail.

The quality of the first stage determines which options reach the next one. If a potentially useful ad is discarded too early, later ranking cannot bring it back. Yet applying an equally complex model to the entire initial pool would be too expensive under strict response-time and computing constraints.

That trade-off becomes harder as ad volumes grow. Advantage+ automation and generative creative tools add more candidates to consider. Andromeda therefore needs to broaden its initial evaluation while maintaining matching quality and a manageable processing cost.

How Multiple Experts Capture User Interests

The updated model uses Retrieval Mixture-of-Experts, or MoE. The general principle behind a mixture-of-experts architecture is to activate only a subset of specialized subnetworks for each example. This allows model capacity to grow without a proportional increase in computation for every request.

Andromeda replaces a single user representation tower with an ensemble of experts. They capture changing interests and context through sequences of ad interactions, user-cohort representations, semantic features derived from large language models and cross-surface co-engagement signals.

The ad representation also incorporates the current serving context. An attention-based routing network selects the most relevant user experts to evaluate a particular candidate. Rather than handling every user–ad pair in the same way, the system draws on model specialization.

Meta describes the resulting matching quality as approaching that of ranking models. Meanwhile, inference requires approximately 300 times fewer floating-point operations, or FLOPs, than the 2024 Andromeda model. This figure should not be read as an equivalent reduction in server costs or an equivalent speedup of the entire advertising system.

Why Move the Entire Retrieval Flow onto a GPU

At billions of candidate evaluations per second, moving data becomes a constraint. Even a small intermediate array for each ad can generate terabytes of memory traffic per second. Fast arithmetic is of limited use if the processor keeps waiting for data.

The NVIDIA GH200 Grace Hopper platform matters here. It connects a Grace CPU and a Hopper GPU through NVLink-C2C, providing a coherent interface with 900 GB/s of bandwidth. Realizing that hardware potential also requires a software pipeline designed around its architecture.

Andromeda combines embedding gathering, user–ad model interaction and score calculation in a single GPU kernel. Embeddings are numerical representations of features. Intermediate results stay in registers and thread-block shared memory instead of being repeatedly written to high-bandwidth memory, or HBM. This follows the broader principle in CUDA optimization guidance: minimize data movement and keep data close to computation.

The network was another obstacle. Targeting checks and final candidate selection previously ran on CPU hosts separately from GPU scoring. Sending long candidate lists between them saturated network connections. Index matching, scoring and top-K selection now run on the same GPU. The server receives a compact request containing user features and a targeting query, then returns only the final shortlist. This reduces network traffic by approximately 100 times.

Moving index matching required preserving complex targeting rules across dozens of feature categories and supporting real-time updates. New, modified and expired ads still need to be handled correctly under the redesigned architecture.

What the NVIDIA Tests Show

Meta and NVIDIA began jointly optimizing pipeline components in early 2026. NVIDIA develops accelerated GPU kernels, while Meta integrates them and benchmarks the full system. Reported component speedups are roughly twofold to fivefold; the routing network improved 5.3 times, rising from 2.4 billion to more than 10 billion ad evaluations per second.

The optimized kernel reaches approximately 85% of peak HBM bandwidth. Routing latency approaches that of a simple dot-product scoring kernel on the same hardware, despite the model capturing more complex relationships. In tests, moving from FP16 to INT8 delivered up to a further twofold throughput increase with comparable accuracy.

Index matching became more than twice as fast, exceeding 100 billion ads per second. Aligning data layouts between stages removed additional overhead. In benchmarks, the full pipeline achieved several times the current production throughput, with further gains expected as optimizations enter production.

These figures describe different processing levels. More than 10 billion evaluations per second refers to the computing kernel on a single GH200; more than 100 billion refers to index matching; and more than one billion refers to the complete production pipeline on a GPU server. Adding them together or presenting one stage’s speed as the speed of the whole system would be misleading. They also measure candidate evaluations, rather than ad impressions delivered to users.

What Changes for Advertisers

The main practical benefit is that a stronger retrieval model can consider more relevant options within the existing response-time budget. The improvement happens inside the advertising infrastructure. It does not, by itself, prescribe a universal budget strategy, campaign structure or guaranteed increase in ROAS.

Further work covers Grace Blackwell and subsequent platform generations. Areas under investigation include CPU-to-GPU resource ratios, deeper interaction models, additional experts and new retrieval pathways. The available performance headroom is intended to support more complex models without comparable infrastructure growth. These directions remain under exploration, so Hopper benchmark results cannot automatically be transferred to future configurations.

Meta has updated Andromeda, the system that selects ads for further evaluation on Facebook and Instagram. Its complete pipeline now processes more than one billion ads per second on a single GPU server. The company puts this at roughly 100 times the throughput of its 2024 system. The improvement comes from a new neural network architecture, fused GPU operations and less data moving between servers.

For advertisers, these changes mean the system can make better matches between people and ads across a larger candidate pool while keeping the same latency budget. Meta attributes a 2.4% increase in Facebook conversions to Andromeda’s development since 2025. That is the company’s reported overall system impact, rather than a promise for every campaign or an effect attributable solely to the 2026 update.

Why Ad Selection Starts with Candidate Retrieval

Ad recommendation involves several stages. First, retrieval reduces a vast pool of eligible ads to a small shortlist within milliseconds. More computationally expensive ranking models then evaluate the surviving candidates in detail.

The quality of the first stage determines which options reach the next one. If a potentially useful ad is discarded too early, later ranking cannot bring it back. Yet applying an equally complex model to the entire initial pool would be too expensive under strict response-time and computing constraints.

That trade-off becomes harder as ad volumes grow. Advantage+ automation and generative creative tools add more candidates to consider. Andromeda therefore needs to broaden its initial evaluation while maintaining matching quality and a manageable processing cost.

How Multiple Experts Capture User Interests

The updated model uses Retrieval Mixture-of-Experts, or MoE. The general principle behind a mixture-of-experts architecture is to activate only a subset of specialized subnetworks for each example. This allows model capacity to grow without a proportional increase in computation for every request.

Andromeda replaces a single user representation tower with an ensemble of experts. They capture changing interests and context through sequences of ad interactions, user-cohort representations, semantic features derived from large language models and cross-surface co-engagement signals.

The ad representation also incorporates the current serving context. An attention-based routing network selects the most relevant user experts to evaluate a particular candidate. Rather than handling every user–ad pair in the same way, the system draws on model specialization.

Meta describes the resulting matching quality as approaching that of ranking models. Meanwhile, inference requires approximately 300 times fewer floating-point operations, or FLOPs, than the 2024 Andromeda model. This figure should not be read as an equivalent reduction in server costs or an equivalent speedup of the entire advertising system.

Why Move the Entire Retrieval Flow onto a GPU

At billions of candidate evaluations per second, moving data becomes a constraint. Even a small intermediate array for each ad can generate terabytes of memory traffic per second. Fast arithmetic is of limited use if the processor keeps waiting for data.

The NVIDIA GH200 Grace Hopper platform matters here. It connects a Grace CPU and a Hopper GPU through NVLink-C2C, providing a coherent interface with 900 GB/s of bandwidth. Realizing that hardware potential also requires a software pipeline designed around its architecture.

Andromeda combines embedding gathering, user–ad model interaction and score calculation in a single GPU kernel. Embeddings are numerical representations of features. Intermediate results stay in registers and thread-block shared memory instead of being repeatedly written to high-bandwidth memory, or HBM. This follows the broader principle in CUDA optimization guidance: minimize data movement and keep data close to computation.

The network was another obstacle. Targeting checks and final candidate selection previously ran on CPU hosts separately from GPU scoring. Sending long candidate lists between them saturated network connections. Index matching, scoring and top-K selection now run on the same GPU. The server receives a compact request containing user features and a targeting query, then returns only the final shortlist. This reduces network traffic by approximately 100 times.

Moving index matching required preserving complex targeting rules across dozens of feature categories and supporting real-time updates. New, modified and expired ads still need to be handled correctly under the redesigned architecture.

What the NVIDIA Tests Show

Meta and NVIDIA began jointly optimizing pipeline components in early 2026. NVIDIA develops accelerated GPU kernels, while Meta integrates them and benchmarks the full system. Reported component speedups are roughly twofold to fivefold; the routing network improved 5.3 times, rising from 2.4 billion to more than 10 billion ad evaluations per second.

The optimized kernel reaches approximately 85% of peak HBM bandwidth. Routing latency approaches that of a simple dot-product scoring kernel on the same hardware, despite the model capturing more complex relationships. In tests, moving from FP16 to INT8 delivered up to a further twofold throughput increase with comparable accuracy.

Index matching became more than twice as fast, exceeding 100 billion ads per second. Aligning data layouts between stages removed additional overhead. In benchmarks, the full pipeline achieved several times the current production throughput, with further gains expected as optimizations enter production.

These figures describe different processing levels. More than 10 billion evaluations per second refers to the computing kernel on a single GH200; more than 100 billion refers to index matching; and more than one billion refers to the complete production pipeline on a GPU server. Adding them together or presenting one stage’s speed as the speed of the whole system would be misleading. They also measure candidate evaluations, rather than ad impressions delivered to users.

What Changes for Advertisers

The main practical benefit is that a stronger retrieval model can consider more relevant options within the existing response-time budget. The improvement happens inside the advertising infrastructure. It does not, by itself, prescribe a universal budget strategy, campaign structure or guaranteed increase in ROAS.

Further work covers Grace Blackwell and subsequent platform generations. Areas under investigation include CPU-to-GPU resource ratios, deeper interaction models, additional experts and new retrieval pathways. The available performance headroom is intended to support more complex models without comparable infrastructure growth. These directions remain under exploration, so Hopper benchmark results cannot automatically be transferred to future configurations.

Meta has updated Andromeda, the system that selects ads for further evaluation on Facebook and Instagram. Its complete pipeline now processes more than one billion ads per second on a single GPU server. The company puts this at roughly 100 times the throughput of its 2024 system. The improvement comes from a new neural network architecture, fused GPU operations and less data moving between servers.

For advertisers, these changes mean the system can make better matches between people and ads across a larger candidate pool while keeping the same latency budget. Meta attributes a 2.4% increase in Facebook conversions to Andromeda’s development since 2025. That is the company’s reported overall system impact, rather than a promise for every campaign or an effect attributable solely to the 2026 update.

Why Ad Selection Starts with Candidate Retrieval

Ad recommendation involves several stages. First, retrieval reduces a vast pool of eligible ads to a small shortlist within milliseconds. More computationally expensive ranking models then evaluate the surviving candidates in detail.

The quality of the first stage determines which options reach the next one. If a potentially useful ad is discarded too early, later ranking cannot bring it back. Yet applying an equally complex model to the entire initial pool would be too expensive under strict response-time and computing constraints.

That trade-off becomes harder as ad volumes grow. Advantage+ automation and generative creative tools add more candidates to consider. Andromeda therefore needs to broaden its initial evaluation while maintaining matching quality and a manageable processing cost.

How Multiple Experts Capture User Interests

The updated model uses Retrieval Mixture-of-Experts, or MoE. The general principle behind a mixture-of-experts architecture is to activate only a subset of specialized subnetworks for each example. This allows model capacity to grow without a proportional increase in computation for every request.

Andromeda replaces a single user representation tower with an ensemble of experts. They capture changing interests and context through sequences of ad interactions, user-cohort representations, semantic features derived from large language models and cross-surface co-engagement signals.

The ad representation also incorporates the current serving context. An attention-based routing network selects the most relevant user experts to evaluate a particular candidate. Rather than handling every user–ad pair in the same way, the system draws on model specialization.

Meta describes the resulting matching quality as approaching that of ranking models. Meanwhile, inference requires approximately 300 times fewer floating-point operations, or FLOPs, than the 2024 Andromeda model. This figure should not be read as an equivalent reduction in server costs or an equivalent speedup of the entire advertising system.

Why Move the Entire Retrieval Flow onto a GPU

At billions of candidate evaluations per second, moving data becomes a constraint. Even a small intermediate array for each ad can generate terabytes of memory traffic per second. Fast arithmetic is of limited use if the processor keeps waiting for data.

The NVIDIA GH200 Grace Hopper platform matters here. It connects a Grace CPU and a Hopper GPU through NVLink-C2C, providing a coherent interface with 900 GB/s of bandwidth. Realizing that hardware potential also requires a software pipeline designed around its architecture.

Andromeda combines embedding gathering, user–ad model interaction and score calculation in a single GPU kernel. Embeddings are numerical representations of features. Intermediate results stay in registers and thread-block shared memory instead of being repeatedly written to high-bandwidth memory, or HBM. This follows the broader principle in CUDA optimization guidance: minimize data movement and keep data close to computation.

The network was another obstacle. Targeting checks and final candidate selection previously ran on CPU hosts separately from GPU scoring. Sending long candidate lists between them saturated network connections. Index matching, scoring and top-K selection now run on the same GPU. The server receives a compact request containing user features and a targeting query, then returns only the final shortlist. This reduces network traffic by approximately 100 times.

Moving index matching required preserving complex targeting rules across dozens of feature categories and supporting real-time updates. New, modified and expired ads still need to be handled correctly under the redesigned architecture.

What the NVIDIA Tests Show

Meta and NVIDIA began jointly optimizing pipeline components in early 2026. NVIDIA develops accelerated GPU kernels, while Meta integrates them and benchmarks the full system. Reported component speedups are roughly twofold to fivefold; the routing network improved 5.3 times, rising from 2.4 billion to more than 10 billion ad evaluations per second.

The optimized kernel reaches approximately 85% of peak HBM bandwidth. Routing latency approaches that of a simple dot-product scoring kernel on the same hardware, despite the model capturing more complex relationships. In tests, moving from FP16 to INT8 delivered up to a further twofold throughput increase with comparable accuracy.

Index matching became more than twice as fast, exceeding 100 billion ads per second. Aligning data layouts between stages removed additional overhead. In benchmarks, the full pipeline achieved several times the current production throughput, with further gains expected as optimizations enter production.

These figures describe different processing levels. More than 10 billion evaluations per second refers to the computing kernel on a single GH200; more than 100 billion refers to index matching; and more than one billion refers to the complete production pipeline on a GPU server. Adding them together or presenting one stage’s speed as the speed of the whole system would be misleading. They also measure candidate evaluations, rather than ad impressions delivered to users.

What Changes for Advertisers

The main practical benefit is that a stronger retrieval model can consider more relevant options within the existing response-time budget. The improvement happens inside the advertising infrastructure. It does not, by itself, prescribe a universal budget strategy, campaign structure or guaranteed increase in ROAS.

Further work covers Grace Blackwell and subsequent platform generations. Areas under investigation include CPU-to-GPU resource ratios, deeper interaction models, additional experts and new retrieval pathways. The available performance headroom is intended to support more complex models without comparable infrastructure growth. These directions remain under exploration, so Hopper benchmark results cannot automatically be transferred to future configurations.

Meta has updated Andromeda, the system that selects ads for further evaluation on Facebook and Instagram. Its complete pipeline now processes more than one billion ads per second on a single GPU server. The company puts this at roughly 100 times the throughput of its 2024 system. The improvement comes from a new neural network architecture, fused GPU operations and less data moving between servers.

For advertisers, these changes mean the system can make better matches between people and ads across a larger candidate pool while keeping the same latency budget. Meta attributes a 2.4% increase in Facebook conversions to Andromeda’s development since 2025. That is the company’s reported overall system impact, rather than a promise for every campaign or an effect attributable solely to the 2026 update.

Why Ad Selection Starts with Candidate Retrieval

Ad recommendation involves several stages. First, retrieval reduces a vast pool of eligible ads to a small shortlist within milliseconds. More computationally expensive ranking models then evaluate the surviving candidates in detail.

The quality of the first stage determines which options reach the next one. If a potentially useful ad is discarded too early, later ranking cannot bring it back. Yet applying an equally complex model to the entire initial pool would be too expensive under strict response-time and computing constraints.

That trade-off becomes harder as ad volumes grow. Advantage+ automation and generative creative tools add more candidates to consider. Andromeda therefore needs to broaden its initial evaluation while maintaining matching quality and a manageable processing cost.

How Multiple Experts Capture User Interests

The updated model uses Retrieval Mixture-of-Experts, or MoE. The general principle behind a mixture-of-experts architecture is to activate only a subset of specialized subnetworks for each example. This allows model capacity to grow without a proportional increase in computation for every request.

Andromeda replaces a single user representation tower with an ensemble of experts. They capture changing interests and context through sequences of ad interactions, user-cohort representations, semantic features derived from large language models and cross-surface co-engagement signals.

The ad representation also incorporates the current serving context. An attention-based routing network selects the most relevant user experts to evaluate a particular candidate. Rather than handling every user–ad pair in the same way, the system draws on model specialization.

Meta describes the resulting matching quality as approaching that of ranking models. Meanwhile, inference requires approximately 300 times fewer floating-point operations, or FLOPs, than the 2024 Andromeda model. This figure should not be read as an equivalent reduction in server costs or an equivalent speedup of the entire advertising system.

Why Move the Entire Retrieval Flow onto a GPU

At billions of candidate evaluations per second, moving data becomes a constraint. Even a small intermediate array for each ad can generate terabytes of memory traffic per second. Fast arithmetic is of limited use if the processor keeps waiting for data.

The NVIDIA GH200 Grace Hopper platform matters here. It connects a Grace CPU and a Hopper GPU through NVLink-C2C, providing a coherent interface with 900 GB/s of bandwidth. Realizing that hardware potential also requires a software pipeline designed around its architecture.

Andromeda combines embedding gathering, user–ad model interaction and score calculation in a single GPU kernel. Embeddings are numerical representations of features. Intermediate results stay in registers and thread-block shared memory instead of being repeatedly written to high-bandwidth memory, or HBM. This follows the broader principle in CUDA optimization guidance: minimize data movement and keep data close to computation.

The network was another obstacle. Targeting checks and final candidate selection previously ran on CPU hosts separately from GPU scoring. Sending long candidate lists between them saturated network connections. Index matching, scoring and top-K selection now run on the same GPU. The server receives a compact request containing user features and a targeting query, then returns only the final shortlist. This reduces network traffic by approximately 100 times.

Moving index matching required preserving complex targeting rules across dozens of feature categories and supporting real-time updates. New, modified and expired ads still need to be handled correctly under the redesigned architecture.

What the NVIDIA Tests Show

Meta and NVIDIA began jointly optimizing pipeline components in early 2026. NVIDIA develops accelerated GPU kernels, while Meta integrates them and benchmarks the full system. Reported component speedups are roughly twofold to fivefold; the routing network improved 5.3 times, rising from 2.4 billion to more than 10 billion ad evaluations per second.

The optimized kernel reaches approximately 85% of peak HBM bandwidth. Routing latency approaches that of a simple dot-product scoring kernel on the same hardware, despite the model capturing more complex relationships. In tests, moving from FP16 to INT8 delivered up to a further twofold throughput increase with comparable accuracy.

Index matching became more than twice as fast, exceeding 100 billion ads per second. Aligning data layouts between stages removed additional overhead. In benchmarks, the full pipeline achieved several times the current production throughput, with further gains expected as optimizations enter production.

These figures describe different processing levels. More than 10 billion evaluations per second refers to the computing kernel on a single GH200; more than 100 billion refers to index matching; and more than one billion refers to the complete production pipeline on a GPU server. Adding them together or presenting one stage’s speed as the speed of the whole system would be misleading. They also measure candidate evaluations, rather than ad impressions delivered to users.

What Changes for Advertisers

The main practical benefit is that a stronger retrieval model can consider more relevant options within the existing response-time budget. The improvement happens inside the advertising infrastructure. It does not, by itself, prescribe a universal budget strategy, campaign structure or guaranteed increase in ROAS.

Further work covers Grace Blackwell and subsequent platform generations. Areas under investigation include CPU-to-GPU resource ratios, deeper interaction models, additional experts and new retrieval pathways. The available performance headroom is intended to support more complex models without comparable infrastructure growth. These directions remain under exploration, so Hopper benchmark results cannot automatically be transferred to future configurations.