Connect with us

Compute & Silicon

Quantization: How AI Models Are Shrunk to Run Faster

Quantization stores a model’s numbers with fewer bits so it uses less memory and power. Here is how it works and what tradeoffs it brings.

Published

on

Green circuit board with a central processor and glowing light

One of the most effective ways to make an AI model cheaper and faster to run does not involve new hardware at all. It involves changing how the model’s numbers are stored. The technique is called quantization.

The basic idea

A trained model is made of billions of numbers called weights. During training, those numbers are usually kept in formats that use 16 or 32 bits each, which allows fine distinctions between values.

Quantization converts those numbers to formats with fewer bits, such as 8 bits or even 4 bits. Think of it like rounding prices to the nearest dime instead of the nearest penny. You lose a little precision, but the list becomes much shorter to write down.

Advertisement

Why fewer bits help

Using fewer bits per number brings several benefits at once:

  • The model takes up less memory, so it can fit on fewer or smaller chips.
  • Less data has to move between memory and the processor, easing the memory bandwidth bottleneck.
  • Many chips can perform more low-precision operations per second than high-precision ones.
  • Less data movement and simpler arithmetic usually mean less energy per answer.

For inference, where a model runs again and again, those savings add up. Quantization is also a big reason capable models can run on laptops and phones.

The tradeoff: accuracy

Rounding numbers introduces small errors. In many cases, the model’s output barely changes. Push the rounding too far, though, and quality can degrade, with the model making more mistakes or giving less coherent answers.

Researchers have developed many techniques to limit that loss:

  • Calibration, which studies typical data flowing through the model to choose the best rounding ranges.
  • Keeping especially sensitive parts of the model at higher precision while compressing the rest.
  • Quantization-aware training, which lets the model adjust to lower precision while it learns.
  • Grouping weights into small blocks, each with its own scale, so rounding fits local values better.

How well a quantized model performs depends on the model, the method and the task. Careful testing is standard practice.

Advertisement

Weights, activations and more

Quantization can be applied to different parts of a model. Weights are the stored parameters. Activations are the intermediate values computed while the model runs. Some approaches compress only weights. Others compress both, which saves more but is harder to do without hurting quality. Working memory such as the caches used during text generation can be compressed too.

Why hardware designers care

Chipmakers increasingly build support for low-precision formats directly into their hardware, because they know so much AI work can run at reduced precision. Each new generation of accelerators tends to add or improve support for smaller number formats.

This interplay between software technique and hardware design is a recurring theme in AI infrastructure. Gains do not only come from faster transistors. They also come from doing the same useful work with less data, less memory and less energy.

A Worked Example: Rounding Eight Numbers

Quantization becomes easier to picture with a tiny example. Imagine a small block of weights:

0.42, -0.17, 0.08, -0.61, 0.33, 0.95, -0.04, -0.29

Suppose these need to be stored as 4-bit signed integers, which can represent whole numbers from -8 to 7. A simple symmetric method works like this:

  1. Find the largest magnitude. Here it is 0.95.
  2. Compute a scale. Divide 0.95 by 7, the largest positive value available, giving a scale of about 0.136.
  3. Divide and round. Each weight is divided by the scale and rounded. For instance, 0.42 becomes about 3.1, which rounds to 3. The value -0.61 becomes about -4.5, which rounds to -4 or -5 depending on the rounding rule.
  4. Store the integers and the scale. Instead of eight 16-bit numbers, the block holds eight 4-bit integers plus one scale value.
  5. Reconstruct when needed. Multiplying each integer by the scale gives an approximation. The stored 3 becomes about 0.41, close to the original 0.42.

The small differences between original and reconstructed values are the quantization error. Using small blocks, each with its own scale, keeps that error low because each scale fits its neighborhood of values more closely. One large outlier in a block, however, can stretch the scale and reduce precision for everything else, which is why outlier handling is an active area of technique.

Choosing a Quantization Approach

Teams deciding how to quantize a model usually weigh a few practical questions:

  • Where will it run? A phone, a laptop and a data center accelerator each support different formats efficiently.
  • How much quality loss is acceptable? A casual chat feature may tolerate more than a coding assistant or a medical documentation tool.
  • Is retraining possible? Post-training methods are quicker. Quantization-aware training can preserve more quality but costs more compute.
  • What is the bottleneck? If memory capacity is the constraint, compressing weights helps most. If arithmetic speed is the constraint, compressing activations too may be needed.

Common mistakes

  • Judging a quantized model only on average benchmark scores, while missing drops on harder or less common tasks.
  • Calibrating with data that does not resemble real usage.
  • Assuming a format will run faster on any hardware, when some chips lack native support and must convert numbers back.
  • Compressing every layer equally, including especially sensitive ones.

What This Means in Everyday Terms

Quantization helps explain why AI features increasingly appear on personal devices and why the cost of running models can fall even without new chips. It is one of the ways the industry extracts more useful work from existing hardware. For readers following AI infrastructure, it is a reminder that progress comes from software and mathematics as well as silicon. This is general educational context, not investment advice; anyone making financial decisions should consult a qualified financial professional.

Questions Readers Ask

Can quantized models be converted back?

The lost precision cannot be recovered from the compressed version alone. Developers typically keep the original higher-precision model as well.

Is quantization the same as distillation?

No. Distillation trains a smaller model to imitate a larger one. Quantization keeps the same model but stores its numbers with fewer bits. The two can be combined.

How do I know whether a quantized model is good enough?

The practical test is to compare it with the original on tasks that resemble real use, not just a single headline benchmark. Many teams also keep a set of tricky examples, such as long documents, math problems or uncommon languages, because those are often where the effects of heavy rounding show up first.

Latest stories

View all

Compute & Silicon

Why a Cold Cache Changes the Meaning of a Fast Response Test

A service that has recently handled a request may respond differently from one meeting it for the first time. Data, compiled work or model state can already be available, so a warm test and a cold test answer different questions.

Published

on

A rack of servers in a server room

Photo: Kevin Ache / Unsplash

A service that has recently handled a request may respond differently from one meeting it for the first time. Data, compiled work or model state can already be available, so a warm test and a cold test answer different questions.

The operating distinction

Caches keep selected information close to where it will be used. Reusing that information can reduce retrieval or preparation work. A cold start may need to rebuild it, load a model or fetch data from a slower location. Neither condition is inherently the only valid test; the expected workload determines the mix.

Advertisement

Preparation can also occur outside the part of the system being timed. A demonstration may load information before starting its stopwatch, while a user experiences that loading as part of the wait. That does not invalidate the demonstration, but it changes the boundary of the claim.

A practical example

Imagine comparing two demonstrations of the same service. One repeats a familiar query while the other begins after a restart. Their response times may differ without any change in the underlying hardware. A fair comparison describes what was already loaded and what preparation was included.

Questions worth asking

A comparison should include preparation that matters to the intended user experience. Keeping warm and cold observations separate makes both useful without presenting the faster result as representative of every possible request.

  • Was the test performed after restart or after repeated requests?
  • Which preparation steps were included in the measured response?
  • How often does the real workload encounter the cold condition?

What the evidence supports

A single fast example cannot establish the experience across restarts, uncommon requests and traffic changes. Conversely, an unusually slow first request should not automatically represent steady operation. Readers need the operating state, the sampling method and the service objective before drawing a conclusion.

A useful comparison can present both first-use and repeated-use observations under clearly described conditions. It should identify whether the test reuses the same request or exercises a broader set of inputs. The reader can then decide which observation resembles the intended application. The aim is an honest description of service behavior, not a contest that rewards whichever test omits the most preparation work.

Advertisement

Treat cache state as part of the test description, alongside the machine and the application.

Quick check

How much did you catch? Tap a question to see the answer.

1. Why do warm and cold tests give different results?

Answer: Data, compiled work or model state may already be loaded in a warm test.

2. Is a cold test the only valid one?

Answer: No. The expected workload decides the right mix.

3. What should a fair comparison describe?

Answer: What was already loaded and what preparation was included in the timing.

Where Warm State Lives in a Modern Service

The word cache can sound like a single component, but a typical service holds several layers of prepared state at once. Each layer can be warm or cold independently, which is one reason response tests can be difficult to compare.

  • Model weights in memory. A model already loaded onto an accelerator is ready to run. A model that must be read from storage first adds loading time.
  • Compiled or optimized code. Some systems prepare optimized execution paths the first time a particular shape of work appears and reuse them afterward.
  • Request-level reuse. Inference systems may reuse computation for repeated prompt prefixes or recently seen inputs.
  • Application data caches. Frequently retrieved documents, records or results may be held in faster memory near the service.
  • Connection and session state. Established network connections, authentication tokens and open sessions avoid setup work for subsequent requests.
  • Storage and operating system caches. Files read recently may already sit in memory, making later reads faster.

Designing a Fairer Response Test

Anyone evaluating a claim about response speed, or designing a test of their own, can work through a short sequence.

  1. Describe the starting state. State whether the test began after a restart, after a period of inactivity or after a run of similar requests.
  2. Define the timing boundary. Record exactly when the clock starts and stops, and which preparation steps fall inside or outside that window.
  3. Vary the inputs. Use a set of requests that resembles the real workload rather than repeating one query.
  4. Report a distribution. Present typical and slower responses, not only the fastest or the average.
  5. Estimate the real cold rate. Consider how often actual users will meet a cold condition, based on traffic patterns, scaling behavior and restarts.
  6. Keep the results labeled. Publish warm and cold observations side by side with their conditions.

Common mistakes

A frequent mistake is running a test many times and reporting the best result, which almost always reflects fully warm conditions. Another is forgetting that automatic scaling can introduce cold starts whenever new capacity is added during a traffic increase. A third is comparing two systems where one test included model loading and the other did not. A fourth is assuming that a cold start happens only after a deliberate restart, when cache eviction under memory pressure can produce similar effects.

What This Means for Buyers and Readers

For a business choosing an inference service or reading a vendor’s performance claim, the cache question is practical. A service that is consistently fast for repeated requests may still feel slow to users who arrive with varied or uncommon requests, or who arrive just after a scaling event. Asking how the provider handles cold conditions, and how often they occur in practice, gives a more realistic sense of what users will experience.

Fast is a description of a particular state. The useful question is how often the real workload is in that state.

Questions Readers Ask

Is a warm test misleading?

Not if it is labeled. Warm results describe steady operation for repeated work, which is useful. They become misleading only when presented as the whole experience.

Can cold starts be eliminated?

They can be reduced by keeping capacity loaded and preparing state in advance, but this usually carries a cost in reserved resources.

Why does the first request after a restart take longer?

The service may need to load models, prepare execution paths, open connections and fill caches before it can respond.

What should a buyer ask a provider?

Ask what conditions their published response figures assume and how the service behaves after restarts, scaling events and long idle periods.

BOSSTOX is coming to iPhone

AI infrastructure explainers and news on iPhone, iPad and Apple Watch.

See BOSSTOX

Continue Reading

Compute & Silicon

An Idle Accelerator Can Still Be Waiting on Useful Work

A low utilization reading does not identify a single problem. An accelerator can be waiting for data, another machine, a scheduling decision or a request from a user. Understanding the waiting state matters more than assigning blame to the chip.

Published

on

Close-up of a circuit board

Photo: Anne Nygård / Unsplash

A low utilization reading does not identify a single problem. An accelerator can be waiting for data, another machine, a scheduling decision or a request from a user. Understanding the waiting state matters more than assigning blame to the chip.

The operating distinction

Useful work passes through several stages before an accelerator starts arithmetic. Data must be available, memory must have room and dependent tasks must finish. A scheduler may also hold capacity for a workload with a deadline. That reserved headroom can look similar to inefficiency on a broad average.

Advertisement

Different kinds of waiting call for different responses. Adding processors will not necessarily repair a slow input pipeline, and changing scheduling can affect one customer's response while improving another's throughput. The resource graph and the service objective need to be read together.

A practical example

Imagine an inference service that keeps spare resources available for sudden requests. Filling every available slot with background work could increase the average utilization while making urgent requests slower. A training job with predictable input and long execution periods would present a different operating problem.

Questions worth asking

Reserved headroom, missing input and inefficient execution are different explanations for unused capacity. The next operational question should follow the observed cause rather than assume that every idle interval represents the same problem.

  • What resource or dependency is the accelerator waiting for?
  • Does the measurement include reserved capacity and maintenance periods?
  • Did higher utilization improve completed work or merely lengthen queues?

What the evidence supports

Utilization needs a definition and a time window. A device-level percentage alone cannot show whether a service met its response target or whether a job completed on schedule. Compare the waiting causes and user outcome before comparing the headline averages.

A useful account identifies the interval when the observation was made and describes the mix of jobs present. Readers can then distinguish an idle system with no demand from a busy service deliberately retaining headroom. Both can show unused device capacity, but they tell different stories about the amount of work waiting and whether customers are receiving the experience the service was designed to provide.

Advertisement

The useful target is productive work within the service requirement, not a percentage detached from its purpose.

Quick check

How much did you catch? Tap a question to see the answer.

1. Does low utilization point to a single problem?

Answer: No. An accelerator may be waiting on data, another machine, scheduling or users.

2. Why can reserved headroom look inefficient?

Answer: It can resemble idle capacity on a broad average.

3. Will adding processors fix a slow input pipeline?

Answer: Not necessarily.

Sorting Idle Time Into Categories

A single utilization line becomes far more useful once the unused intervals are split into a few named categories. The categories do not need to be elaborate. They only need to separate explanations that call for different responses, so that a review meeting can discuss causes instead of a number.

  • No demand. Nothing was waiting. The service was available and nobody needed it during that interval.
  • Deliberate headroom. Capacity was held back on purpose so urgent or latency-sensitive requests could start quickly.
  • Input starvation. Work existed, but data, model weights or request batches were not arriving fast enough to keep the device busy.
  • Coordination waits. The device was ready but waiting on another machine, a synchronization step or a dependent task.
  • Scheduling gaps. Work was queued, yet placement, priority or fragmentation kept it from landing on available hardware.
  • Planned unavailability. Maintenance, testing, updates or recovery activity removed the device from service.

Even a rough split changes the conversation. If most idle time belongs to deliberate headroom, the question becomes whether the reserve is sized correctly for the response target. If most belongs to input starvation, the attention moves to storage, networking or preprocessing. Neither answer requires buying or removing accelerators.

A Step-by-Step Review of Unused Capacity

An operator or analyst trying to understand a low reading can work through a short sequence before drawing conclusions.

  1. Fix the time window. State whether the figure covers an hour, a day, a month or a reporting quarter. Overnight lulls and weekend demand can dominate a long average.
  2. State what counts as busy. Some measurements record whether any work is scheduled on the device, while others try to estimate how much of its arithmetic capability is in use. The two can tell very different stories.
  3. Separate the workload mix. Training, batch processing and interactive inference have different shapes. Averaging them together hides the behavior of each.
  4. Place the waits on a timeline. Line up idle intervals against data transfers, checkpoint writes, job starts and request bursts. Patterns often appear where they would never show in a single percentage.
  5. Check the outcome. Ask whether jobs finished on schedule and whether response targets were met during the same window.
  6. Choose the response that matches the cause. Only then decide whether the answer is more capacity, better scheduling, a faster input path or no change at all.

Common mistakes

The most frequent error is treating utilization as a goal in itself. A team rewarded for a higher figure can fill reserved headroom with low-priority work, which raises the average while making urgent requests wait. Another mistake is comparing utilization across facilities or providers without confirming that the definitions match. A third is ignoring the cost of reaching a higher number, such as longer queues or more frequent preemption of jobs that then have to restart.

What This Means for Readers Following the Industry

Utilization figures appear in earnings discussions, vendor presentations and infrastructure announcements. For a reader trying to understand the business, the figure is a starting point for questions rather than a verdict. A high number may reflect strong demand, or it may reflect a fleet with too little headroom for spikes. A low number may signal weak demand, or a service that keeps capacity in reserve because its customers expect fast responses.

The more useful question is usually what the figure was designed to show and what it leaves out.

This is general education about how infrastructure metrics are read, not a view on any company or security. Anyone weighing a financial decision should consult a qualified financial professional about their own situation.

Questions Readers Ask

Is a fully utilized accelerator always the efficient outcome?

Not necessarily. A device kept constantly busy may leave nothing available for sudden requests. Whether that is efficient depends on the service the device is meant to support.

Can idle time be measured precisely?

It can be measured, but the result depends on the definition, the sampling interval and the tools used. Short waits can disappear inside coarse samples.

Why would a provider intentionally leave capacity unused?

To meet response commitments, absorb failures, allow maintenance or accept new work without delay. Each of those purposes has value that an average cannot display.

What is the single best follow-up question?

Ask what the accelerator was waiting for during the idle intervals. The answer points toward the next useful decision.

BOSSTOX is coming to iPhone

AI infrastructure explainers and news on iPhone, iPad and Apple Watch.

See BOSSTOX

Continue Reading

Compute & Silicon

Custom AI ASICs vs. GPUs

Custom AI ASICs can deliver strong efficiency on a narrow workload, while GPUs remain valuable because their software ecosystem and flexibility absorb uncertainty.

Published

on

Close-up of a graphics card heatsink

The custom-ASIC-versus-GPU debate is often framed as a contest between a cheaper chip and a more expensive one. That misses the real decision. A GPU is a general-purpose accelerator with a large software ecosystem. An application-specific integrated circuit is a narrower machine designed around a workload, a power target, or a system architecture.

The winner depends on whether the workload is stable enough to justify specialization. Training, inference, recommendation, search, and video pipelines do not stress hardware in exactly the same way. A model that is efficient today can change after a new architecture, quantization method, context length, or customer demand appears.

What the GPU buys

A GPU’s first advantage is optionality. It can run different model families, numerical formats, kernels, and frameworks. It can be reassigned when demand shifts. That matters for a cloud provider serving many customers and for a company whose AI roadmap is still moving.

Advertisement

The second advantage is software. Mature compiler tools, libraries, kernels, orchestration systems, and developer familiarity reduce the risk between purchasing hardware and putting it into production. A GPU’s list price is not the whole cost; neither is an ASIC’s lower unit cost if engineers spend months rewriting software.

The third advantage is scale. GPUs can be linked into standard or proprietary fabrics, with established server, cooling, and support ecosystems. A customer can often rent them before it owns them, which converts a large capital decision into a more flexible operating expense.

The GPU’s weaknesses are also clear. The chip may carry circuitry for workloads the customer does not use. Its memory and network design may be optimized for a broad market rather than one application. High demand can create shortages, and power density can push a facility’s cooling and electrical systems beyond their original design.

What the ASIC buys

A custom ASIC can remove general-purpose features and use the saved area or power for the specific operations that matter. It may deliver more performance per watt, better latency, or lower cost per inference when the workload is predictable.

An ASIC can also be part of a broader system strategy. Google and other hyperscalers have built custom accelerators because owning the silicon roadmap can reduce dependence on a merchant vendor and let the company tune hardware to its software stack. Intel described an expanded 2026 collaboration with Google around custom ASIC-based infrastructure processing units that offload networking, storage, and security functions from host CPUs. That is a reminder that “custom AI silicon” is not limited to a chip that performs matrix multiplication.

Advertisement

The costs appear before production

ASIC design requires architecture, verification, physical design, software, packaging, testing, and manufacturing commitments. A company must pay those costs before it knows how many units it will ship.

The minimum efficient volume can be high. If a custom chip saves $1,000 per accelerator but costs $300 million to develop and qualify, the business needs more than 300,000 units before ignoring financing, support, and inventory risk. If the workload changes before that volume is reached, the savings may disappear.

Advanced packaging and high-bandwidth memory can add another constraint. Micron reported in March 2025 that its HBM output for calendar 2025 was sold out and that it was already discussing 2026 demand with customers. A custom ASIC that cannot secure memory, substrates, and packaging is not a complete product.

A practical comparison must use system metrics

Compare the two options at the system level. Measure useful work per rack, tokens per second per watt, cost per million tokens, time to train, memory capacity, software migration effort, and failure-recovery behavior.

Do not compare a peak chip number from a data sheet with a measured production result from another platform. Ask whether the benchmark uses the same model, batch size, sequence length, precision, compiler version, and utilization target.

A GPU may win on time to market even if an ASIC wins on energy. The value of six months of earlier service can be material for a cloud provider. Conversely, an inference fleet with a stable model and high utilization may benefit from a specialized accelerator.

The market is moving toward heterogeneous systems

The likely outcome is not one winner. General-purpose GPUs will remain useful for training, new models, and workloads with uncertain demand. Custom ASICs will take more of the predictable inference, networking, storage, and recommendation work. CPUs, DPUs, and specialized accelerators will handle other parts of the system.

Intel’s Gaudi 3 product materials emphasize standard Ethernet, open software, and a migration path for customers moving from GPU-based systems. That positioning shows the market’s pressure point: customers want acceleration without locking every layer of the stack to one vendor.

The decision rule is simple. Use GPUs when flexibility, software, and deployment speed are worth paying for. Use an ASIC when the workload is stable, volume is large, memory and packaging are secured, and the organization can support a separate software and operations path.

The best analysis treats silicon as a fleet decision, not a chip-shopping exercise. The question is which architecture produces reliable, adaptable work at the lowest total cost over the period in which the workload is likely to remain stable.

There is also a strategic lock-in cost. A GPU fleet may be portable across cloud providers and model frameworks, while a custom accelerator can require a dedicated compiler, kernel library, scheduler, and support team. That lock-in may be acceptable when demand is large and stable. It is dangerous when the business case assumes a model architecture will remain unchanged for five years. The best ASIC programs preserve a fallback path: compatible APIs, exportable model formats, or enough GPU capacity to absorb new workloads while the specialized stack catches up.

Sources

  1. Micron, “Fiscal Q2 2025 Investor Materials,” March 20, 2025: investors.micron.com
  2. TSMC, “2025 Annual Report,” 2026: investor.tsmc.com
  3. Intel, “Intel Gaudi AI Accelerator Products,” accessed October 2026: intel.com
  4. Intel, “Intel and Google Deepen Collaboration to Advance AI Infrastructure,” April 9, 2026: newsroom.intel.com
  5. IEA, “Energy and AI,” April 10, 2025: iea.org

BOSSTOX Research articles are drafted with AI assistance and reviewed by our editors before publication. See our editorial standards.

Continue Reading

Trending