Picture driving towards Vice City in GTA 6. Across the water, you can recognise a hotel long before you can read the sign over its door. Open-world streaming gives that lettering a deadline. It has to be ready by the time you can read it from the road.
Rockstar released Grand Theft Auto VI on 19 November 2026 for PlayStation 5 and Xbox Series X|S. Two of its technology leads have described how the team made Vice City fit both in console memory and into the time available to draw a frame. Their answers also explain a familiar problem in web application development. A map can open quickly, then lose the street names the moment someone starts moving. Rockstar’s launch announcement
What Rockstar changed for open-world streaming
David Hynd, Rockstar’s vice president of technology, described three changes behind the increase in people and traffic. The team rebuilt the systems that choose how much detail to show, spread more work across threads so processors could handle it concurrently, and changed the priority given to assets competing for limited memory. The last part matters on a busy street. Every texture kept for the block behind you takes up space that an approaching junction could use. Restart’s interview with Rockstar’s developers
Consider two views of the same hotel. Across the bay, a simplified model can carry the outline of its roof and the broad colour of its facade. At the entrance, the shape of the doorway matters because a player can walk up to it. Level of detail governs which representation you see. Streaming moves the data for those representations into and out of memory as the view changes. Hynd described how the team prioritised that work. The interview leaves the exact distances at which GTA 6 switches models open.
Our browser driving game makes a similar decision on a smaller scale. In Free Ride, we generate ground locally in 160-metre chunks and keep at least 120 metres of road ready ahead of the car. The renderer draws the visible strip of terrain with a small margin on either side. The wheels need solid ground ready for the next hill well before the camera has to draw it. Play Rattletrap Ridge
What a hotel sign costs in memory
Start with one surface on that Vice City hotel. Up close, the player might be able to pick out every letter on a sign. Across the bay, the whole sign might occupy six pixels. A common way to serve both views is to prepare the same texture at successively smaller sizes. Each version is a mip level. The renderer can sample a smaller version when the sign occupies less of the screen. That also helps prevent narrow strokes from flickering as the camera moves.
How much memory does a sharper sign actually cost? Take an illustrative square texture, 4,096 pixels on each side, in Microsoft's BC7 format. BC7 stores each four-by-four block of texture elements in 16 bytes. The top level therefore has 1,024 blocks in each direction. Multiply 1,024 by 1,024 by 16 and you get 16,777,216 bytes, or 16 MiB. Here is what happens as the dimensions fall. These are texture payload sizes for the example, independent of Rockstar's choice of format. Microsoft's BC7 format specification
| Texture dimensions | Memory for that level alone |
|---|---|
| 4,096 by 4,096 | 16 MiB |
| 2,048 by 2,048 | 4 MiB |
| 1,024 by 1,024 | 1 MiB |
| 512 by 512 | 256 KiB |
The first row explains why a small change in viewing distance can have a large cost. If the smaller levels are already resident, bringing in the finest level adds the full 16 MiB. Imagine eight different signs competing for that upgrade at a junction. Loading the finest level of all eight would add 128 MiB of texture data before allowing for any other surface. If all eight shops share one sign texture, the renderer can reuse one resident copy. Counting shop fronts alone gives a poor estimate of the working set.
There is a finer choice than bringing in an entire level. Microsoft's sampler feedback example records which regions of a texture the renderer needs. An application can use that information to bring in selected tiles while continuing to draw from detail it already holds. Think of a hotel atlas with the entrance sign in one corner and roof fixtures in another. A pedestrian outside the entrance has given the engine a good reason to fetch the lettering. The roof fixtures can wait. This is a documented example of how a graphics API solves the prioritisation problem Hynd described. Rockstar has not identified this API as its implementation. Microsoft's explanation of sampler feedback
Why a road test needs a boat and a plane
A drive down the main road gives the engine a useful hint about what comes next. The buildings ahead grow gradually, and a bend may hide the next block until the driver reaches it. Approach the same district by water and its waterfront is visible from farther away. Fly over it and an entire strip of rooftops may come into view at once. The assets have stayed in the same places. Their deadlines have changed.
Paul MacPherson, Rockstar's senior vice president of environment art, told Restart that the team regularly traversed the map in different vehicles and from different angles. The artists were looking for stretches that felt empty or too busy, then working with design and narrative to make each area feel right. That playtest has an engineering use too. If QA only repeats journeys along roads, even an excellent road test gives little evidence about the first wide view from the water. Restart's interview with MacPherson and Hynd
To turn that observation into a repeatable test, choose a landmark visible by two routes. Mark the point on each route where a reader should first be able to recognise its silhouette, then mark where a smaller feature needs to become legible. Capture both approaches after a fresh start. Repeat after visiting the landmark once. A miss confined to the first approach points towards work that the second pass could reuse. A miss at the same viewing angle on every pass is a reason to examine the rule that selects the finer representation. Save the camera position with each capture so another engineer can reproduce the transition.
A sudden jump across a map is a third case. Epic documents a concrete solution in Unreal Engine's World Partition system. A developer can place a streaming source at a teleport destination, wait until the surrounding cells are ready, then move the player and remove the extra source. For a dispatch map, the equivalent action is clicking a vehicle in another city. Its destination is known as soon as the operator clicks. That gives the application a head start on loading the area and a clear moment to retire requests for the old view. Epic's World Partition documentation
Why GTA 6 targets 30 frames per second
Rockstar’s senior vice president of technology, Jonathon Ashcroft, explained that the console release targets 30 frames per second to accommodate both world simulation and rendering. That gives the game about 33.3 milliseconds between frames. A target of 40 fps would leave 25 milliseconds. Finding those extra 8.3 milliseconds on the path that determines when a frame finishes is a substantial engineering change. TechRadar’s interview with Rockstar
Take a junction full of cars. Lowering the resolution can reduce the graphics processor’s work drawing the scene. The processor updating vehicle movement still has to work out where the cars belong. If that update is already taking most of the interval, reducing the number of pixels buys little extra time before the next frame. That is one reason a screenshot alone tells you so little about the cost of a street.
There are two symptoms worth telling apart. A sign that sharpens late while the camera moves smoothly suggests checking when the detailed asset became available and when the renderer chose it. A camera that pauses as you pass the sign calls for a frame trace. Both can happen at the same corner, with different fixes.
What an average frame rate hides
Suppose a capture contains 30 displayed frames over exactly one second. One version spaces them evenly, about 33.3 milliseconds apart. Another has 28 intervals of about 33.3 milliseconds, one of about 16.7 and one of 50. Both produce essentially the same frames-per-second figure over the sample. In the second capture a car moving sideways at 600 screen pixels per second travels 30 pixels during the long interval. At the regular cadence it travels about 20. The driver sees a bigger jump at the junction even though the headline figure has barely moved. These are illustrative timings, independent of any console benchmark.
When a hitch repeats, compare the slow frame with the adjacent good ones. Start with what changed on screen. Did the vehicle round a corner, did a crowd enter view, or did the camera hold still while a new surface appeared? Then inspect the work on the path to presenting that frame. A texture can be ready in storage and still need time to reach graphics memory. An animation update can hold up a frame even when every visible texture is already there. Record which stage finished last before choosing between faster reads, fewer updates or less drawing work.
How far ahead should a moving map load?
Suppose you are building a dispatch map. An operator follows a van through a city, then zooms in to read the name of the street at its next turn. The blue route line should survive the transition even if the labels take longer. MapLibre’s engineering guide describes how to split large geographic datasets, serve them as vector tiles and vary the detail shown at different zoom levels. Roads and labels can then have different delivery and drawing costs. MapLibre’s guide to large datasets
Here is a sizing exercise you can repeat with your own numbers. Suppose the leading edge of the visible map moves at 800 screen pixels per second at a fixed zoom. A tile takes 300 milliseconds from the moment the application requests it until it can be drawn. The viewport advances 240 pixels during that wait. Fetching the tile when its edge first appears means it is already late.
| Assumed time until the tile can be drawn | Distance the view travels at 800 px/s |
|---|---|
| 150 ms | 120 px |
| 300 ms | 240 px |
| 600 ms | 480 px |
These are example inputs for a web map. Estimate your own look-ahead distance by multiplying pan speed in pixels per second by time to readiness in seconds. At 300 ms, starting 240 pixels ahead leaves zero margin. Try the request slightly earlier and measure how many speculative tiles the operator actually sees. A quick reversal can make a whole row of prefetched tiles useless while the tiles behind the van become urgent again.
When loading farther ahead makes the map slower
Look-ahead only buys time if the application can finish work as quickly as movement creates it. Imagine that a pan reveals four new tiles every quarter of a second. That is 16 tiles per second. If a worker can prepare 12 tiles per second on the target phone, the queue grows by four every second of sustained movement. After three seconds it has accumulated 12 tiles, a full second of work at that rate. Starting earlier can hide a short burst. Extending the route eventually exposes the deficit again.
The choice of fix depends on what is waiting. If the worker spends most of its time preparing features that are invisible at the current zoom, simplify the data sent at that zoom. If the application keeps preparing tiles for a street the operator has already left, drop obsolete work before it enters the expensive stage. If the route line must stay visible while building footprints can arrive later, give those layers independent deadlines where the data format permits it. MapLibre's guidance on simplifying geometries and restricting the zoom levels at which features appear gives a team concrete changes to test. MapLibre's large-data guide
A useful counterexample is a driver doubling back. A loader that immediately evicts everything behind the vehicle may request the same intersection twice in a few seconds. Keeping the previous block can make that turn cheap, but it consumes memory otherwise available ahead. Replay a straight journey and a U-turn under the same memory limit. Count tiles prepared, tiles shown and tiles prepared a second time. If a new policy improves the straight run by a few milliseconds and leaves the U-turn blank for half a second, those three counts help explain where the gain came from.
Test the turn where detail disappears
- Replay one route. Pick a fixed zoom, drag across the same junction, pause, then reverse. Run it with an empty cache and again with the tiles cached. Use the slowest device you support.
- Time a tile through to its first appearance. Record when it is requested, when its bytes arrive and when its features become visible. A long wait for the response points towards transfer or server work. A larger gap after arrival is a reason to inspect decoding, label placement and drawing.
- Find the expensive frame. In Chrome DevTools, record the pan and line up the Frames and Network tracks. A 60 Hz display presents a new refresh roughly every 16.7 ms, and the browser needs some of that time for its own work. If the tiles arrived early but the pan still catches, inspect the main-thread activity during that frame before changing your preload distance. Chrome DevTools Performance reference
Read a trace before changing the loader
Consider two hypothetical captures of the same missing street name. In the first, the operator reaches the turn at 400 ms. The request starts at 310 ms, the response finishes at 450 ms and the label appears at 475 ms. Preparation after delivery took only 25 ms. Starting the same request 100 ms earlier, with the same transfer and preparation times, would put the label on screen at 375 ms. That is a plausible scheduling fix to test against another cold run.
In the second capture the request starts at 20 ms and the response finishes at 120 ms, yet the label still appears at 475 ms. There are 355 ms between delivery and first appearance. Inspect the worker queue, label placement and the frame in which the map incorporated the result. Also check whether the style intentionally suppresses overlapping labels. A street name withheld by a collision rule could remain absent even if every tile were in memory before the pan began. Keep the map style and zoom fixed when comparing the two runs.
Report the result in terms an operator would recognise. For example, record how many of 20 repeated turns showed the street name before the van reached it, alongside the longest late arrival and the device tested. Keep first visits and repeat visits in separate rows. Eight successful warm runs say little about the person opening an unfamiliar district for the first time. The trace explains the delay. That route-level count tells the team whether the change actually helped someone finish the task.
For the dispatch operator, a practical success condition is simple. At the quickest supported pan on the target phone, the route stays visible through the turn and the next street name is legible by the time the van reaches it. If a map in your product misses that turn, bring a recording of the route and the device it happened on. Book a free consultation.
Frequently asked questions
Does open-world streaming require an internet connection?
Asset streaming describes when an application brings data into memory. An installed game can read assets from local storage. A browser map can fetch its tiles over the network or reuse ones it already has in its cache.
Why can objects pop into detail even when the frame rate looks steady?
A renderer can keep drawing a simpler version on time and switch when more detail is available. A late switch can also come from the distance threshold chosen by the developer. Logging both the asset’s readiness and the moment of the switch helps distinguish those causes.
How does speeding up storage help a game like GTA 6?
Sony says GTA 6 uses the PS5’s SSD to shorten loading times. Quicker reads get the asset’s bytes to the engine sooner. The engine still needs memory to hold the asset and enough processing time to prepare and draw it.
Sources
- Rockstar Games’ launch announcement, published by Take-Two
- Restart’s interview with Rockstar’s technology and design leads
- TechRadar’s questions to Rockstar about the 30 fps target
- PlayStation’s account of the PS5 hardware features used by GTA 6
- MapLibre GL JS documentation on large geographic datasets
- Google’s guide to rendering performance and display refresh rates
- Chrome DevTools documentation on frame and network traces
- Microsoft's BC7 texture format
- Microsoft's DirectX sampler feedback example
- Epic's World Partition streaming documentation
Originally published on znovastudio.com.
Top comments (0)