
Discussion of the AI car usually stops at TOPS. NVIDIA's own figures show why that misses the tighter constraint: from DRIVE Orin to DRIVE Thor, peak AI compute rises about 3.9 times, while peak memory bandwidth rises about 1.3 times.
This post builds a vehicle-level budget for 2026, 2028 and 2030. In our illustrative model, sustained DRAM demand grows about 4.8 times, from 132 GB/s to 631 GB/s, while installed peak bandwidth grows about 3.4 times, from 342 GB/s to 1,152 GB/s, even with LPDDR6 at its top rate. Every input is on the page for you to replace.
Key takeaways
- Vendor figures put Thor at up to 1,000 INT8 TOPS and 273 GB/s, against 254 TOPS and 205 GB/s for Orin. Both are peak claims.
- In our model, usable bandwidth runs short before capacity does. The 2030 driving domain uses 85% of usable bandwidth but only about 35% of installed capacity.
- Reasoning text is the line that can break the budget. A 40-token trace on every 10 Hz cycle from an 8.2 billion parameter backbone would need about 3,280 GB/s on its own, on a read-all-weights assumption.
- The cockpit deserves its own budget. A 7 billion parameter assistant on a 128-bit LPDDR5X-9600 bus takes the 2028 cockpit to 101% of usable bandwidth in our model, even though the vehicle total shows only 72%.
The figures we can anchor to
We could not read a primary automaker memory budget, so this model leans on chip maker disclosures and JEDEC workshop material. Treat vendor numbers as claims, and compute numbers as peak values.
| Item | Published figure | Status |
|---|---|---|
| NVIDIA DRIVE AGX Orin | Up to 254 INT8 TOPS, 205 GB/s LPDDR5, 32 GB on the developer kit | NVIDIA claim. Its Orin developer deck rounds bandwidth to "up to 200 GB/s". The width and rate work out at 204.8 GB/s if the bus is 256 bits, which is our assumption |
| NVIDIA DRIVE AGX Thor | Up to 1,000 INT8 TFLOPS, 273 GB/s LPDDR5X, 64 GB on the developer kit | NVIDIA claim, on the developer kit |
| Hyperion 10 reference platform | Dual Thor, 14 cameras, nine radars, one lidar, 12 ultrasonics | NVIDIA reference design |
| Samsung automotive LPDDR5X | Up to 9,600 Mb/s, and 307.2 GB/s on a 256-bit bus | Samsung claim, with ASIL-D described as expected |
| LPDDR6, per a Samsung talk at the JEDEC workshop | Up to 14.4 Gb/s, two 12-DQ sub-channels, 288 bits per burst carrying 256 data bits. 153.6 GB/s on a 96-bit bus | Workshop material. We could not read the paid standard, JESD209-6 |
The 288-bit burst matters when you compare generations. LPDDR6 moves 16 metadata bits and 16 DBI or link ECC bits in-band, so only 8/9 of the raw rate is payload. At 14.4 Gb/s, 24 DQ gives 43.2 GB/s raw and 38.4 GB/s of payload. Twelve such channels use 288 pins, the same count as sixteen LPDDR5X channels with their 16 DQ and 2 mask pins each. That gives 460.8 GB/s of payload against 307.2 GB/s at 9,600 Mb/s, or 1.5 times, for the same data and mask pin count. That is our arithmetic, and no vendor has announced this configuration.
Automotive LPDDR6 is still early. The Korea Herald reported in July 2026 that SK hynix showed an automotive part at Mobile World Congress in March. We have not confirmed that from a primary source, and we found no published automotive speed grade.
How the budget works
Four rules drive almost every line:
- Sensors and displays. Pixels times bytes per pixel times frame rate, multiplied by the number of times each frame touches DRAM on its way through capture, ISP, scaling and the network.
- Neural networks. Weight bytes times inference rate, plus feature traffic. Weights too large for on-chip memory have to come from DRAM on each pass.
- Token generation. Tokens per second times bytes read per token. Arm's llama-roofline guide gives the same relation and measures it: 13.8 GB/s divided by 346 MB per token caps a Raspberry Pi 5 at about 40 tokens per second, however fast the cores are.
- Usable supply. Installed peak times a derating factor for refresh, bank conflicts, bus turnarounds and error protection. We use 0.70 for scheduling efficiency and 0.90 for inline ECC, giving 0.63. Both are illustrative. Synopsys explains the ECC cost: each access needs extra commands for the check bits, and the penalty depends on traffic pattern.
The inputs
Everything in this table is an illustrative assumption about a reference vehicle unless the row says otherwise. It does not describe any real car or supplier plan. We used no market forecasts.
| Input | 2026 | 2028 | 2030 |
|---|---|---|---|
| Cameras at 12 bits per pixel (1.5 B), 30 fps, 4 DRAM touches | 12 at 5 MP | 14 at 8 MP (count from Hyperion 10) | 16 at 8 MP |
| Driving model weights, INT8, at 30 Hz | 1 GB | 3 GB | 4 GB |
| Reasoning and action model at 10 Hz | None | None | 10.5 GB (parameter count from NVIDIA's model card, INT8 is our assumption) |
| Feature traffic as a multiple of weight traffic | 1 | 1 | 1 |
| Cockpit assistant traffic while speaking | 30 GB/s | 70 GB/s (Samsung slide: 7B model, 10 to 20 tokens/s) | 100 GB/s |
| Display pixels at 4 B, 60 Hz, 5 touches | 10 MP | 15 MP | 20 MP |
| Driving domain memory, installed peak | One Orin class, 205 GB/s | Two Thor class, 546 GB/s | Two 288-pin LPDDR6 at 14.4 Gb/s, 921.6 GB/s |
| Cockpit memory, installed peak | 128-bit at 8,533 Mb/s, 136.5 GB/s | 128-bit at 9,600 Mb/s, 153.6 GB/s | 144-pin LPDDR6 at 14.4 Gb/s, 230.4 GB/s |
Worked lines, so you can check them. The 2028 camera path is 14 × 8 MP × 1.5 B × 30 fps = 5.04 GB/s, and four touches make 20.16 GB/s. For the link-level view, our piece on how automotive cameras and sensor fusion evolved works through camera data rates. The 2028 driving model is 3 GB × 30 Hz × 2 = 180 GB/s, with the factor of 2 being weights plus features. The 2030 reasoning and action model is 10.5 GB × 10 Hz × 2 = 210 GB/s. Cockpit graphics in 2028 are 15 MP × 4 B × 60 Hz × 5 = 18 GB/s.
The budget
| GB/s, sustained | 2026 | 2028 | 2030 |
|---|---|---|---|
| Camera path | 10.8 | 20.2 | 23.0 |
| Radar, lidar, IMU | 1.0 | 2.0 | 2.0 |
| Driving model | 60.0 | 180.0 | 240.0 |
| Reasoning and action model, no text | 0 | 0 | 210.0 |
| Planning, OS, logging | 10.0 | 15.0 | 20.0 |
| Driving domain demand | 81.8 | 217.2 | 495.0 |
| Driving domain usable (peak × 0.63) | 129.2 | 344.0 | 580.6 |
| Share of usable | 63% | 63% | 85% |
| Cockpit assistant | 30.0 | 70.0 | 100.0 |
| Cockpit graphics | 12.0 | 18.0 | 24.0 |
| Cockpit CPU and apps | 8.0 | 10.0 | 12.0 |
| Cockpit demand | 50.0 | 98.0 | 136.0 |
| Cockpit usable (peak × 0.63) | 86.0 | 96.8 | 145.2 |
| Share of usable | 58% | 101% | 94% |
| Vehicle demand | 131.8 | 315.2 | 631.0 |
| Vehicle installed peak | 341.5 | 699.6 | 1,152.0 |
| Vehicle usable | 215.1 | 440.7 | 725.8 |
| Share of usable | 61% | 72% | 87% |
Three things stand out. First, cameras are small: 23 GB/s in 2030 is under 5% of driving domain demand. What the network does with those frames is what stresses DRAM. Second, 2028 is the pinch for the cockpit, because the assistant arrives before LPDDR6 does. Assistant traffic is bursty, so a real design might shed graphics or slow token output during speech. Third, the vehicle row hides this completely: 72% looks comfortable.
Test the feature multiple and derating. At 0.55 derating, the 2030 driving domain reaches 98% of usable. At 0.75 it falls to 72%. Doubling feature traffic to a multiple of 2 adds 225 GB/s and takes it to 124%. If automotive LPDDR6 only reaches 12.8 Gb/s, the data rate in the title of Samsung's ISSCC 2026 paper, the installed peak falls to 819.2 GB/s and the share rises to 96%.
As an outside check, an arXiv preprint from Peking University and Shanghai Jiao Tong University (June 2026) models a Level 4 stack on 128-bit LPDDR5 chips with 102 GB/s each. It needs 3 to 5 chips for its 180 to 300 TMAC/s FP16 workload, so 306 to 510 GB/s is installed. Image backbones and multi-camera fusion take 69.7% of average bandwidth, and fusion alone peaks at 280.2 GB/s. Its chip count is set by compute on its own design, so this is provisioned bandwidth, not a measured requirement. Even so, our 217 GB/s for 2028 is the same order of magnitude. The 280.2 GB/s peak also exceeds one Thor's 273 GB/s, so peak bursts need a plan separate from this average budget.
The expensive line: reasoning text
NVIDIA's Alpamayo 1 is a useful public reference. The model card lists an 8.2 billion parameter backbone and a 2.3 billion parameter action expert, stored in BF16, with a 24 GB GPU minimum. At BF16, 10.5 billion parameters is 21 GB, so one read of the weights at 10 Hz is 210 GB/s, or 77% of one Thor's peak, before any feature traffic. The card offers the model for non-commercial use, not as a production stack, and does not describe an INT8 version. Our INT8 row is a planning assumption.
The paper's runtime table shows where time goes. On an RTX 6000 Pro Blackwell workstation GPU, the trajectory-only baseline takes 29 ms. Adding a 40-token reasoning trace takes it to 99 ms, with 70 ms, or 71% of the total, spent decoding that text. The paper lists reasoning on demand as future work, since the current design traces every input.
In a memory budget, that choice is large. Autoregressive decoding reads the weights once per token, as Arm's guide describes. Here is the 2030 driving domain, taking the 495.0 GB/s base and 580.6 GB/s usable from above. We assume all 8.2 GB of INT8 backbone is read per token, which overstates things because the vision encoder is not re-read. It also ignores the key-value cache, which works the other way.
| Reasoning policy (illustrative) | Added GB/s | Domain demand | Share of usable |
|---|---|---|---|
| No text in the vehicle | 0 | 495.0 | 85% |
| 40 tokens every 10 seconds | 32.8 | 527.8 | 91% |
| 40 tokens every second | 328.0 | 823.0 | 142% |
| 40 tokens every 10 Hz cycle | 3,280.0 | 3,775.0 | 650% |
| Every cycle, on a 0.5B backbone with the same action expert | 200.0 | 541.0 | 93% |
Shrinking the backbone cuts the reasoning and action line from 210 GB/s to 56 GB/s (2.8 GB × 10 Hz × 2), and the text costs 0.5 GB × 400 tokens/s = 200 GB/s. NVIDIA's public benchmark gives a sense of the price: the 10B model reaches a minADE of 0.849 m at 6.4 s against 0.913 m for the 0.5B version, about 7.0% better. Roughly a sixteenth of the backbone gives up about 7% on that open-loop metric. Whether that is a good trade is a safety question this budget cannot answer.
Our recommendation: treat text generation as a scheduled resource in the memory plan.
Capacity and compute are not the first wall
Capacity comes out easier. This table is illustrative, apart from the installed 32 GB and 64 GB developer kit sizes and the parameter count, and it assumes 15 buffered frames per camera.
| Driving domain, GB | 2026 | 2028 | 2030 |
|---|---|---|---|
| Model weights | 1.00 | 3.00 | 14.50 |
| Features and key-value buffers | 4.00 | 8.00 | 12.00 |
| Camera frame buffers | 1.35 | 2.52 | 2.88 |
| OS, middleware, maps, logging | 8.00 | 12.00 | 16.00 |
| Total | 14.35 | 25.52 | 45.38 |
| Installed | 32 | 128 | 128 |
Keeping the full model in BF16 lifts 2030 to 55.88 GB, still within 128 GB. Production memory sizes may differ from developer kits. Our note on the AI memory squeeze and DRAM allocation explains why automotive buyers should watch allocation as well as price.
Compute looks easier still. Thor's 1,000 INT8 TOPS gives 273 GB/s ÷ 1,000 TOPS = 0.27 bytes per operation at peak, against 205 ÷ 254 = 0.81 for Orin. Engines can only stay busy at that ratio by reusing data on-chip, which token decode does poorly. The preprint's 180 to 300 TMAC/s sits well below two Thors' 2,000 TOPS claim, though FP16 MACs and INT8 operations are not like for like. We assume no 2030 compute figure, because we found none we could read. An August 2026 review preprint takes the same broad view for large language models: memory, not arithmetic, is the main limit.
What to do with this
- Budget per memory domain, because bandwidth does not pool across chips. Zonal architecture concentrates software in a few central computers, so a shortfall in one domain has few places to go.
- Ask for sustained figures. Ask suppliers for measured bandwidth under your traffic mix with safety mechanisms on, including inline ECC, and put that in place of our 0.63.
- Write down the reasoning policy. How many tokens, how often, on which backbone? That one decision moved the table above from 85% to 650%.
- Keep LPDDR6 pin planning open. Compare payload rather than raw rate, because of the 8/9 factor.
- Separate average and peak. This budget is average traffic. Deadlines depend on bursts, so queue analysis has to follow it, and your latency analysis sets the final margin.
Decisions like these get locked early. Our SoC lifecycle on The Semiconductor Hub shows where they fall in a programme.
Why not simply use GDDR or HBM?
This budget does not rule them out. Samsung's article notes that other interfaces can offer higher I/O bandwidth, typically at a higher cost per bit. A fair comparison needs temperature grades, qualification and cost data, which we have not modelled.
Sources
- NVIDIA, "Accelerate Autonomous Vehicle Development with the NVIDIA DRIVE AGX Thor Developer Kit", NVIDIA Technical Blog, 3 September 2025.
- NVIDIA, DRIVE AGX Orin platform for developers (slide deck, PDF dated 15 October 2025).
- NVIDIA, "NVIDIA Makes the World Robotaxi-Ready With Uber Partnership to Support Global Expansion", 28 October 2025, and the NVIDIA DRIVE Hyperion product page.
- NVIDIA, Alpamayo-R1-10B model card, Hugging Face, released 3 December 2025.
- Wang et al. (NVIDIA), "Alpamayo-R1: Bridging Reasoning and Action Prediction for Generalizable Autonomous Driving in the Long Tail", arXiv:2511.00088, v2, 7 January 2026.
- Seunghyun Moon (Samsung), "LPDDR6 Key Architecture", JEDEC LPDDR6 Workshop, 8 April 2025.
- JEDEC, "JEDEC Releases New LPDDR6 Standard to Enhance Mobile and AI Memory Performance", press release on JESD209-6, 9 July 2025.
- Kim et al. (Samsung), "A 16Gb 12.8Gb/s LPDDR6 SDRAM...", ISSCC, February 2026. Title only.
- Samsung Semiconductor, "Samsung's 12nm-Class Automotive LPDDR5X: DRAM for Safety-Critical Centralized Automotive Systems", tech blog, undated.
- Zhang et al., "Isolation-aware Scheduling Framework for DNN-based End-to-End Autonomous Driving System on Tile-based Accelerators", arXiv:2606.10303, 9 June 2026. Preprint.
- Arm, "Understand why token generation is limited by memory bandwidth", Learning Path on llama-roofline, undated.
- Patel and Singh, "AI Hardware Accelerators for Large Language Models: Architectures and the Memory Wall", arXiv:2608.28048, 28 August 2026. Preprint, abstract read.
- Vadhiraj Sankaranarayanan (Synopsys), "Error Correction Code (ECC) in DDR Memories", 19 October 2020.
- The Korea Herald, "Cars become the next memory chip battleground", 15 July 2026. Press report, not confirmed from a primary source.
The budget uses illustrative inputs for a reference vehicle unless a row names a source. Vendor figures are the companies' own claims as published on the dates shown, and arXiv preprints may not have completed peer review.

Your comments will be moderated before it appears here.