Computed

Gemma 3 1B Instruct on a Jetson AGX Orin (64GB)

Yes — the unquantized weights fit with room for 32k tokens of context.

Change the numbers →Gemma 3 1B Instruct on every card Jetson AGX Orin (64GB) across every model

Verdict

Fits at FP16

48.0 GiB usable of 64 GB

Weights (Q4_K_M)

576 MiB

1.9 GiB at FP16 · 1000M params · 1 of 6 quant rows are measured files

KV cache

26.0 KiB per token at FP16

26 layers × 1 KV heads × 256 dimensions

Speed ceiling

100 tok/s

204.8 GB/s ÷ bytes read per token

Every quant, against this card

Weight bytes are the size of the real published file wherever one exists — 1 of these6 rows are measured from the published checkpoint, the rest computed from the parameter count. Max context is what the KV cache can grow to in whatever memory the weights leave behind, capped at the 32k tokens this model was trained to address.

Bars are the weight bytes at each precision; the dashed line is the usable memory of a Jetson AGX Orin (64GB). Drawn from the computed byte counts, not sketched.

Precision	Bits/weight	Weights	Source	Fits	Max context	Ceiling
FP16 / BF16	16.00	1.9 GiB	measured file	yes	32,768 (model cap)	100 tok/s
Q8_0	8.50	1013 MiB	nominal	yes	32,768 (model cap)	185 tok/s
Q6_K	6.56	782 MiB	nominal	yes	32,768 (model cap)	237 tok/s
Q5_K_M	5.67	676 MiB	nominal	yes	32,768 (model cap)	272 tok/s
Q4_K_M	4.83	576 MiB	nominal	yes	32,768 (model cap)	316 tok/s
Q3_K_M	3.91	466 MiB	nominal	yes	32,768 (model cap)	384 tok/s

What the context actually costs

The KV cache is 2 · layers · kv_heads · head_dim bytes per token per element — 2 · 26 · 1 · 256 · 2 B = 26.0 KiB. It depends on the key/value head count, not the attention-head count and not the parameter count. This model shares 1 KV head across 4 query heads, which divides the cache by 4 against multi-head attention.

Sliding-window attention. 22 of this model's 26 layers attend to a 512-token window and stop growing there; only the remaining 4 keep scaling with context. That is why the cache figures below flatten out — and why a calculator that ignores the layer pattern over-states this model's long-context footprint several times over.

Context	KV cache (FP16)	KV cache (8-bit)	Plus Q4_K_M weights
4,096	27 MiB	14 MiB	603 MiB
8,192	43 MiB	22 MiB	619 MiB
32,768	139 MiB	70 MiB	715 MiB

The speed ceiling, and where it comes from

Generating one token reads every active weight from memory once. At FP16 / BF16 that is 1.9 GiB, plus a pass over the KV cache. A Jetson AGX Orin (64GB) moves 204.8 GB/s, so the arithmetic ceiling is 100 tok/s. Treat it as a bound, not an estimate: attention overhead, kernel launches and imperfect memory access keep real runtimes at roughly 60–80% of it, and nothing pushes past it.

Where these numbers come from

The model

Parameters: 999,885,952
Layers: 26
Attention / KV heads: 4 / 1
Head dimension: 256
Trained context: 32,768
Checkpoint as published: 1.9 GiB

Read from unsloth/gemma-3-1b-it — an ungated mirror of google/gemma-3-1b-it, whose own config cannot be fetched without an access token. The parameter count is the Hub's own total over the tensor shapes, not a figure taken from the model's name.

The accelerator

Memory: 64 GB LPDDR5 unified
Bandwidth: 204.8 GB/s
Assumed usable: 75% → 48.0 GiB

Capacity and bandwidth from the vendor's specification. The usable fraction is an assumption, not a spec: unified memory is shared with the OS and the display, and the GPU working-set cap is raisable on Apple silicon with `sudo sysctl iogpu.wired_limit_mb`.

Questions this pairing answers

How much VRAM does Gemma 3 1B Instruct need?

1.9 GiB for the weights at FP16 — 1000M parameters at two bytes each — and 576 MiB at Q4_K_M. The KV cache is on top of that and is not a fixed number: this model spends 26.0 KiB per token of context on its full-attention layers, so 8,192 tokens costs a further 43 MiB. A Jetson AGX Orin (64GB) makes 48.0 GiB of its 64 GB available on the assumption below.

Can a Jetson AGX Orin (64GB) run Gemma 3 1B Instruct?

Yes — the unquantized weights fit with room for 32k tokens of context. That is the weights and the KV cache together, against 48.0 GiB of usable memory.

How fast will Gemma 3 1B Instruct run on a Jetson AGX Orin (64GB)?

No faster than 100 tokens/second at FP16 / BF16, and in practice below it. Decoding is memory-bound: every token reads all 1.9 GiB of weights plus the cache, and this card moves 204.8 GB/s. That division is the ceiling — no kernel, runtime or driver beats it, and a real runtime typically reaches 60–80% of it.

Why does the context length change how much memory Gemma 3 1B Instruct needs?

Because the KV cache holds one key and one value vector per token, per layer, for the whole conversation, and it is allocated separately from the weights. This model has 26 layers and 1 key/value head of 256 dimensions, shared across 4 query heads — grouped-query attention, which divides the cache by 4. That works out at 26.0 KiB per token, except that 22 of the 26 layers use a 512-token sliding window and stop growing there. Parameter count tells you nothing about this number.

The same model on a different card

A different model on the same card

All 62 models on Jetson AGX Orin (64GB) →

Method and limits. Weight bytes are the byte size of the real published file wherever one exists, and the model's exact parameter count times the published llama.cpp bits-per-weight where it does not. That distinction is on every row above and it matters at both ends: a sub-1B model's Q4_K_M file runs a third larger than the nominal figure because k-quants keep its embedding tables at higher precision, and an already-4-bit release cannot be quantized upward at all. The KV cache is 2 · layers · kv_heads · head_dim · bytes per token, summed over layers with each sliding-window layer capped at its window. The speed figure is a roofline bound, not a benchmark: bandwidth divided by bytes read per token, which no runtime exceeds and every runtime falls short of. The usable fraction of card memory is an assumption: 75% here, which is what this lane assumes for unified memory, where the OS and window server share the same pool — the other class assumes 92%, so it is not one number applied to every device. That is the only assumed input on this page; every other figure is computed from the model config and the card's published specification. Nothing on this page is written by a language model. Architecture from the model's published config, fetched 2026-08-07.