Compare
· Qwen3.6-35B-A3B · llama.cpp
One GPU or a split?
Same 35B GGUF , llama.cpp Vulkan , f16 KV . Left is one R9700 ; right is a split across both cards.
On this sweep, one R9700 usually decodes faster than a split of the same 35B. Split still cuts VRAM per card.
Two full copies are a different layout —
35B on two cards .
When a second card helps:
Dual / multi-GPU .
Hands-on llama.cpp flags:
/setup/multi-gpu .
Daily driver · 35B llama.cpp · 262K
Measured peak system RAM on the local desktop : 42.9 GB (Linux GiB labeled GB). Measured, pending re-verify after RAM fix.
Breakdown
Offloaded weights ·
0 GiB
Separable embed in RAM ·
0 GiB
KV in RAM at selected ·
0 GiB
KV in RAM at full context ·
0 GiB
Runtime / backend ·
0.85 GiB
OS / desktop baseline ·
8.00 GiB
Headroom ·
16.0 GiB
Measured verdict
For the 35B (UD-Q5_K_XL), one card is faster at decode. A split only wins on decode for Q6_K past short context.
Winner counts from the measured cards: one GPU 12 · split 4 · tie 0. The split prefills long prompts faster (+14% at 128K and +29% at 256K, from the UD-Q5_K_XL cards). A second card pays off for Flash-class models, for running two models at once, and for going beyond 262K context.
Counts and percents are derived from the 35B llama.cpp cards on this page, not typed in by hand.
Winner counts on this sweep: one GPU 12 · split 4 · tie 0 · incomplete 0.
Q6_K 1-GPU 4K uses the 23:55 card (103.23 tok/s). The earlier 82.62 card stays on /results.
Overnight sweep, 2026-10-12
Measured, pending re-verify after RAM fix. DDR5-5200 later proved unstable; a kernel oops at 00:14 CT aborted the sweep. All 33 cards predate the oops. Each card already notes one transient bad read in earlier sha256 passes.
Flash / Strata , two cards
Measured two-card Strata Flash
Flash-class files are the job a second card is for. These rows already exist in the repo — they are not a 35B split.
Q3 stays a labeled estimate on this page.
Two-card Flash on Strata from official run cards and the lab log. Empty stays empty.
Quant
Context
Decode
Source
UD-Q6_K_XL
8K
32 tok/s
Measured on our rig, 2026-10. Two-card Flash Q6, Strata, DDR5-3600, 32.0 tok/s at 8K. Pair with 32.8 at DDR5-5200.
UD-Q6_K_XL
8K
32.8 tok/s
Measured on our rig, 2026-10. Two-card Flash Q6 at 262K load. Strata 32.8 tok/s at 8K. DDR5-5200.
UD-Q6_K_XL
32K
32.8 tok/s
Measured on our rig, 2026-10. Two-card Flash Q6 at 262K load. Strata 32.8 tok/s at 32K.
UD-Q6_K_XL
100K
31.8 tok/s
Measured on our rig, 2026-10. Two-card Flash Q6 at 262K load. Strata 31.8 tok/s at 100K.
UD-Q6_K_XL
256K
32.6 tok/s
Measured on our rig, 2026-10. Prior first-party Flash Q6: Strata ~32.6 tok/s. Card count was not published; recorded on the two-R9700 host. A different run from the 32.8 / 31.8 two-card points.
UD-Q5_K_XL
short
65 tok/s
Strata UD-Q5_K_XL decode, code
UD-Q5_K_XL
short
52 tok/s
Strata UD-Q5_K_XL decode, short story
UD-Q5_K_XL
128K
47.6-54.3 tok/s
Strata UD-Q5_K_XL decode at 131K
Empty
Nothing matches that filter
We will not invent a row. Clear a chip or open the run cards.
UD-Q5_K_XL · short
One R9700 decodes faster by 19.61 tok/s.
1× prefill
2761.6 tok/s
Split prefill
2637.3 tok/s
1× VRAM / total
25.45 GB · 25.97 GB
Split VRAM / total
13.28 / 13.11 GB · 26.39 GB
1× system RAM
37.32 GB
Split system RAM
37.6 GB
UD-Q5_K_XL · 32K
One R9700 decodes faster by 15.72 tok/s.
1× prefill
1937.9 tok/s
Split prefill
1916 tok/s
1× VRAM / total
26.16 GB · 26.7 GB
Split VRAM / total
13.77 / 13.64 GB · 27.41 GB
1× system RAM
37.83 GB
Split system RAM
38.16 GB
UD-Q5_K_XL · 128K
One R9700 decodes faster by 8.49 tok/s.
1× prefill
942.8 tok/s
Split prefill
1071.1 tok/s
1× VRAM / total
28.22 GB · 28.82 GB
Split VRAM / total
15.47 / 15.48 GB · 30.94 GB
1× system RAM
40.18 GB
Split system RAM
41.02 GB
UD-Q5_K_XL · 256K
One R9700 decodes faster by 2.37 tok/s.
1× prefill
519.2 tok/s
Split prefill
671.7 tok/s
1× VRAM / total
28.73 GB · 29.34 GB
Split VRAM / total
17.8 / 17.99 GB · 35.74 GB
1× system RAM
42.9 GB
Split system RAM
44.71 GB
Q5_K_M · short
One R9700 decodes faster by 21.54 tok/s.
1× prefill
2710.8 tok/s
Split prefill
2647.2 tok/s
1× VRAM / total
24.16 GB · 24.68 GB
Split VRAM / total
12.71 / 12.41 GB · 25.13 GB
1× system RAM
35.83 GB
Split system RAM
37.17 GB
Q5_K_M · 32K
One R9700 decodes faster by 17.30 tok/s.
1× prefill
1902.6 tok/s
Split prefill
1920.9 tok/s
1× VRAM / total
24.87 GB · 25.39 GB
Split VRAM / total
13.2 / 12.95 GB · 26.13 GB
1× system RAM
36.46 GB
Split system RAM
37.8 GB
Q5_K_M · 128K
One R9700 decodes faster by 9.78 tok/s.
1× prefill
940.3 tok/s
Split prefill
1067.3 tok/s
1× VRAM / total
26.93 GB · 27.43 GB
Split VRAM / total
14.98 / 14.79 GB · 29.76 GB
1× system RAM
38.65 GB
Split system RAM
40.73 GB
Q5_K_M · 256K
Split across both cards decodes faster by 1.79 tok/s.
1× prefill
541.7 tok/s
Split prefill
664.6 tok/s
1× VRAM / total
28.56 GB · 29.15 GB
Split VRAM / total
17.45 / 17.3 GB · 34.54 GB
1× system RAM
42.79 GB
Split system RAM
46.21 GB
Q4_K_M · short
One R9700 decodes faster by 24.41 tok/s.
1× prefill
2733.9 tok/s
Split prefill
2660.4 tok/s
1× VRAM / total
20.84 GB · 21.43 GB
Split VRAM / total
10.83 / 10.85 GB · 21.69 GB
1× system RAM
33.74 GB
Split system RAM
33.74 GB
Q4_K_M · 32K
One R9700 decodes faster by 19.31 tok/s.
1× prefill
1902.1 tok/s
Split prefill
1933.1 tok/s
1× VRAM / total
21.55 GB · 22.17 GB
Split VRAM / total
11.34 / 11.39 GB · 22.72 GB
1× system RAM
34.58 GB
Split system RAM
34.39 GB
Q4_K_M · 128K
One R9700 decodes faster by 10.36 tok/s.
1× prefill
942.1 tok/s
Split prefill
1078.4 tok/s
1× VRAM / total
23.61 GB · 24.27 GB
Split VRAM / total
13.03 / 13.23 GB · 26.26 GB
1× system RAM
36.5 GB
Split system RAM
38.04 GB
Q4_K_M · 256K
One R9700 decodes faster by 4.89 tok/s.
1× prefill
558.5 tok/s
Split prefill
673 tok/s
1× VRAM / total
26.36 GB · 27.03 GB
Split VRAM / total
15.28 / 15.74 GB · 31.02 GB
1× system RAM
39.05 GB
Split system RAM
40.84 GB
Q6_K · short
One R9700 decodes faster by 19.73 tok/s.
1× prefill
2623.6 tok/s
Split prefill
2572.1 tok/s
1× VRAM / total
28.78 GB · 29.26 GB
Split VRAM / total
15.01 / 14.63 GB · 29.64 GB
1× system RAM
40.15 GB
Split system RAM
38.22 GB
Q6_K · 32K
Split across both cards decodes faster by 4.24 tok/s.
Partly offloaded to DDR5 on 1 GPU
1× prefill
1683 tok/s
Split prefill
1890.7 tok/s
1× VRAM / total
28.68 GB · 29.17 GB
Split VRAM / total
15.5 / 15.16 GB · 30.66 GB
1× system RAM
44.01 GB
Split system RAM
38.98 GB
Q6_K · 128K
Split across both cards decodes faster by 3.53 tok/s.
Partly offloaded to DDR5 on 1 GPU
1× prefill
800.3 tok/s
Split prefill
1059.7 tok/s
1× VRAM / total
28.78 GB · 29.25 GB
Split VRAM / total
17.21 / 17 GB · 34.21 GB
1× system RAM
45.43 GB
Split system RAM
41.41 GB
Q6_K · 256K
Split across both cards decodes faster by 4.42 tok/s.
Partly offloaded to DDR5 on 1 GPU
1× prefill
471.6 tok/s
Split prefill
666.2 tok/s
1× VRAM / total
28.81 GB · 29.3 GB
Split VRAM / total
19.46 / 19.51 GB · 38.96 GB
1× system RAM
43.69 GB
Split system RAM
45.15 GB
Q3 estimate
UD-Q3_K_XL
Q6 / Q5 / Q4 system RAM above is from local desktop cards.
UD-Q3_K_XL is a labeled estimate (Hugging Face listing for Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf (16.8 GB)) — not a local desktop run.
Decode and prefill stay not measured. Full Q3 column:
Q6 / Q5 / Q4 + Q3 estimate .
Minimum system RAM needed
32 GB kit
Estimate · 24.6 GiB raw · 4K context
RAM needed at full context
32 GB kit
Estimate · 24.7 GiB raw · 256K full context
Estimate. Offloaded weights + KV in RAM + runtime + OS/desktop + headroom, rounded up to a 32 GB DIMM kit.
Breakdown
Offloaded weights ·
0 GiB
Separable embed in RAM ·
0 GiB
KV in RAM at selected ·
0 GiB
KV in RAM at full context ·
0 GiB
Runtime / backend ·
0.64 GiB
OS / desktop baseline ·
8.00 GiB
Headroom ·
16.0 GiB
Card sheet:
R9700 measured page
·
Quant picker:
Q6 vs Q5 vs Q4
·
run cards
·
calculator
AMDai Lab Notes digest
Email only — no account. Double opt-in. About once a week, and only when run cards, guides, ROCm problems, or Stack Lab actually changed. Not a daily drip.
Issues may include one labeled Sponsored mention.
Slots and rates:
Sponsor .