Home / Builds / B70 vs 5080

Run cards

B70 vs 5080 — run cards

2026-09-21 · rumble · dual-GPU · Experience→Silicon

  • 2026-09-21
  • two-box run
  • 30 run cards
  • mandatory columns

Lane × GPU scan matrix on top. Evidence = full run cards below (schema SoT). Numbers from Neo — unknowns stay unknown/n/a. No invented tok/s.

Different weight sets — B70 rows are GGUF via llama.cpp SYCL; 5080 rows are Ollama tags on CUDA. Not an A/B of the same GGUF.

Lane × GPU matrix

Scannable pass/fail/blocked. Hover lane labels for definitions. Each chip links the first matching run card. Describe ≠ image-gen ≠ video-gen.

Lane by GPU outcome matrix
LaneB70 SYCL5080 CUDA
textpass llama-climixed Ollama
describefail llama-mtmd-clipass Ollama
image-genblocked n/a — no SYCL Comfy pathblocked n/a — ComfyUI/weights not installed
video-genblocked n/a — no SYCL Comfy pathblocked n/a — ComfyUI/weights not installed

Image-gen / video-gen stay blocked (not installed). Logo-card keepers are side/background chrome only — not inside matrix cells. No invented tok/s. Locked hexes. Art: media-create@673d219 mascots/lanes-bot/keepers/.

Hardware / backends

  • B70: Intel Arc Pro B70 32GB · llama.cpp SYCL · Serving llama-cli / llama-mtmd-cli
  • 5080: RTX 5080 16GB · Ollama CUDA
  • No LiteLLM alias retargets; no purges; no gen installs

llama.cpp CLI (SYCL)

Same text prompts as the Ollama (CUDA) table; hardware is a column. Tasks are prompts (hover headers).

Model Hardware bugfix tooljson refactor
qwen3-coder-30b-a3b B70 pass pass pass
gemma4-26b-a4b B70 pass pass pass
gpt-oss-20b B70 pass pass pass
qwen38-27b-iq4xs B70 pass pass pass

Serving: llama-cli · Engine: llama.cpp SYCL

Ollama (CUDA)

Same text prompts as the llama.cpp CLI (SYCL) table; hardware is a column. Tasks are prompts (hover headers).

Model Hardware bugfix tooljson refactor
qwen38-27b-iq4xs 5080 pass pass pass
gemma4-12b-agent-hi 5080 pass fail fail
gemma4-12b-vision-agent 5080 pass pass fail
llama32-3b 5080 pass pass pass

Serving: Ollama · Engine: llama.cpp CUDA

Full run cards (all columns)

Mandatory Experience→Silicon fields per llm-run-card-schema. Source: CSV · handoff MD. Weights shown without local machine paths — open the CSV for stack detail. When a perf set is present, lead with engine tok/s_gen (decode); wall_secs is wall clock for the quality task and only belongs with that perf set (+ outcome) — never as a lone speed claim. Unknowns stay n/a/unknown; columns are not dropped.

Scroll horizontally for Silicon columns (UI → Hardware → Ctx → tok/s → wall_secs → notes). Heading says all columns because the markup keeps them — use the scrollbar under the table.

B70 cards (15)

B70 run cards
run_id date host gpu lane model task result UI Harness Tools/MCP Gateway Serving Engine Weights Hardware Ctx MTP/spec load_s ttft_ms tok/s_gen tok/s_prefill concurrency custom_suite wall_secs notes
2026-09-21-phase-b-b70-qwen3-coder-30b-a3b-bugfix 2026-09-21 i9rumble B70 text qwen3-coder-30b-a3b bugfix pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-qwen3-coder-30b-a3b (UD-Q4_K_XL) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/bugfix; score=pass 13.1 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-qwen3-coder-30b-a3b-tooljson 2026-09-21 i9rumble B70 text qwen3-coder-30b-a3b tooljson pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-qwen3-coder-30b-a3b (UD-Q4_K_XL) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/tooljson; score=pass 12.9 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-qwen3-coder-30b-a3b-refactor 2026-09-21 i9rumble B70 text qwen3-coder-30b-a3b refactor pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-qwen3-coder-30b-a3b (UD-Q4_K_XL) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/refactor; score=pass 12.9 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-gemma4-26b-a4b-bugfix 2026-09-21 i9rumble B70 text gemma4-26b-a4b bugfix pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-gemma4-26b-a4b (UD-Q4_K_XL) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/bugfix; score=pass 13.2 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-gemma4-26b-a4b-tooljson 2026-09-21 i9rumble B70 text gemma4-26b-a4b tooljson pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-gemma4-26b-a4b (UD-Q4_K_XL) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/tooljson; score=pass 13.2 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-gemma4-26b-a4b-refactor 2026-09-21 i9rumble B70 text gemma4-26b-a4b refactor pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-gemma4-26b-a4b (UD-Q4_K_XL) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/refactor; score=pass 13.1 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-gpt-oss-20b-bugfix 2026-09-21 i9rumble B70 text gpt-oss-20b bugfix pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-gpt-oss-20b (F16) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/bugfix; score=pass 12.3 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-gpt-oss-20b-tooljson 2026-09-21 i9rumble B70 text gpt-oss-20b tooljson pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-gpt-oss-20b (F16) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/tooljson; score=pass 13.7 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-gpt-oss-20b-refactor 2026-09-21 i9rumble B70 text gpt-oss-20b refactor pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-gpt-oss-20b (F16) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/refactor; score=pass 13.2 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-qwen38-27b-iq4xs-bugfix 2026-09-21 i9rumble B70 text qwen38-27b-iq4xs bugfix pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-qwen38-27b (UD-IQ4_XS) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/bugfix; score=pass 12.9 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-qwen38-27b-iq4xs-tooljson 2026-09-21 i9rumble B70 text qwen38-27b-iq4xs tooljson pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-qwen38-27b (UD-IQ4_XS) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/tooljson; score=pass 14.5 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-qwen38-27b-iq4xs-refactor 2026-09-21 i9rumble B70 text qwen38-27b-iq4xs refactor pass CLI Neo run script (outs-20260921-085213) none none llama-cli llama.cpp SYCL unsloth-qwen38-27b (UD-IQ4_XS) B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown (included in wall secs) unknown unknown (quality-only row; see Phase A speed card for tg128) unknown (see Phase A pp512) 1 phase-b-text/refactor; score=pass 13 Quality pass/fail only. Do not treat wall secs as tok/s. Different GGUF set than 5080 Ollama tags.
2026-09-21-phase-b-b70-gemma4-26b-a4b-img_describe 2026-09-21 i9rumble B70 image_describe gemma4-26b-a4b img_describe fail CLI llama-mtmd-cli none none llama-mtmd-cli llama.cpp SYCL + mmproj gemma4-26b-a4b UD-Q4_K_XL + mmproj-BF16 B70 32GB SYCL; VRAM used=unknown allocated=unknown; filled=unknown off unknown n/a n/a n/a 1 phase-b-vision/img_describe; score=fail 0.3 SYCL abort / empty out (exit -1073740791 on earlier attempts; exit 1 / invalid -st on mtmd). Separate lane from text. Not image gen.
2026-09-21-B70-image_gen-blocked 2026-09-21 i9rumble B70 image_gen none smoke blocked n/a n/a n/a none n/a n/a — no SYCL Comfy path none installed B70 32GB n/a n/a n/a n/a n/a n/a n/a image_gen; score=blocked n/a BLOCKED: no ComfyUI/A1111; no Qwen-Image 2.1; no Wan/LTX/Hunyuan. Survey lean 5080 CUDA only if installed later.
2026-09-21-B70-video_gen-blocked 2026-09-21 i9rumble B70 video_gen none smoke blocked n/a n/a n/a none n/a n/a — no SYCL Comfy path none installed B70 32GB n/a n/a n/a n/a n/a n/a n/a video_gen; score=blocked n/a BLOCKED: no ComfyUI/A1111; no Qwen-Image 2.1; no Wan/LTX/Hunyuan. Survey lean 5080 CUDA only if installed later.

5080 cards (15)

5080 run cards
run_id date host gpu lane model task result UI Harness Tools/MCP Gateway Serving Engine Weights Hardware Ctx MTP/spec load_s ttft_ms tok/s_gen tok/s_prefill concurrency custom_suite wall_secs notes
2026-09-21-phase-c-5080-qwen38-27b-iq4xs-bugfix 2026-09-21 i9rumble 5080 text qwen38-27b-iq4xs bugfix pass CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) qwen3.8-27b-ud-iq4xs:latest RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/bugfix; score=pass 25.6 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-qwen38-27b-iq4xs-tooljson 2026-09-21 i9rumble 5080 text qwen38-27b-iq4xs tooljson pass CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) qwen3.8-27b-ud-iq4xs:latest RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/tooljson; score=pass 6.1 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-qwen38-27b-iq4xs-refactor 2026-09-21 i9rumble 5080 text qwen38-27b-iq4xs refactor pass CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) qwen3.8-27b-ud-iq4xs:latest RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/refactor; score=pass 4.6 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-gemma4-12b-agent-hi-bugfix 2026-09-21 i9rumble 5080 text gemma4-12b-agent-hi bugfix pass CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) gemma4-12b-agent-hi:latest RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/bugfix; score=pass 19.7 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-gemma4-12b-agent-hi-tooljson 2026-09-21 i9rumble 5080 text gemma4-12b-agent-hi tooljson fail CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) gemma4-12b-agent-hi:latest RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/tooljson; score=fail 4.8 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-gemma4-12b-agent-hi-refactor 2026-09-21 i9rumble 5080 text gemma4-12b-agent-hi refactor fail CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) gemma4-12b-agent-hi:latest RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/refactor; score=fail 17.7 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-gemma4-12b-vision-agent-bugfix 2026-09-21 i9rumble 5080 text gemma4-12b-vision-agent bugfix pass CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) gemma4-12b-vision-agent:latest RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/bugfix; score=pass 18.3 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-gemma4-12b-vision-agent-tooljson 2026-09-21 i9rumble 5080 text gemma4-12b-vision-agent tooljson pass CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) gemma4-12b-vision-agent:latest RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/tooljson; score=pass 4 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-gemma4-12b-vision-agent-refactor 2026-09-21 i9rumble 5080 text gemma4-12b-vision-agent refactor fail CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) gemma4-12b-vision-agent:latest RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/refactor; score=fail 4.9 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-llama32-3b-bugfix 2026-09-21 i9rumble 5080 text llama32-3b bugfix pass CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) llama3.2:3b RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/bugfix; score=pass 11.2 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-llama32-3b-tooljson 2026-09-21 i9rumble 5080 text llama32-3b tooljson pass CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) llama3.2:3b RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/tooljson; score=pass 0.5 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-llama32-3b-refactor 2026-09-21 i9rumble 5080 text llama32-3b refactor pass CLI Invoke-Bakeoff.ps1 /api/generate none none Ollama llama.cpp CUDA (Ollama) llama3.2:3b RTX 5080 16GB CUDA; VRAM used=unknown (not sampled per prompt) allocated=unknown (Ollama default/modelfile); filled=unknown off unknown (may be warm) unknown unknown (quality-only; wall_secs recorded) unknown 1 phase-c-text/refactor; score=pass 0.4 Ollama tags ≠ B70 GGUF set. Quality row only — do not invent tok/s.
2026-09-21-phase-c-5080-gemma4-12b-vision-agent-img_describe 2026-09-21 i9rumble 5080 image_describe gemma4-12b-vision-agent img_describe pass CLI Invoke-Bakeoff + /api/chat (thinking fallback) none none Ollama llama.cpp CUDA (Ollama) gemma4-12b-vision-agent:latest RTX 5080 16GB CUDA; VRAM used≈8.1GB resident at keep-alive sample (not exact peak) allocated=unknown; filled=unknown off unknown unknown unknown unknown 1 phase-c-vision/img_describe; score=pass 3.2 PASS 3.2s. /api/generate returned empty; score from /api/chat message.content|thinking. Describe ≠ image gen.
2026-09-21-5080-image_gen-blocked 2026-09-21 i9rumble 5080 image_gen none smoke blocked n/a n/a n/a none n/a n/a — ComfyUI/weights not installed none installed RTX 5080 16GB n/a n/a n/a n/a n/a n/a n/a image_gen; score=blocked n/a BLOCKED: no ComfyUI/A1111; no Qwen-Image 2.1; no Wan/LTX/Hunyuan. Survey lean 5080 CUDA only if installed later.
2026-09-21-5080-video_gen-blocked 2026-09-21 i9rumble 5080 video_gen none smoke blocked n/a n/a n/a none n/a n/a — ComfyUI/weights not installed none installed RTX 5080 16GB n/a n/a n/a n/a n/a n/a n/a video_gen; score=blocked n/a BLOCKED: no ComfyUI/A1111; no Qwen-Image 2.1; no Wan/LTX/Hunyuan. Survey lean 5080 CUDA only if installed later.

Takeaway

  • llama.cpp CLI (SYCL) on B70 — text: 12/12 pass. Describe: fail (llama-mtmd-cli). Gen lanes blocked.
  • Ollama (CUDA) on 5080 — text: mixed (gemma4-12b variants fail some tooljson/refactor). Describe: pass (3.2s wall). Gen lanes blocked.
  • Speed metrics (tok/s, TTFT, load) mostly unknown on these quality rows — see Phase A speed card when comparing decode.

← All Builds Open run-card CSV Tok mirror

Sister site: toklanes.com · Schema: run-card schema