llamaperf

Intel Arc A770 16GB reports

6 reports. These are individual setups, not a controlled benchmark. Intel Arc A770 16GB speeds and models →

Speeds are as each source reported them. GPU count, offloading and concurrent requests change them, and prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing two reports.

Tone: mixedIntel hardware
generation:
69.7 tokens/s
prompt processing (prefill):
2074.2 tokens/s
quant:
Q4_0 (GGUF)

User benchmarks Llama 2 7B Q4_0 on an Arc A770 16GB with arcint's llama.cpp patches and flash attention on, reaching 2074.21 t/s prefill and 69.71 t/s decode in llama-bench (512 prefill, 128 decode, 5 repetitions). Both builds are at llama.cpp commit 3e4dfa8 on a Ryzen 5700X, Debian 13, xe driver; the A770 sits on PCIe 3.0 x4 behind the chipset with its clock pinned at 2000 MHz. Stock SYCL fp16 with flash attention decodes faster at 73.66 t/s but prefills at 1457.24 t/s; arcint leads prefill by 10.5% on the A770 and 31% on the B60, while SYCL wins decode by 5.4% and 4.4%. Through arcint's own server the A770 gives 1963 / 70.5. On the B60, Qwen3.6-35B-A3B Q4_K_M runs at 2332 / 84.9 and Qwen3.8-27B Q4_K_M at 844 / 19.9; the 27B at full 262,144 context with MTP drafts decodes 47.1 t/s at 3,599 prompt tokens down to 12.5 t/s at 234,553.

Nov 5, 2026

Gemma 4 12B

Intel Arc A770 16GB · llama.cpp · 128,000 ctx

Tone: positiveIntel hardware
generation:
20.0 tokens/s
prompt processing (prefill):
1100.0 tokens/s
quant:
QAT
kv:
Q8
multilingualrp

User reports Gemma 4 12B QAT at 20 t/s decode and 1100 t/s prefill on a single Intel Arc A770 16GB. Setup is llama.cpp with QAT weights, Q8 KV cache and a Q4 MTP head, giving 128k context with about 5GB VRAM headroom; the card is in a 13-year-old Haswell desktop with under 6GB RAM headroom. Four models run together at 14-15GB VRAM via llama-swap, with LiteLLM routing prompts through Laya to thinking and non-thinking aliases of the 12B. The user notes SYCL/llama.cpp lacks Alchemist optimizations, making the A770 the bottleneck, and that GPU-resident models still use 4-5GB of system RAM as compute buffer.

Oct 26, 2026
Tone: positiveIntel hardware
generation:
49.0 tokens/s
prompt processing (prefill):
48.0 tokens/s
quant:
Q6_K (GGUF)
kv:
q4_1
flash attention:
on
codingtool-usevisionlong-contextagentic

User reports Qwen3.5-9B at 49 t/s generation and 48 t/s prompt processing on an Intel Arc A770 16GB. Setup is llama.cpp b9521 (Vulkan) with Q6_K weights, q4_1 KV cache, 256K context, flash attention, vision mmproj and MTP speculative decoding on a single slot. The same model under WSL2 with Q8_0 KV cache and 128K context reached about 40 t/s without vision. Qwopus3.5-4B-Coder hit 64 t/s generation and 100 t/s prompt at 96K context, and Gemma 4 12B reached 22 t/s generation and 74 t/s prompt at 128K context.

Oct 7, 2026
Intel hardware
generation:
14.4 tokens/s
quant:
Q4_K_M (GGUF)
coding

User reports Qwen3.6-27B-A3B-Coder at 14.4 t/s decode on an Intel Arc A770 16GB, scoring 10/10 on a Lua CSV parser acceptance task. Setup is llama.cpp with the SYCL backend and GGUF Q4_K_M weights, a comparison point against arcint's own AWQ IR serving on the same card. The figure is described as a rough bound rather than a directly comparable measurement, since the quantisation and engine differ from arcint's production configuration.

Oct 5, 2026
Intel hardware
generation:
33.2 tokens/s
quant:
Q8_0 (GGUF)

User benchmarks Qwen3-1.7B at 33.2 tok/s on an Intel Arc A770 16GB using llama.cpp with the SYCL backend. Setup is llama.cpp SYCL with Q8_0 quantisation, single-stream (n_parallel=1). OpenVINO via OVMS reached 65.4 tok/s on the same model. Ten models were tested in total, with OpenVINO faster than SYCL on every model.

Sep 23, 2026
Tone: mixedIntel hardware
generation:
12.0 tokens/s
quant:
Q3_K_S (GGUF)

User reports Qwen3.8 27B at 12 t/s on an Intel Arc A770 using pure K-quants. Setup uses the SYCL backend, where I-quant mixes like Unsloth's UD-Q3_K_XL run at 7 t/s while pure K-quants like Bartowski's Q3_K_S run at 12 t/s. The user notes SYCL is bottlenecked by inefficient I-quant handling. The user is looking for the smallest K-quant-only Qwen3.8 27B at q3, with Bartowski's Q3_K_S at 12.7 GB as the current top contender.

Sep 14, 2026