TensorSharp Beats llama.cpp in DeepSeek V4 Flash Multi-GPU Prefill Benchmark
fuzhongkai · reddit · 2026-08-01
The open-source inference engine TensorSharp recently added support for multi-GPU and multi-node inference. The author tested its performance running the DeepSeek-V4-Flash-0731 quantized model against llama.cpp.
Test Environment & Model:
- Hardware: 4x Nvidia A40 GPUs (CUDA 12.8)
- Model: DeepSeek-V4-Flash-0731-UD-Q8KXL (from unsloth)
Benchmark Results (TensorSharp CUDA backend vs llama.cpp):
- Prefill (16K): 836 tok/s (TensorSharp optimal) vs 558 tok/s (llama.cpp)
- Decode (short): 31.5 tok/s vs 35.3 tok/s
- Decode (16K): 28.5 tok/s vs 32.2 tok/s
TensorSharp shows a significant speed advantage in long-context prefill, though it is slightly slower than llama.cpp during decoding.
Related event: TensorSharp Engine Doubles DeepSeek Inference Speed(4 posts)→
More from Infra
- NVIDIA quietly delists the entry-level RTX 5050 from its marketplace, following the RTX 5090 — chemist_slime · 2026-10-16
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16