Sciforium Boosts DeepSeek-V4.1-Flash to 1,129 Tokens Per Second on AMD GPUs
Sciforium has achieved a significant inference speed breakthrough for the DeepSeek-V4.1-Flash AI model, reaching 1,129 output tokens per second at the P90 percentile using four AMD Instinct MI355X GPUs. This performance represents an increase from 835 tokens per second, accomplished through traffic-tuned speculative decoding, extending draft token counts to seven, and optimizing kernels specifically for AMD hardware. The optimization utilized the model's original weights without architectural modifications, demonstrating that parameter tuning and kernel engineering can substantially enhance large language model efficiency.
Editorial responsibility
- No named human review is recorded for this page.
- Reports are grouped by semantic similarity and deterministic rules. Language models may assist titles, summaries, translation and cross-source analysis; the page reads the event directly, while its address stays stable when the title changes.
- Summary covers the current reports
Reporting timeline
RT @rohanpaul_ai: Sciforium shows DeepSeek-V4.1-Flash reaching 1,129 output tokens/sec at P90 on 4 AMD Instinct MI355X GPUs, keeping DeepSe…
According to data attributed to Sciforium and shared by user rohanpaul_ai on X, the DeepSeek-V4.1-Flash AI model has demonstrated a performance benchmark of 1,129 output tokens per second at the P90 percentile. This result was achieved using a cluster of four AMD Instinct MI355X graphics processing units. The report highlights significant inference speed capabilities for this specific model configuration, suggesting high efficiency in token generation under the tested conditions. The source text indicates that this performance metric helps maintain or extend the competitive standing of the DeepSeek series against other large language models. No additional context regarding the testing environment, hardware specifications beyond the GPU count, or comparative analysis with other systems is provided in the atomic post.
Read source4 AMD GPUs. 1,129 output tokens/sec 🤯 @sciforium sped up DeepSeek-V4.1-Flash from 835 → 1,129 tok/s by: > tuning the drafter on customer traffic > drafting 7 tokens instead of 5 > optimizing kernels with AMD All using DeepSeek's original
User @sciforium has achieved a significant performance increase for the DeepSeek-V4.1-Flash AI model by utilizing four AMD GPUs. The output speed was improved from 835 tokens per second to 1,129 tokens per second through three specific optimizations: tuning the drafting mechanism based on customer traffic data, increasing the draft token count from five to seven, and optimizing kernels specifically for AMD hardware. Notably, these enhancements were implemented using the original weights provided by DeepSeek without modifying the base model architecture. This development highlights the potential for hardware-specific kernel optimization and inference parameter tuning to maximize efficiency in large language model deployments.
Read sourceSciforium shows DeepSeek-V4.1-Flash reaching 1,129 output tokens/sec at P90 on 4 AMD Instinct MI355X GPUs, keeping DeepSeek's original weights. The gain comes from speculative decoding done better. Sciforium's Traffic-tuned Drafter is tuned on customer traffic, so
Sciforium has demonstrated a significant performance breakthrough for the DeepSeek-V4.1-Flash model, achieving an output rate of 1,129 tokens per second at the P90 percentile. This result was attained using four AMD Instinct MI355X GPUs while retaining the model's original weights without modification. The performance gain is attributed to advanced speculative decoding techniques, specifically utilizing Sciforium's Traffic-tuned Drafter. This drafter is optimized using actual customer traffic data, which increases the acceptance rate of proposed tokens and allows the draft length to be extended to seven tokens. Additionally, kernel-level optimizations fused small operations like mHC and projections, and combined all-reduce operations with compute tasks. At this generation speed, producing 10,000 tokens requires approximately nine seconds. This development highlights the potential of traffic-specific tuning in speculative decoding to maximize inference efficiency on high-performance hardware.