Unverified50% confidenceBenchmarkExact time
The R9700 prefill throughput peaks at 3619 tok/s at a depth of 16000 tokens, drops back to 3192 tok/s at a depth of 64000 tokens, a decay of approximately 12%.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/17/2026
First Seen
Valid until: 12/16/2026
Sources
Related Entities
Related Claims
UnverifiedWhen running 4-bit DeepSeek R1 on the M3 Ultra, a 1,000-token context generates approximately 19 tokens per second using 392GB, dropping to approximately 15 tokens per second using 410GB at a 16,000-token context.75% similarUnverifiedThe context window of DeepSeek-R1 7B can typically reach 128K tokens, but the actual usable length locally is limited by VRAM capacity.68% similarUnverifiedThrough optimization of the KV Cache, the cache required by DeepSeek per token has dropped to the 1/1000 level, and model memory and SSD requirements have been reduced to one-quarter and one-eighth of the previous generation respectively.64% similarUnverifiedFor the LLaMA-2 70B model, when processing a context of 4096 tokens, the KV Cache requires approximately 20 GB of system memory at FP16 precision.64% similarUnverifiedThe Q8 main model occupies 162 GB, preventing the drafter from fitting into unified memory, rendering DSpark acceleration ineffective; after moving three layers of MoE experts back to the CPU, output only improved from 22.87 to 23.49 tokens/second (2.7% faster), while prefill dropped from 95.38 to 76.65.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/930097API
curl https://kongchang.com/api/v1/knowledge/claims/930097MCP
get_claim(id=930097)