Update README.md

README.md

@@ -312,7 +312,9 @@ Sample sizes: Carbon & GENERator n=500. Evo2-7B n=150 at 16k, n=100 at 32k, n=10
 
 ### Inference efficiency
 
-> TODO: add Ed's benchmarks
+Carbon models run natively in vLLM and thus generate DNA sequences over 150 times faster than the Evo2 family of models. Below we show the results of a throughput benchmark, where 1080 base-pairs are used for prefill and decode with increasing number of input sequences. All models except Evo2 40B were run on a H100 GPU, with the batch size of the Evo2 models tuned to the largest possible size that fits in VRAM.
+
+![](figures/inference.png)
 
 ## Training