@@ -237,7 +237,8 @@ and the additional bandwidth reduction from LVQ-4x4 and LVQ-4x8 provides a signi
237237| **LVQ-4x8 ** | 2.5x | 1.3x | **3.1x ** | 1.6x | **4.7x ** | 2.1x |
238238+--------------------+----------+----------+----------+----------+----------+----------+
239239
240- **Click on the triangles ** to see the throughput vs recall curves comparing several LVQ settings for each dataset.
240+ **Click on the triangles ** to see the throughput vs recall curves comparing SVS with several LVQ settings, as well as float32
241+ and float16 encodings, for each dataset.
241242
242243.. collapse :: deep-96-1B
243244
@@ -273,6 +274,37 @@ and the additional bandwidth reduction from LVQ-4x4 and LVQ-4x8 provides a signi
273274
274275|
275276
277+ **Click on the triangles ** to see the throughput vs recall curves comparing the highest performing SVS-LVQ setting
278+ (a Pareto curve built with results from LVQ-8, LVQ-4x4, LVQ4x8 and LVQ8x8), SVS using float32 and float16 encodings,
279+ as well as other methods for each dataset.
280+
281+ .. collapse :: deep-96-1B
282+
283+ Results for the deep-96-1B dataset
284+
285+ .. image :: ../figs/bench_largeScale_ablation_and_other_methods_deep-1B.png
286+ :width: 700
287+ :alt: deep-96-1B compression evaluation results
288+
289+ .. collapse :: t2i-200-100M
290+
291+ Results for the t2i-200-100M dataset
292+
293+ .. image :: ../figs/bench_largeScale_ablation_and_other_methods_text2image-100M.png
294+ :width: 700
295+ :alt: t2i-200-100M compression evaluation results
296+
297+
298+ .. collapse :: DPR-768-10M
299+
300+ Results for the DPR-768-10M dataset
301+
302+ .. image :: ../figs/bench_largeScale_ablation_and_other_methods_dpr-10M.png
303+ :width: 700
304+ :alt: DPR-768-10M compression evaluation results
305+
306+ |
307+
276308.. |copy | unicode :: U+000A9 .. COPYRIGHT SIGN
277309.. |reg | unicode :: U+00AE .. REGISTERED
278310
0 commit comments