Schestex commited on
Commit
62b484c
·
verified ·
1 Parent(s): 6132d8e

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +30 -3
README.md CHANGED
@@ -335,10 +335,37 @@ I44 is designed to retain much of that performance benefit while excluding the i
335
  ### Current I44 benchmark status
336
 
337
  ```text
338
- D48 + O44 + I44 performance: pending final benchmark result
339
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
340
 
341
- The README will be updated with the measured I44 throughput and fraction of the I64 speed ceiling retained once that run is frozen.
 
 
 
 
 
 
 
 
 
 
 
 
342
 
343
  ---
344
 
 
335
  ### Current I44 benchmark status
336
 
337
  ```text
338
+ ## D48 + O44 + I44 performance
339
+
340
+ The I44 candidate was benchmarked against D48 + O44 and the performance-only I64 ceiling on the same physical RTX PRO 4000 Blackwell GPU.
341
+
342
+ | Test | Metric | D48 + O44 | I44 | I44 gain | I64 ceiling | Ceiling retained |
343
+ |---|---|---:|---:|---:|---:|---:|
344
+ | pp512 | prefill | 840.81 | 973.42 | +15.77% | 1047.02 | 64.3% |
345
+ | pp2048 | prefill | 850.04 | 998.64 | +17.48% | 1086.12 | 62.9% |
346
+ | pp8192 | prefill | 790.06 | 945.86 | +19.72% | 1029.31 | 65.1% |
347
+ | tg256 | decode | 19.06 | 20.52 | +7.68% | 21.19 | 68.5% |
348
+ | tg1024 | decode | 18.72 | 19.78 | +5.64% | 20.33 | 65.5% |
349
+ | pp2048+tg512 | prefill | 765.80 | 899.07 | +17.40% | 988.97 | 59.7% |
350
+ | pp2048+tg512 | decode | 18.70 | 19.68 | +5.23% | 20.37 | 58.7% |
351
+ | pp8192+tg512 | prefill | 754.14 | 886.13 | +17.50% | 964.10 | 62.9% |
352
+ | pp8192+tg512 | decode | 18.52 | 19.40 | +4.74% | 20.08 | 56.3% |
353
+
354
+ Summary:
355
 
356
+ - **Mean I44 prefill gain:** +17.576%
357
+ - **Mean I44 decode gain:** +5.822%
358
+ - **Mean prefill I64 ceiling retained:** 63.0%
359
+ - **Mean decode I64 ceiling retained:** 62.2%
360
+
361
+ Artifact details:
362
+
363
+ ```text
364
+ artifact bytes = 16,371,739,140
365
+ runtime weights = 16,358,539,264
366
+ workspace = 159,981,568
367
+ SHA256 = 09babb880c5d0ea112005ec6701ab7e6a074683318c0ac47cefbe60b80d55040
368
+ ```
369
 
370
  ---
371