Add card
Browse files
CARD.md
CHANGED
|
@@ -133,9 +133,21 @@ tensor load on this 300 W part) and 183 TFLOP/s at the best sample; the
|
|
| 133 |
throttle rejection is the point, not a defect: the instrument reports
|
| 134 |
the sag instead of averaging it away.
|
| 135 |
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 139 |
|
| 140 |
## Limits
|
| 141 |
|
|
|
|
| 133 |
throttle rejection is the point, not a defect: the instrument reports
|
| 134 |
the sag instead of averaging it away.
|
| 135 |
|
| 136 |
+
GeForce RTX 3070 Ti Laptop (sm86, 46 SMs, 4 MB L2), measured through the
|
| 137 |
+
published `get_kernel` artifact in the certification container:
|
| 138 |
+
|
| 139 |
+
| entry | value |
|
| 140 |
+
|---|---|
|
| 141 |
+
| clock under sustained FMA load | 1.785 GHz, 0.0% sag |
|
| 142 |
+
| HBM bandwidth | triad 419-422 GB/s; read-only 436; write-only 436; random 16 B gather 105 |
|
| 143 |
+
| L2 bandwidth | 3,332 GB/s |
|
| 144 |
+
| dependent-load latency | smem 29.0 cy; L1 43.9; L2 233; HBM 502 (0.57 ns/cy) |
|
| 145 |
+
| FMA pipes | f32 18.7 TFLOP/s; f64 0.27 TFLOP/s |
|
| 146 |
+
| MMA sustained | fp16 84.0, bf16 84.0, tf32 41.7, int8 327 (TOP/s); fp8 reported unsupported (sm86, correctly) |
|
| 147 |
+
| atomics fp32 curve | 0.56 Gop/s at 1 address up to 244 Gop/s at one per thread |
|
| 148 |
+
|
| 149 |
+
The full 23-test suite passes on both parts: on the Ada from source and
|
| 150 |
+
on the 3070 Ti through the published variants.
|
| 151 |
|
| 152 |
## Limits
|
| 153 |
|
README.md
CHANGED
|
@@ -133,9 +133,21 @@ tensor load on this 300 W part) and 183 TFLOP/s at the best sample; the
|
|
| 133 |
throttle rejection is the point, not a defect: the instrument reports
|
| 134 |
the sag instead of averaging it away.
|
| 135 |
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 139 |
|
| 140 |
## Limits
|
| 141 |
|
|
|
|
| 133 |
throttle rejection is the point, not a defect: the instrument reports
|
| 134 |
the sag instead of averaging it away.
|
| 135 |
|
| 136 |
+
GeForce RTX 3070 Ti Laptop (sm86, 46 SMs, 4 MB L2), measured through the
|
| 137 |
+
published `get_kernel` artifact in the certification container:
|
| 138 |
+
|
| 139 |
+
| entry | value |
|
| 140 |
+
|---|---|
|
| 141 |
+
| clock under sustained FMA load | 1.785 GHz, 0.0% sag |
|
| 142 |
+
| HBM bandwidth | triad 419-422 GB/s; read-only 436; write-only 436; random 16 B gather 105 |
|
| 143 |
+
| L2 bandwidth | 3,332 GB/s |
|
| 144 |
+
| dependent-load latency | smem 29.0 cy; L1 43.9; L2 233; HBM 502 (0.57 ns/cy) |
|
| 145 |
+
| FMA pipes | f32 18.7 TFLOP/s; f64 0.27 TFLOP/s |
|
| 146 |
+
| MMA sustained | fp16 84.0, bf16 84.0, tf32 41.7, int8 327 (TOP/s); fp8 reported unsupported (sm86, correctly) |
|
| 147 |
+
| atomics fp32 curve | 0.56 Gop/s at 1 address up to 244 Gop/s at one per thread |
|
| 148 |
+
|
| 149 |
+
The full 23-test suite passes on both parts: on the Ada from source and
|
| 150 |
+
on the 3070 Ti through the published variants.
|
| 151 |
|
| 152 |
## Limits
|
| 153 |
|