Leaker HXL published the results of the AIDA64 cache & memory benchmark performed on an eight-core APU sample Cezanne with a 4.7 GHz boost. There was also a sample of the “PRO” series, a processor designed for the business-class segment. The PRO versions are not said to be in regular retail sales (which is the usual state, it only ends in reports), but we have not yet learned anything about the release of the standard desktop version (not the PRO). Which in the current situation can still not be expected until half a year ago.
However, we can compare the latency and throughput of both cache and memory, although we must keep in mind that test Cezanne a Renoir did not take place on identical memories (it is practically impossible to ensure the results of two unavailable desktop APUs tested on the same memories).
- Renoir: DDR4-4200 (18-19-19-35 CR2) dual-channel
- Cezanne: DDR4-4400 (18-23-23-26 CR1) dual channel
So the memories are a little different, but you can’t say that some are better than the other (or vice versa).
| read | write | copy | latency | |
|---|---|---|---|---|
| DDR4 | 64215 MB / s | 63228 MB / s | 58845 MB / s | 62,8 ns |
| 67234 MB / s | 49694 MB / s | 60345 MB / s | 54,9 ns | |
| +4,7 % | -21,4 % | +2,5 % | -12,6 % | |
| L1 | 2132,8 GB/s | 1070,7 GB/s | 2131,7 GB/s | 0,9 ns |
| 2317,8 GB/s | 1171,6 GB/s | 2336,3 GB/s | 0,9 ns | |
| +8,7 % | +9,4 % | +9,6 % | ±0 % | |
| L2 | 1072,0 GB/s | 998,75 GB/s | 1072,7 GB/s | 2,7 ns |
| 1173,4 GB/s | 1153,4 GB/s | 1130,3 GB/s | 2,6 ns | |
| +9,5 % | +15,5 % | +5,4 % | -3,7 % | |
| L3 | 668,11 GB/s | 747,85 GB/s | 650,38 GB/s | 9,5 ns |
| 740,80 GB/s | 683,83 GB/s | 737,93 GB/s | 11,2 ns | |
| +10,9 % | -8,6 % | +13,5 % | +17,9 % |
The first line always contains the data for Renoir, in the second for the sample Cezanne. The difference is marked in color (green if the sample is Cezanne better, red if worse). Of the 16 partial tests, 12 are Cezanne better, in one it’s about felt and in three it came out better Renoir. This may be due to the fact that the sample Cezanne it is not yet fully tuned in all respects. The felt for L1 latency is then a consequence of the fact that AIDA64 expresses latency to only one decimal place, so for example 0.85 ns and 0.94 ns will be rounded and reported as 0.9 ns. Therefore, if the difference is not at least 10%, AIDA64 will not distinguish it.
All improvements that exceed 9.3% (which is the maximum possible difference in clock frequency between 4.3GHz Renoir a 4,7GHz Cezanne) is due to improvements or tuning of the architecture. This would include L1 cache latency, writing to the L2 cache, and reading and copying within the L3 cache.
–
