Biwin Black Opal X570 PRO Review – A new player on the market with an interesting product and minor weaknesses

1 - Introduction and Technical Data 2 - Teardown with Component and Material Analysis 3 - Streaming and Root Cause Analysis of Drops 4 - Synthetic Benchmarks 5 - Workstation Benchmarks 6 - Temperatures, Power Consumption and Conclusion

Foreword to the synthetic benchmarks

The direct comparison between the synthetic results from e.g. CrystalDiskMark and the practical measured values from the AJA System Test reveals a central weakness of many conventional SSD benchmarks: They show theoretically achievable peak performance under ideal conditions, but do not allow any conclusions to be drawn about the behavior under real continuous load or with continuous access patterns. Even when continuously writing large amounts of data in capture mode, there were numerous, sometimes massive drops in the transfer rate. Such drops into the single-digit MB/s range indicate internal limitations due to flash management, cache saturation or inconsistent FTL management. The situation was similar in the read test: initially stable, but later increasingly fragmented, with pronounced jitter and constant fluctuations in the actual transfer rate.

This discrepancy is no coincidence, but a consequence of the fundamental orientation of synthetic benchmarks. Tools such as CrystalDiskMark test short, predictable access patterns with low variance. The block sizes, queue depth and access directions are known, the test period is limited and the data volumes to be measured usually fit completely in the pSLC cache, DRAM buffer or even dedicated controller area. Modern SSD controllers in particular (and also the firmware) recognize these typical benchmarks based on the access patterns and behave in a specifically “optimized” way in order to deliver particularly good results in these tests. In practice, this means that benchmark performance is achieved at the cost of targeted prioritization of synthetic I/O structures that have little in common with real workloads.

The decisive factor is not how high the numerical value is under laboratory conditions, but how consistent and stable an SSD behaves in realistic scenarios – and this is precisely where tools such as the AJA System Test do much more valuable work. So if you only rely on green bars, you risk ending up with a drive that shines on paper but reacts unreliably in operation. Incidentally, the different drive letter in the second test run (E: instead of D:) was purely coincidental, as an external SSD was also connected via USB during the second run. The different drive size results from the image that was loaded. For particularly attentive readers, this should suffice as a small side note to forestall any queries. The test results remain unaffected by this.

CrystalDiskMark

The screenshots shown document a detailed synthetic performance comparison of the Biwin X570 PRO 4TB SSD using CrystalDiskMark 6.0.2, in each case with test files of different sizes and in two clearly defined operating states. Not only the behavior in the brand-new state is recorded, but also the performance behavior after real load, specifically by restoring a 1.6 TB system image and subsequent use as a system drive with a typical workstation load. The measurement results thus provide a valuable insight into the consistency of the controller, the durability of the cache mechanisms and the behavior under fragmentation or mapping load.

The first four screenshots show the SSD in its virgin state, completely empty, without partitioning and directly after initialization. The sequential read and write rates are consistently at a very high level and hardly scale down even with increasing test file sizes. The sequential read rates level off at 14.4 GB/s, which almost corresponds to the maximum specified value for the SM2508 controller. The sequential write rates also remain stable at around 13.4 GB/s, which indicates an actively working and previously unsaturated pSLC cache structure.

The 4K values at low queue depth (Q1T1) are within the expected range, while very decent throughput values are achieved at 4K Q32T1 and Q8T8. It is striking that the values from 1 GiB to 32 GiB remain practically identical, which indicates that all benchmarks take place entirely in the DRAM-supported pSLC area and no physical reorganization has yet been necessary in the TLC area.

In direct comparison, the four screenshots below show the performance behavior after an intensive usage phase. Here, a 1.6 TB system image was previously loaded with Macrium Reflect and the SSD was used as a system drive under real workload. Numerous I/O-intensive benchmarks were carried out, which heavily fragment the SSD internally and permanently stress the cache mechanisms. In terms of sequential performance, the SSD remains impressively constant. Read values of 13.8 to 14.1 GB/s are only slightly below the new state, and the write performance of around 13.4 GB/s also remains stable. This shows that neither thermal aging nor flash wear are directly recognizable in the sequential performance.

However, an interesting behavior is revealed in the area of 4K random accesses. With 4K Q8T8 in particular, there is a noticeable jump in the read rate from the previous 3000 MB/s to values of over 4000 MB/s. This improvement can be explained by the active mapping data structures actually used: With an occupied and used SSD, the FTL management can react much more efficiently to data patterns that resemble real workloads. It is conceivable that additional cache paths or optimized path finding in the page and block tables take effect here, which are not activated in the idle state.

The remaining random values for 4K Q32T1 and Q1T1 remain stable and within tolerance. Slight fluctuations in the range of ±3% are normal for SSDs under real workload and do not indicate inconsistent performance. Thus, the Biwin X570 PRO shows a very constant and responsive performance, both in the ideal synthetic state and after real system load and heavy utilization. The fact that there are no drops in write rates under continuous load and even larger test buffers such as 32 GiB are processed without any noticeable drops in performance speaks for very efficient cache and garbage management, which is not overwhelmed even under heavy load.

ATTO benchmark

At first glance, the ATTO benchmarks also show a consistently impressive performance of the Biwin X570 PRO 4TB, both when brand new and after prolonged use. However, especially in comparison with the previously discussed, more realistic AJA results, it once again becomes clear how unreliable synthetic benchmarks are as the sole basis for performance evaluations. The ATTO results also document the behavior under optimal conditions rather than the actual behavior in continuous operation under realistic loads.

The first two screenshots show the SSD in its virgin state with a test file size of initially 1 GB, then 32 GB. As expected, the transfer rates increase continuously as the I/O size increases. Write speeds of over 12 GB/s and read rates above 13.5 GB/s are achieved from as little as 128 KB. The maximum values stabilize at around 12.5 GB/s write and almost 14 GB/s read, regardless of the selected file size. This indicates an aggressively working SLC cache, low fragmentation and optimum access to a completely empty NAND structure.

The two screenshots below show the result of the same test after intensive use of the SSD with system operation and a large number of real applications, including a restored 1.6 TB image. Again, 1 GB and 32 GB test files were used, the progression remains almost identical to the measurement results in the brand-new state. The transfer rates also increase quickly with larger I/O sizes and reach the same maximum values as before. Even after prolonged use and real operation with more complex data structures, the linear scaling of the transfer rates in ATTO remains virtually unaffected.

This is precisely the problem with synthetic benchmarks such as ATTO: Just like CrystalDiskMark, they deliver typical values that result from controlled access patterns. The test structure is predictable, access to the file system takes place in a strictly linear framework, the amount of data is limited and can be processed in the pSLC cache and DRAM buffer without any problems. The use of “Direct I/O” and “Bypass Write Cache” also contributes more to testing the behavior of the SSD in the most ideal way possible – but not under conditions that occur in everyday operation.

The conspicuous consistency between empty and used SSDs in the ATTO test once again underlines the point that such benchmarks often do not allow any statement to be made about long-term performance, thermal effects, cache overflows or management load. There is also sufficient evidence that many SSD firmwares react specifically to typical benchmark patterns such as those of ATTO or CrystalDiskMark – i.e. prioritize resources internally, time cache mechanisms more aggressively or even postpone reorganizations in order to deliver good figures.

A comparison with the AJA results therefore shows once again: Where ATTO draws stable bars, the SSD stumbles during continuous read and write operation, showing numerous dips and inconsistent performance. These dips have a real impact on workloads that exceed the SLC size or do not pay attention to optimal block sizes and even load distribution. ATTO therefore provides a clean snapshot of ideal performance, but nothing more. For everyday, workstation or production scenarios, the informative value remains limited. If you want to know how an SSD behaves in real life, you have to test beyond ATTO.