top of page

Reviews on R7608A in ITG Marketing Booth 

Even with an ultra-fast SSD, the speed does not come out at all in Explorer?! We asked the developers of "FastCopy" what was necessary to achieve a "practical speed" of over 50 GB/s

April 11, 2025

Credit: Samsung

001_s.jpg

Mr. Hiroaki shiramizu, director and CEO of FastCopy laboratory, interviewed at ITG marketing

The speed of PC storage increases year by year. On the 5.0-Gen PCle SSD, the faster models reach speeds that are just around the corner of 15GB /s. But did you know that, in fact, this performance is rarely seen in Windows explorer? If you use explorer to copy files, you'll get stuck at a few GB/s, no matter how fast the internal SSD is.

Benchmark testing does give you the ability to see specs, but that's just benchmark testing. As long as the actual file operations are done in explorer, they are much slower to use, so fast storage is a treasure.

This isn't just about the existing PCle 4.0 or PCle 5.0. Raid CARDS with multiple SSDS for faster speeds, or external SSDS connected to Thunderbolt 5. Even if faster standards come along, Windows explorer will still be a bottleneck. The higher the storage specs, the bigger the gap between expectations, and the more obvious the impact will be.

Mr. Hiroaki shiramizu, who developed FastCopy, was frustrated by this situation. With its original ideas, it achieves reliable file copy and backup while drawing out the original performance of storage. We asked him what kind of technology is packed in there.

What is FastCopy?

There was a demonstration of the comfort of 8K video editing at that time. Even if the operation during editing becomes comfortable due to faster storage, it cannot be said that it becomes really comfortable unless the exchange of material data and finished data (copy processing) that occurs before and after it is also accelerated. If you use explorer to transfer data, the performance will be much slower than what the storage has.

This is where FastCopy, developed by Mr. Shiramizu, plays an active role. The show also demonstrated fast file transfer using FastCopy, as well as an external HighPoint RocketStor enclosure with an SSD for Samsung data centers. 6542aw "and Lenovo ThinkStation P8 showed high performance.

Fastcopy is a Windows tool to copy (or delete) specified files and folders to the desired location.

Personal use is free, and commercial use of the "Pro version" requires a paid license. It has been used not only in video production sites but also in major companies and government agencies both in Japan and overseas. Shiramizu is also known as the developer of IP Messenger, a LAN message exchange software that appeared in the Windows 95 era, and some readers may have been a long-time supporter of him.

In addition to the ability to simply copy, the Pro version also features validation, which increases reliability by generating hashes and comparing data to check if it is correctly cobyed.

Fastcopy (Windows) It was a long-running tool since the XP era, and shiramizu realized that explorer was not able to take full advantage of HDD's performance, and he felt that there were some unsuitable parts for backup work other than speed.  From such dissatisfaction, he started development with the motivation of "wanting to bring out 100% of the performance, to backup without stress".

003_s.jpg

Custom model of LenovoThinkStation P8 exhibited at Inter BEE 2024

DSC01513_s.jpg

External CDP enclosure "RocketStor 6542AW"

exhibited at Inter BEE 2024

007_s.jpg

If you try to copy files using Windows explorer on a 7GB /s PCle 4.0 compatible SSD, it will only run 2.5GB /s or more

005_s.jpg

Fastcopy performed file copying at a speed of

about 14.5GB/s, which was quite fast.

004_s.jpg

Custom model of LenovoThinkStation P8 exhibited at Inter BEE 2024

DSC02334_s.jpg

The benchmark's sequential read is about 27GB/s

dc1_s.png

I prepared two SSDs with a read/write speed of over 5GB /s, and the speed of copying data with Windows explorer was only about 2.5GB /s.

Why does FastCopy copy data so fast?

"I don't know exactly how it works internally," he said of the reason why Windows explorer is so slow. Explorer in the xp-era behaves like one is running while the other is resting, or OS file caching is rather overhead for large data copies." He saw it as the cause. Although parallel processing has been performed since Windows vista, the reason why the current explorer does not show the speed is "direct 1/0 is not used, asynchronous 1/0 is issued few, etc. The possibility is high."

006_s.jpg

There seems to be very little asynchronous 1/O processing in explorer

So I created FastCopy based on the idea of running each read and write on a dedicated thread, and then taking a large buffer capacity. It took about six months from the start to form as a software, until it was released in 2004.

Since then, it has continued to expand its functions and performance by increasing the speed and adding validation in response to user requests. The use of asynchronous 1/0 processing and direct 1/o(processing that does not involve the OS's cache mechanism) to communicate directly with the device are also innovations.

In addition, considering the time it takes to seek an HDD, we can devise an algorithm to write to the place where it is the shortest, or we can copy from one drive to another and from the same drive, or use different methods for HDD and SSD. And so on, the process is optimized for the situation.

Also, especially the NVMe For example, in the case of the SCSI system, it is important to note that the SCSI system is designed to be "at least a few hundred times more efficient than the SCSI system". In addition, if a large number of queues are issued to the HDD, it seems to be a problem, but in the case of Windows, it is possible to determine whether the storage is SSD or HDD from the application side, not limited to FastCopy to the target storage It is possible to switch the operation together. Shiramizu speculates that explorer doesn't use these techniques very much.

Samsung working together for even faster speed the latest FastCopy prototype achieves over 50GB /s measured

The latest version of FastCopy, v5.8.1, released in January 2025, enables high-speed processing by utilizing these various techniques. However, shiramizu currently has an undisclosed alpha test version called "v6.0.0" running.

008_s.jpg

Customized model of Lenovo ThinkStation P8 provided as verification equipment

As you can see in the photos from that time, when using FastCopy for copying files of about 20GBfor external enclosures, the average transfer speed was about 14.5GB /s. It is much faster than explorer, but the benchmark software has about 27GB/s read and about 24GB/s write, which is not the same.

In order to find out the reason for the lack of storage specifications, bai shui used Lenovo ThinkStation P8 to conduct verification. The AMD Threadripper Pro7975WX is equipped with the R7608A RAID card from HighPoint, which is equipped with eight Samsung 990 PRO chips.

009_s.jpg

RAID card from HighPoint R7608A

DSC05778_s.jpg
DSC05779_s.jpg

Equipped with a large heat sink to cool SSD and RAID controller

The 8x 990 PRO 2tb SSD that came with it

010a_s.jpg

Versions have shown that in ultra-fast 1/0 environments with transfer speeds of several tens of GB/s, the asynchronous read/write process of the data is not performed in itself, but in an earlier stage. We found that synchronous processing in s is a bottleneck. One CPU core was taken up 100%, and it was impossible to make additional asynchronous 1/0 issues.

This allows a typical Windows RAID to run up to 28GB/s per core on a HighPoint RAID card environment. 0(stripe volume) was supposed to only produce a lower 13GB /s.

In order to solve this problem, Mr. Shiramizu "made a thread dedicated to 1/0 processing" so that multiple cores can be processed simultaneously. It also implemented a mechanism that "the 1/0 completion notification may be disorganized, not in the order in which instructions are issued, but if a thread completes somewhere, it immediately moves to the next 1/0 processing." Add a pool from which multiple threads can simultaneously issue asynchronous read-write instructions.

In the end, the core part of the coby processing, such as "greatly modifying the conventional design and structure, and also modifying the library", was updated so that almost all of it was written. Although it was almost full scratch, there was a reason why it did not come up to a certain speed, and it was actually completed after 35 days of operation time.

010_s.jpg

The number in brackets following 52.154 on the TransRate column (109.977) is the total speed of both read and write. What is it?

It was also confirmed that data was transferred from Windows task manager over 55GB/s

"Up until then, I had only been using equipment that could produce up to 20GB /s, so I didn't realize what would happen with this ultra-fast 1/0 processing." "However, there is no doubt that the speed of ordinary users' devices will increase steadily in the future, and I think this effort was meaningful in terms of creating a path to make the most of that dynamic," says shiramizu.

AKIBA PC Hotline!

@watch_akiba

A monster machine exhibited at ITG's marketing booth at ASK*FES 2025 with 8 samsung 990 pro achieving over 58gb /s. He demonstrated file transfer over 50gb /s. There is no shortage of opportunities to see this level of PC, so please come to the booth

bottom of page