by
Zak Killian
—
Thursday, April 28, 2022, 04:31 PM EDT
It wasn't that long ago that we were all using regular old spinny disks for storage. Time was, you really wanted to use dedicated hardware for RAID arrays to get the best performance. Plugging them into your motherboard and using software—or worse, firmware—RAID was seen as the inferior, poor man's option. Using a proper RAID card got you access to superior performance as well as improved reliability.
These days, those in the know stick to software RAID, because hardware RAID cards simply don't cut it with multiple NVMe SSDs. Solutions like ZFS have become popular among the audience that traditionally would have used hardware RAID. Still, managing large arrays using software RAID can put a heavy load on the CPU—so much so that in systems with slower processors, it can become the bottleneck for I/O performance.
Enter GRAID Technology. A small startup in Santa Clara, California, GRAID wowed storage nerds last year when it released the SupremeRAID SR-1000 PCIe card. GRAID's technology is a sort of fusion of software and hardware RAID. Essentially, it works like software RAID, but instead of running on the CPU, it runs on a powerful GeForce GPU on the SupremeRAID card.
Last year's SupremeRAID SR-1000 is based on a lowly NVIDIA T1000 card, and yet it can deliver absolutely astonishing storage performance: over 30 GB/sec in sequential reads and nearly 9 million random read IOPS from a RAID5 array.
The original SupremeRAID 1010 card. The new model is at the top of this page.
Of course, as with anything in computing, no matter how fast something is, there will be someone who wants it to be even faster. Naturally, that means a new version of the GRAID card, known simply as the SupremeRAID SR-1010. The latest version leverages a newer NVIDIA A2000 GPU based on the GA106 GPU that powers the GeForce RTX 3060.
GRAID SupremeRAID SR-1010 Performance Specs
It's easy to forget that GA106 is a monstrously-powerful processor. Sure, it may seem unassuming next to its more-potent GeForce siblings, but we're still talking about a microprocessor that can perform some 12.7 trillion half-precision floating point operations per second, and many more if you can make use of the chip's onboard tensor processors. We don't know exactly what GRAID's algorithm looks like, because it's completely proprietary. We do know it leverages AI in some fashion, though, so it's probably making heavy use of the Ampere GPU's tensor compute units.
Whatever magic it's doing, it seems to work. GRAID says that the new card can offer read performance up to 110 GB/second in throughput, and up to 19 million IOPS. Write performance is drastically improved, too—GRAID says the new card will do up to 1.1M random write IOPS and 22GB/sec write throughput in RAID 5. This performance is possible because the GRAID card is just managing the array—none of the stored data actually passes through the PCIe interface to the GRAID card itself.
If you're fascinated by this idea, or an existing GRAID fan looking to upgrade, the company says that general availability of the new model will begin on May 1 for folks who place orders beforehand.