Gigabyte Promises 219,000 TBW for New AI TOP 100E SSD

[TheLostSwede]

Gigabyte has quietly added a new SSD to its growing lineup and this time around it's something quite different. The drive is part of Gigabyte's new AI TOP (Trillions of Operations per Second) and was announced at Computex with little fanfare. At the show, the company only announced that it would have 150x the TBW compared to regular SSDs and that it was built specifically for AI model training. What that 150x means in reality is that the 2 TB version of the AI TOP 100E SSD will deliver no less than 219,000 TBW (TeraBytes Written), whereas most high-end 2 TB consumer NVMe SSDs end up somewhere around 1,200 TBW. The 1 TB version promises 109,500 TBW and both drives have an MTBF time of 1.6 million hours and a five-year warranty.

Gigabyte didn't reveal the host controller or the exact NAND used, but the drives are said to use 3D NAND flash and both drives have a LPDDR4 DRAM cache of 1 or 2 GB depending on the drive size. However, the pictures of the drive suggest it might be a Phison based reference design. The AI TOP 100E SSDs are standard PCIe 4.0 drives, so the sequential read speed tops out at 7,200 MB/s with the write speed for the 1 TB SKU being up to 6,500 MB/s, with the 2 TB SKU slightly behind at 5,900 MB/s. No other performance figures were provided. The drives are said to draw up to 11 Watts in use, which seems very high for PCIe 4.0 drives. No word on pricing or availability as yet.



View at TechPowerUp Main Site | Source

 

[Chrispy_]

Presumably this is SLC with pricing to match.

I'm unaware of any developments in NAND technology that permit such a huge increase in endurance, so it's likely a combination of SLC (at huge cost) as well as some weasel-worded caveats that mean the TBW guarantee only applies to AI workloads which are likely very small delta changes per page of NAND, and the firmware is working in a database mode tracking changes rather than rewriting pages.

I'd love to be wrong, and that Gigabyte have found a way to increase endurance, but they're neither NAND manufacturers, nor SSD controller manufacturers, meaning that they're building this with existing solutions we already know about.

The alternative is that they're just using marketing vagueness to obfuscate "150x the TBW of regular SSDs", in that by regular SSD's they mean bottom-of-the-barrel QLC DRAM-less drives rated at a pathetic 300TBW per TB of capacity. When your base is pathetic, it's easy to inflate your claims

I'm cynical because I have 25 years experience of this bullshit and 99/100 times my cynicism is justified.

 

[Dr. Dro]

This is quite the breakthrough, are they using SLC NAND for this, I wonder? The only other drive I know that has even remotely the same proportion of capacity to endurance is the original Intel X25-E from 2008 (50 nm SLC, 2 PBW for 64 GB model). This one coming in at 219 PBW for 2 TB.

 

[R0H1T]

Presumably this is SLC with pricing to match.

I'm thinking MLC or TLC with too much spare capacity for over provisioning, I doubt anyone's making SLC for them!

The only other drive I know that has even remotely the same proportion of capacity to endurance is the original Intel X25-E from 2008 (50 nm SLC, 2 PBW for 64 GB model).

You can get similar numbers by sparing 50(?) percent of drive for OPing.

Memblaze P7946 6 TB Specs

6 TB, TLC, PCIe 5.0 x4, U.2

www.techpowerup.com

Solidigm D7-P5810 800 GB Specs

800 GB, pSLC, PCIe 4.0 x4, U.2

www.techpowerup.com

 

[AnarchoPrimitiv]

Random performance figures would give us a better idea if this was an all SLC drive, perhaps that's why they're withholding them. Even so, most enterprise drives are measured in DWPD and current SLC enterprise offerings can be in the 60 - 100 range, which for a 800GB Dapustor 2900P that has 100 DWPD equates to 149,000 TBW, so GIgabyte is claiming a big improvement over even that.....so we'll see

 

[TF-GrayWizard]

Wasn't there a high endurance nvme drive from team group when chia mining was suddenly popular that was not released in the end or had a super short launch window before been EOLed.

iirc it was micron 64L FortisMAX that had 10,000 P/E cycles.

 

[TheLostSwede]

Random performance figures would give us a better idea if this was an all SLC drive, perhaps that's why they're withholding them. Even so, most enterprise drives are measured in DWPD and current SLC enterprise offerings can be in the 60 - 100 range, which for a 800GB Dapustor 2900P that has 100 DWPD equates to 149,000 TBW, so GIgabyte is claiming a big improvement over even that.....so we'll see

Tom's calculated that the claimed TBW figure would allow you to fill the 2 TB drive for 17 hours a day for five years, so not quite a DWPD figure, but close.

 

[R0H1T]

It was TLC, the pSLC mode had up to 4x P/E cycles.

Wasn't there a high endurance nvme drive from team group when chia mining was suddenly popular that was not released in the end or had a super short launch window before been EOLed.

iirc it was micron 64L FortisMAX that had 10,000 P/E cycles.

Teamgroup T-Create Expert 2 TB Specs

2 TB, TLC, PCIe 3.0 x4, M.2 2280

www.techpowerup.com

 

[P4-630]

Teamgroup T-Create Expert 2 TB Specs​

2 TB, TLC, PCIe 3.0 x4, M.2 2280

A 12-Year warranty though!..

 

[R0H1T]

At those write speeds it'll probably take 20 years to exhaust the rated endurance. Even if it's 100% busy half of the time

Just guesstimating, so don't go full scientific calculator mode on me

 

[chrcoluk]

Presumably this is SLC with pricing to match.

I'm unaware of any developments in NAND technology that permit such a huge increase in endurance, so it's likely a combination of SLC (at huge cost) as well as some weasel-worded caveats that mean the TBW guarantee only applies to AI workloads which are likely very small delta changes per page of NAND, and the firmware is working in a database mode tracking changes rather than rewriting pages.

I'd love to be wrong, and that Gigabyte have found a way to increase endurance, but they're neither NAND manufacturers, nor SSD controller manufacturers, meaning that they're building this with existing solutions we already know about.

The alternative is that they're just using marketing vagueness to obfuscate "150x the TBW of regular SSDs", in that by regular SSD's they mean bottom-of-the-barrel QLC DRAM-less drives rated at a pathetic 300TBW per TB of capacity. When your base is pathetic, it's easy to inflate your claims

I'm cynical because I have 25 years experience of this bullshit and 99/100 times my cynicism is justified.

It could be achieved with some changes like PLP, ignoring disk flushes, firmware behavioural changes and an appropriate amount of over provisioning. Also bear in mind consumer 3D TLC drives are likely usually under rated to avoid the enterprise market using them.

I agree with you though I expect a non standard price for these drives.

 

[evernessince]

I'm unaware of any developments in NAND technology that permit such a huge increase in endurance, so it's likely a combination of SLC (at huge cost) as well as some weasel-worded caveats that mean the TBW guarantee only applies to AI workloads which are likely very small delta changes per page of NAND, and the firmware is working in a database mode tracking changes rather than rewriting pages.

The way AI training works is the entire model and associated data is dumped to disk at a specified interval. This is to ensure training can be resumed and quality can be compared between different points in the training process. Picking the best model is usually a matter of finding the local minima and selecting the best from among those so you need to keep a lot of these backups on hand.

Space wise you are talking 11 - 14 GB per backup for something small like a LoRA. It'll be much larger for a checkpoint. Typically when I'm training a LORA, it quickly eats all the space on my P5800X 800GB. I can fill it up in 8 hours.

Endurance wise the drive in the article is pretty impressive, almost as good as Optane if the specs are true. It's on the level with cache drives so I expect it to be pricey.

 

[R0H1T]

It it's up to a 2x premium it could be digestible. Anything above & they should sell it to the next Chia farmers!

 

[TheLostSwede]

Seems like the 2 TB version vill cost around US$312 in Vietnamn. The same shop charges as much for Gigabyte's 2 TB Aorus Gen5 14000 drive.

Ổ cứng SSD GIGABYTE AI TOP 100E 2TB NVMe PCIe Gen 4.0 x4 3D NAND M.2 2280 (AI100E2TB)

Ổ cứng SSD GIGABYTE AI TOP 100E 2TB NVMe PCIe Gen 4.0 x4 3D NAND M.2 2280 (AI100E2TB)

mygear.vn

 

[AnarchoPrimitiv]

Personally, I would absolutely love an all SLC consumer drive, something like 250-500GB just to load your OS and a handful of apps on....enterprise SLC drive get extremely close to optane in their random performance figures

Tom's calculated that the claimed TBW figure would allow you to fill the 2 TB drive for 17 hours a day for five years, so not quite a DWPD figure, but close.

I used an online calculator that converts drive endurance figures

 

[Chuuei]

Seems like the 2 TB version vill cost around US$312 in Vietnam. The same shop charges as much for Gigabyte's 2 TB Aorus Gen5 14000 drive.

Pretty sure they just guessed a price based on existing products


pSLC mode for modern flash gives you similar P/E cycles as old SLC, around 100k. With 4TB raw TLC you get 1.3 TB pSLC which in theory could get you around 100k TBW but the workload must be almost optimal to realize it in practice.

I think more likely we see 8TB and 16TB raw QLC NAND, which gives you 2 TB resp. 4 TB of pSLC and allows more overprovisioning to get the advertised TBW.

 

[remixedcat]

isn't DRAM less SSD seen as better? what is the reasoning behind that? I've heard a lot of people saying that's better for some reason but I generally feel the opposite.

 

[Dr. Dro]

isn't DRAM less SSD seen as better? what is the reasoning behind that? I've heard a lot of people saying that's better for some reason but I generally feel the opposite.

No, it's the other way around. DRAMless SSDs tend to be slower or they use system memory as a cache (HMB), invariably, SSDs with a built-in DRAM cache tend to be faster and more reliable than the drives tht don't have one

 

[remixedcat]

No, it's the other way around. DRAMless SSDs tend to be slower or they use system memory as a cache (HMB), invariably, SSDs with a built-in DRAM cache tend to be faster and more reliable than the drives tht don't have one

Wonder if mine are...hmmm

 

[Yraggul666]

Explosion and/or fire after how many TBW?

Seriously now: i'd certainly buy something like this if it was made by some other brand. The fact that i could really rely on a ssd to this extent would really give me huge peace of mind.

 

[R0H1T]

Explosions w/SSD really

Btw the NAND used in most of these brands are from the same top 3/4 companies.

 

[Chrispy_]

Wonder if mine are...hmmm

On Windows the HBM implementation is usually good enough that DRAMless doesn't matter. The WD SN580 and SN770 are both DRAMless and perform very well indeed - there are a few more expensive drives with DRAM that fall behind them in the charts, so DRAMless isn't really a performance issue where the OS supports it.

DRAMless is a problem when you're using the drive as a member disk in a NAS/SAN RAID group attached to a hardware controller that doesn't permit HMB.

Explosions w/SSD really

I think you missed the joke:

 

[TheLostSwede]

isn't DRAM less SSD seen as better? what is the reasoning behind that? I've heard a lot of people saying that's better for some reason but I generally feel the opposite.

The only apparent benefit is a cooler running SSD, everything else tend to be inferior to drives with DRAM, which generally isn't a cache, but instead stores the mapping table of the NAND.

 

[remixedcat]

On Windows the HBM implementation is usually good enough that DRAMless doesn't matter. The WD SN580 and SN770 are both DRAMless and perform very well indeed - there are a few more expensive drives with DRAM that fall behind them in the charts, so DRAMless isn't really a performance issue where the OS supports it.

DRAMless is a problem when you're using the drive as a member disk in a NAS/SAN RAID group attached to a hardware controller that doesn't permit HMB.


I think you missed the joke:

WD PCSN520. and SK Hynix BC901 1024GB are the nvme drives I have I dunno if they have DRAM or not and I don't wanna take apart the laptops to find out cuz it's too hot for that lol

Ok I know the SK Hynix drive is DRAM-less now... but that leaves the WD one

 

[Chrispy_]

WD PCSN520. and SK Hynix BC901 1024GB are the nvme drives I have I dunno if they have DRAM or not and I don't wanna take apart the laptops to find out cuz it's too hot for that lol

Ok I know the SK Hynix drive is DRAM-less now... but that leaves the WD one

I suspect not. The SN520 is a very low-performance Gen3x2 SSD that is better than a SATA drive but not by much. Either way, it's normally used in small spaces and is probably a M.2 2242 size that's physically too small to have room for a DRAM chip anyway. It's also so old that it predates HMB, so it's just a slow SSD, period. That probably doesn't matter too much though - even a SATA drive is fast enough that most consumer applications don't really benefit from anything faster. If you were running an i9 or Ryzen9 and doing heavy write-intensive workloads then sure, it'd feel like a turd - but for everyday booting, gaming, web-browsing it's likely to be plenty fast enough.

The BC901 is another low-end drive that's DRAMless for sure, but at least it's Gen4x4 and supports HMB so the lack of DRAM isn't a big deal.