Curious why they use a PCIe Switch - does the Mac Pro not support bifurcation? The Xeon definitely has enough PCIe lanes, so using something like a $50 ASUS M.2 X16 card would be a cheaper option (and one that gives x16 instead of x8 speed), especially since the RAID is software-side anyway.
That said, the price isn't bad for a card that uses a PCIe switch, I just question why you'd buy a card with a PCIe Switch if you don't have a lack of lanes available - the Cascade Lake Xeon W's have 64 PCIe 3.0 lanes across the board.
The Mac Pro has one PCIe x16 slot, but two PCIe x8 slots: https://www.apple.com/uk/mac-pro/specs/ Realistically, you probably won't be using 16x of bandwidth, so using the 8x slots is more efficient. It means you could use four cards at least at half capacity, which isn't possible with a 2 x 4x card.
Yeah, but the x16 slot is in addition to the 2 x16 for Graphics cards if I read this correctly (2x MPX, which is 1x16 and 1x8, and then in addition 1x16 and 2x8, and then one x4 with Apple's IO card).
So unless you have something else, there an x16 slot for four x4 drives. Of course, bifurcation would mean you have to use x16 for four drives, while a bridge allows sticking as many drives as you want into any PCIe slot. I'm not criticizing this product for existing, I'm just wondering why you'd opt for a Bridge if you have the slot free.
All because it's in the CPU, doesn't mean Apple can be bothered to support it. eg - for displays, Displayport MST has been around for an age, but apple don't support it.
The latest Mac Pros don't have enough PCIe lanes coming from the CPU, even if it is a respectable 64 lanes. Throwing the extra 24 from the PCH doesn't do much for this equation since that is bottlenecked by the DMI uplink. The main driving factor for the need of so many switches is Thunderbolt which gets fed four PCIe lanes per pair of ports.
There is no guarantee that the slot this expansion card goes into has all 16 physical lanes routed to it. Thus to be able to guarantee usage of all four M.2 lots is that it has to sit behind a PCIe switch, even if it takes a performance hit due to the bandwidth reduction. (There is a sliver of hope for RAID1 writes as the PCIe switch might be able to duplicate the write command there instead of having it sent individual for each M.2 drive. This comes down to the specific PCIe switch used and software support.)
The use of a switch also permits it to be used in the previous generation of Mac Pros that only have PCI 2.0 rated slots while allowing the full bandwidth of a single M.2 to be used. IE 4x PCIe 3.0 M.2 -> 8x PCIe 2.0 slot
Would be interesting to see actual real-world performance benchmarking on something like this, especially as compared to regular consumer/prosumer SSD drives (max storage capacity aside...)
6 gig per sec... And now let's return to the real world where formatted read/write in c/c++ is 2-3 ***megabytes*** per second, and in most Fortran 50-100. Unformatted I/O in C, Python etc faster but still 10 MB/s, in Fortran 100-300 MB/s. Only HDF and Silverfrost Fortran can reach 4 GB/s. Funny, even own tools of HDF groplups like H5dump converts H5 files to text with turtle speed 1/3000 of the power of this drive with 1/1000 for unformatted output. Hahahaha
Very underwhelming. Its max sequential read speed is ~500 MB/s slower than just *two* Aura P12 SSDs, while its max sequential write speed is just ~775 MB/s faster than two Aura P12 SSDs (all set in RAID 0 mode apparently). No word about read/write IOPS either, which suggests they are abysmal, i.e. probably worse than those of the individual Aura P12 SSDs.
Neither the seq. read speed nor the seq. write speed is bottlenecked from the PCIe 3.0 x8 link (the read is certainly not bottlenecked at all), so where is the bottleneck? In the ASM2824 switch? I have no idea, can a PCIe switch become a bottleneck? Maybe because they weirdly had it kill half the available PCIe lanes (16 incoming lanes from the SSDs, 8 outgoing lanes to the PCIe slot)? Or is OWC's software RAID implementation the bottleneck? Or is MacOS itself the bottleneck? I think I will lose sleep in the coming days contemplating this question :)
This is so stupid. I raided P3700 nvme ssds 5 years ago .... and now, when apple has gotten some decent storage, the speeds are "INSANE". You can get over half this speed with ONE PCI-e 4.0 drive today. Fucking apple retards. People are SO stupid.
That said, some of Apple’s gear has legs on. I recently retrofitted a 2013 MacBook Air, Apple’s cheapest laptop, with an Adata m.2 nvme, and got 1300 mb/s rw. Owner is very happy with it. I don’t know many other 7-year old laptops still in regular use for work.
This OWC card has a PCIe switch, so it doesn't rely on the host motherboard supporting PCIe bifurcation and being able to provide four lanes per SSD. It'll work in any x8 or x16 slot and isn't limited to two M.2 drives when used in an x8 slot.
Yeah, I was wondering the same... bought a Sabrent 1TB Gen4x4 drive; it does a steady 3.9GB Write & 4.3GB Read in a new 3900X/X570 build and it only set me back about $167 after taxes on Amazon (had a $11 coupon which pretty much covered the tax bill). 2TB must be getting scare though, price has jumped up to ~$500 on Amazon and I'm sure it was around $350 when I was shopping for the 1TB early last week. Some X570 boards are bundling in a 4 NVMe slot Gen4x16 add-in card... whoopie!
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
24 Comments
Back to Article
MenhirMike - Monday, December 23, 2019 - link
Curious why they use a PCIe Switch - does the Mac Pro not support bifurcation? The Xeon definitely has enough PCIe lanes, so using something like a $50 ASUS M.2 X16 card would be a cheaper option (and one that gives x16 instead of x8 speed), especially since the RAID is software-side anyway.That said, the price isn't bad for a card that uses a PCIe switch, I just question why you'd buy a card with a PCIe Switch if you don't have a lack of lanes available - the Cascade Lake Xeon W's have 64 PCIe 3.0 lanes across the board.
GreenReaper - Monday, December 23, 2019 - link
The Mac Pro has one PCIe x16 slot, but two PCIe x8 slots: https://www.apple.com/uk/mac-pro/specs/Realistically, you probably won't be using 16x of bandwidth, so using the 8x slots is more efficient. It means you could use four cards at least at half capacity, which isn't possible with a 2 x 4x card.
MenhirMike - Monday, December 23, 2019 - link
Yeah, but the x16 slot is in addition to the 2 x16 for Graphics cards if I read this correctly (2x MPX, which is 1x16 and 1x8, and then in addition 1x16 and 2x8, and then one x4 with Apple's IO card).So unless you have something else, there an x16 slot for four x4 drives. Of course, bifurcation would mean you have to use x16 for four drives, while a bridge allows sticking as many drives as you want into any PCIe slot. I'm not criticizing this product for existing, I'm just wondering why you'd opt for a Bridge if you have the slot free.
MenhirMike - Monday, December 23, 2019 - link
To answer myself, the Apple Afterburner card is meant for that free x16 slot. That makes sense.dontlistentome - Monday, December 23, 2019 - link
All because it's in the CPU, doesn't mean Apple can be bothered to support it. eg - for displays, Displayport MST has been around for an age, but apple don't support it.Kevin G - Thursday, December 26, 2019 - link
The latest Mac Pros don't have enough PCIe lanes coming from the CPU, even if it is a respectable 64 lanes. Throwing the extra 24 from the PCH doesn't do much for this equation since that is bottlenecked by the DMI uplink. The main driving factor for the need of so many switches is Thunderbolt which gets fed four PCIe lanes per pair of ports.There is no guarantee that the slot this expansion card goes into has all 16 physical lanes routed to it. Thus to be able to guarantee usage of all four M.2 lots is that it has to sit behind a PCIe switch, even if it takes a performance hit due to the bandwidth reduction. (There is a sliver of hope for RAID1 writes as the PCIe switch might be able to duplicate the write command there instead of having it sent individual for each M.2 drive. This comes down to the specific PCIe switch used and software support.)
The use of a switch also permits it to be used in the previous generation of Mac Pros that only have PCI 2.0 rated slots while allowing the full bandwidth of a single M.2 to be used. IE 4x PCIe 3.0 M.2 -> 8x PCIe 2.0 slot
boeush - Monday, December 23, 2019 - link
Would be interesting to see actual real-world performance benchmarking on something like this, especially as compared to regular consumer/prosumer SSD drives (max storage capacity aside...)SanX - Monday, December 23, 2019 - link
6 gig per sec... And now let's return to the real world where formatted read/write in c/c++ is 2-3 ***megabytes*** per second, and in most Fortran 50-100. Unformatted I/O in C, Python etc faster but still 10 MB/s, in Fortran 100-300 MB/s. Only HDF and Silverfrost Fortran can reach 4 GB/s. Funny, even own tools of HDF groplups like H5dump converts H5 files to text with turtle speed 1/3000 of the power of this drive with 1/1000 for unformatted output. HahahahaSlash3 - Tuesday, December 24, 2019 - link
Alright.patmanRR - Tuesday, December 24, 2019 - link
No, it is not 3mb/s.timecop1818 - Tuesday, December 24, 2019 - link
Lulz, 8TB model is half price of what Apple wants for 8TB upgrade on the trashPro.Santoval - Tuesday, December 24, 2019 - link
Very underwhelming. Its max sequential read speed is ~500 MB/s slower than just *two* Aura P12 SSDs, while its max sequential write speed is just ~775 MB/s faster than two Aura P12 SSDs (all set in RAID 0 mode apparently). No word about read/write IOPS either, which suggests they are abysmal, i.e. probably worse than those of the individual Aura P12 SSDs.Neither the seq. read speed nor the seq. write speed is bottlenecked from the PCIe 3.0 x8 link (the read is certainly not bottlenecked at all), so where is the bottleneck? In the ASM2824 switch? I have no idea, can a PCIe switch become a bottleneck? Maybe because they weirdly had it kill half the available PCIe lanes (16 incoming lanes from the SSDs, 8 outgoing lanes to the PCIe slot)? Or is OWC's software RAID implementation the bottleneck? Or is MacOS itself the bottleneck? I think I will lose sleep in the coming days contemplating this question :)
s.yu - Tuesday, December 24, 2019 - link
Ah, I was wondering that.The_Assimilator - Tuesday, December 24, 2019 - link
Hey look it's another piece of overpriced junk from OWC. Must be nice having a captive market that's also stupid and rich.Korguz - Tuesday, December 24, 2019 - link
then just dont buy it... pronlem solvedKorguz - Tuesday, December 24, 2019 - link
problem solved....Spunjji - Friday, December 27, 2019 - link
*facepalm*Korguz - Saturday, December 28, 2019 - link
huh???JKJK - Tuesday, December 24, 2019 - link
This is so stupid. I raided P3700 nvme ssds 5 years ago .... and now, when apple has gotten some decent storage, the speeds are "INSANE".You can get over half this speed with ONE PCI-e 4.0 drive today.
Fucking apple retards.
People are SO stupid.
Tomatotech - Tuesday, December 24, 2019 - link
This isn’t from Apple.That said, some of Apple’s gear has legs on. I recently retrofitted a 2013 MacBook Air, Apple’s cheapest laptop, with an Adata m.2 nvme, and got 1300 mb/s rw. Owner is very happy with it. I don’t know many other 7-year old laptops still in regular use for work.
sandtitz - Wednesday, December 25, 2019 - link
ASRock Ultra Quad from yesteryear had double the performance. Price without drives: $80 vs $250. Why would any PC owner choose this OWC model?Billy Tallis - Thursday, December 26, 2019 - link
This OWC card has a PCIe switch, so it doesn't rely on the host motherboard supporting PCIe bifurcation and being able to provide four lanes per SSD. It'll work in any x8 or x16 slot and isn't limited to two M.2 drives when used in an x8 slot.Freeb!rd - Thursday, December 26, 2019 - link
Yeah, I was wondering the same... bought a Sabrent 1TB Gen4x4 drive; it does a steady 3.9GB Write & 4.3GB Read in a new 3900X/X570 build and it only set me back about $167 after taxes on Amazon (had a $11 coupon which pretty much covered the tax bill). 2TB must be getting scare though, price has jumped up to ~$500 on Amazon and I'm sure it was around $350 when I was shopping for the 1TB early last week. Some X570 boards are bundling in a 4 NVMe slot Gen4x16 add-in card... whoopie!flyingpants265 - Sunday, December 29, 2019 - link
Remember Gigabyte i-RAM? That was so cool, but I guess we don't even need that anymore.