While impressive, you are throwing away most of the platform - half your memory channels, most of your PCIe, half the TDP headroom (this is is the real shocker), etc. - to fit in the size and layout constraints. ATX is just the wrong form-factor, and if you have to I feel like going single-socket is much more practical.
Actually, after looking at the Intel Ice Lake Xeon price list, where, for example, 2 dual-socket 16-core CPUs have the same price as 1 single-socket 32-core CPU, the main advantage of this MB is the better cooling of the cores (370 W TDP vs. 205 W TDP), which allows a much higher all-core clock frequency, e.g. 3.3 GHz vs. 2.9 GHz.
Because the Epyc Milan CPUs have a much higher clock frequency at the same number of active cores and the same power consumption, the only way to exceed their performance with Ice Lake Xeon is to allow a more generous TDP, like with this motherboard.
Because the Ice Lake Xeons with few cores, i.e. in the 16 ... 32 cores range, have now a more decent price, not much different from AMD, with a MB like this you might be able to match or exceed the Epyc price/performance, but with the price of a much higher power consumption.
One thing being thrown away is memory bandwidth. This board only has 8 of the 16 memory channels present which can be an issue for some select workloads. A solution to this that might still work is to leverage SO-DIMM slots. Put two slots end to end where you have a single full sized slot on this board. The chip package looks like this might be feasible from a trace layout stand point too. You won't get as much capacity per DIMM but the bandwidth side would be boosted by the channel increase.
I suspect that SuperMicro could support higher TDP chips on a board this physical size but they'd have to add more CPU power headers and a more robust VRM. Those take up board space but perhaps it is time to split the VRM on both sides of a socket for Ice Lake-SP. In terms of CPU cooling, liquid would be the only way to go outside of super high air flow server chassis. Being a standard ATX board does imply some usage outside of a typical rackmount server.
The one "nice" thing that has happened is that as software licensing has shifted from a per socket model to a per core model. The quotations are there as overall prices have generally increased moving to the per core model, these licensing models have not penalizing multisocket systems as much. Thus the the theory adding a second socket vs. a single larger socket system for greater performance could also lead to a better performance per dollar scenario. Have to really read the fine print and do the math as this is far from universal.
Regardless, this board is relatively dense and has its niche that'll pay for it. I wonder how many PCB layers it has.
You'd be surprised at the wide variety of requirements in enterprise. Those that don't need memory capacity/bandwidth or PCIe and just focused on raw compute will pay for a customized system. AMD announced on Milan that they included 4-channel and 6-channel memory interleaving for customers that need only 4-channel or 6-channel memory, saving those customers money in design and server density.
The general point is true, but AMD had 4 channel support before Milan (*you* wrote about that! https://www.anandtech.com/show/16529/amd-epyc-mila... Milan added 6 channel, I guess for folks migrating from Intel and keeping their memory.
I can see some use if you need AVX-512, but the lack of TDP Headroom seems like an issue then. Even for a workstation, a SSI-EEB Board wouldn't make the tower case THAT much bigger. Dual Socket ATX seems like a weird compromise, even more so than HEDT ITX.
I'd like to see SuperMicro show their chops by making the unicorn - a reasonably priced mini-ITX board with the various things people have been asking for. Trying to resist giving my shopping list here (but it has 4 RAM slots and multiple NVME slots on it).
Dunno how Supermicro could assess demand though. One way could be to have a poll of various options, and it requires a $10 deposit on your favoured option to vote. At the end of the poll, losing options are refunded to the voters, and these who put $10 into the winning option get a $15 voucher towards the purchase price. Just an idea.
Nevermind. As a server company, Supermicro and consumer value don't go together. Their customers buy IT time at $100 per hour for setting up servers.
As you can get a single socket EPYC motherboard (ASRock ROMED8-2T) in ATX format with 7 PCIe4.0 slots 8 memory channels 2x10G ethernet and more cores available than the dual Xeon's why bother with this board. This board looks as though it was created for bragging rights - "Hey Look - we can get 2 Xeons on an ATX board".
Too much was sacrificed by making this board ATX instead of EATX for very little purpose - almost all ATX style cases can take an EATX motherboard. If you are paying the money for a pair of Xeons then saving a few dollars by crippling their capabilities seems a very bad bargin.
This is odd - what is the recommended case for something like this? It seems like that would affect the market significantly. You would need wind tunnel cooling for this.
Rack servers are wind tunnels. Besides, 185W per socket isn't *that* much. A couple of 360 mm rads should be about right. Just make sure you have enough airflow.
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
22 Comments
Back to Article
alfalfacat - Monday, May 10, 2021 - link
While impressive, you are throwing away most of the platform - half your memory channels, most of your PCIe, half the TDP headroom (this is is the real shocker), etc. - to fit in the size and layout constraints. ATX is just the wrong form-factor, and if you have to I feel like going single-socket is much more practical.AdrianBc - Monday, May 10, 2021 - link
Going single-socket is of course much more practical ... if you are using an Epyc.This motherboard is the only way to have an ATX MB with Intel CPUs having a higher performance than an ATX MB with AMD CPUs.
The memory is the same as for a single-socket MB, so you lose only 3 PCIe slots, while doubling the number of CPU cores.
If you are happy with that, there may be some cases when this MB is an acceptable choice.
AdrianBc - Monday, May 10, 2021 - link
Actually, after looking at the Intel Ice Lake Xeon price list, where, for example, 2 dual-socket 16-core CPUs have the same price as 1 single-socket 32-core CPU, the main advantage of this MB is the better cooling of the cores (370 W TDP vs. 205 W TDP), which allows a much higher all-core clock frequency, e.g. 3.3 GHz vs. 2.9 GHz.Because the Epyc Milan CPUs have a much higher clock frequency at the same number of active cores and the same power consumption, the only way to exceed their performance with Ice Lake Xeon is to allow a more generous TDP, like with this motherboard.
Because the Ice Lake Xeons with few cores, i.e. in the 16 ... 32 cores range, have now a more decent price, not much different from AMD, with a MB like this you might be able to match or exceed the Epyc price/performance, but with the price of a much higher power consumption.
Kevin G - Monday, May 10, 2021 - link
One thing being thrown away is memory bandwidth. This board only has 8 of the 16 memory channels present which can be an issue for some select workloads. A solution to this that might still work is to leverage SO-DIMM slots. Put two slots end to end where you have a single full sized slot on this board. The chip package looks like this might be feasible from a trace layout stand point too. You won't get as much capacity per DIMM but the bandwidth side would be boosted by the channel increase.I suspect that SuperMicro could support higher TDP chips on a board this physical size but they'd have to add more CPU power headers and a more robust VRM. Those take up board space but perhaps it is time to split the VRM on both sides of a socket for Ice Lake-SP. In terms of CPU cooling, liquid would be the only way to go outside of super high air flow server chassis. Being a standard ATX board does imply some usage outside of a typical rackmount server.
The one "nice" thing that has happened is that as software licensing has shifted from a per socket model to a per core model. The quotations are there as overall prices have generally increased moving to the per core model, these licensing models have not penalizing multisocket systems as much. Thus the the theory adding a second socket vs. a single larger socket system for greater performance could also lead to a better performance per dollar scenario. Have to really read the fine print and do the math as this is far from universal.
Regardless, this board is relatively dense and has its niche that'll pay for it. I wonder how many PCB layers it has.
antonkochubey - Monday, May 10, 2021 - link
>A solution to this that might still work is to leverage SO-DIMM slots.RDIMM memory modules don't exist in SO-DIMM form-factor, and I'm not sure those CPUs even support UDIMM's.
bhtooefr - Thursday, May 13, 2021 - link
SORDIMMs do exist: https://industrial.apacer.com/en-ww/DRAM/DDR4-SORD...Ian Cutress - Monday, May 10, 2021 - link
You'd be surprised at the wide variety of requirements in enterprise. Those that don't need memory capacity/bandwidth or PCIe and just focused on raw compute will pay for a customized system. AMD announced on Milan that they included 4-channel and 6-channel memory interleaving for customers that need only 4-channel or 6-channel memory, saving those customers money in design and server density.Zizy - Monday, May 10, 2021 - link
The general point is true, but AMD had 4 channel support before Milan (*you* wrote about that! https://www.anandtech.com/show/16529/amd-epyc-mila... Milan added 6 channel, I guess for folks migrating from Intel and keeping their memory.mode_13h - Tuesday, May 11, 2021 - link
> You'd be surprised at the wide variety of requirements in enterprise.This could be aimed more at embedded or appliance applications.
MenhirMike - Monday, May 10, 2021 - link
I can see some use if you need AVX-512, but the lack of TDP Headroom seems like an issue then. Even for a workstation, a SSI-EEB Board wouldn't make the tower case THAT much bigger. Dual Socket ATX seems like a weird compromise, even more so than HEDT ITX.mode_13h - Tuesday, May 11, 2021 - link
> a SSI-EEB Board wouldn't make the tower case THAT much bigger.Even EATX would make more sense than this.
Tomatotech - Monday, May 10, 2021 - link
I'd like to see SuperMicro show their chops by making the unicorn - a reasonably priced mini-ITX board with the various things people have been asking for. Trying to resist giving my shopping list here (but it has 4 RAM slots and multiple NVME slots on it).Dunno how Supermicro could assess demand though. One way could be to have a poll of various options, and it requires a $10 deposit on your favoured option to vote. At the end of the poll, losing options are refunded to the voters, and these who put $10 into the winning option get a $15 voucher towards the purchase price. Just an idea.
Nevermind. As a server company, Supermicro and consumer value don't go together. Their customers buy IT time at $100 per hour for setting up servers.
edzieba - Monday, May 10, 2021 - link
If you want ITX boards with ludicrous socketry, then ASRockRack are who you want to look to rather than Supermicro.RU482 - Monday, May 10, 2021 - link
Supermicro dual socket ATX boards have been gimped in one way or another since at least X9. This isn't anything new.Duncan Macdonald - Monday, May 10, 2021 - link
As you can get a single socket EPYC motherboard (ASRock ROMED8-2T) in ATX format with 7 PCIe4.0 slots 8 memory channels 2x10G ethernet and more cores available than the dual Xeon's why bother with this board. This board looks as though it was created for bragging rights - "Hey Look - we can get 2 Xeons on an ATX board".Too much was sacrificed by making this board ATX instead of EATX for very little purpose - almost all ATX style cases can take an EATX motherboard. If you are paying the money for a pair of Xeons then saving a few dollars by crippling their capabilities seems a very bad bargin.
mode_13h - Tuesday, May 11, 2021 - link
> Too much was sacrificed by making this board ATX instead of EATX for very little purposeThere was probably a very specific customer or regulatory requirement they were trying to meet.
As a generic product, you're right that it doesn't make a lot of sense.
bananaforscale - Tuesday, May 11, 2021 - link
The only use case is probably needing AVX512. I can't see this making any sense otherwise.mode_13h - Tuesday, May 11, 2021 - link
Even then, it's an odd fit as most AVX-512 workloads probably want lots of memory bandwidth.strtj - Monday, May 10, 2021 - link
This is odd - what is the recommended case for something like this? It seems like that would affect the market significantly. You would need wind tunnel cooling for this.mode_13h - Tuesday, May 11, 2021 - link
> You would need wind tunnel cooling for this.They only support 2x 185 W CPUs. That's not bad, in terms of server cooling. For workstations, water cooling would probably make more sense.
bananaforscale - Tuesday, May 11, 2021 - link
Rack servers are wind tunnels. Besides, 185W per socket isn't *that* much. A couple of 360 mm rads should be about right. Just make sure you have enough airflow.msroadkill612 - Wednesday, May 12, 2021 - link
Intercore data hops of mm on an epyc single socket, could stretch to inches on this rig.It sounds a corner case, custom code solution.