I basically tried to assemble a bit of an AMD "Stompinator", a powerhouse that still yields some efficiency, while avoiding single massive bottlenecks (not done yet due to PCIe limits).
It roughly chunks out something like 110000-120000 Cr/day, but also consumes 310W.
That yields about 371Cr/W and might be the presently fastest AMD based Host on Einstein.
As it is severely limited by the PCIe lanes and also by CPU performance, next month I'll modify it one last time to afford PCIe 2.0 16/16/8 with a different board (AM3+) and CPU (likely 6core FX-6300).
Wattage will increase a tad but performance should increase by ~40% as is.
In the end, it may run 2x HD7850 and 1x HD7790 or 3x HD7850.
That should place the rig in the Top 5 I hope, without busting my personal Wattage limits (absolute max. I want to invest = 500W, less is better and looks like I'll be good on that).
Are you sure you are limited by CPU? The total CPU consumption of Einstein+Linux (The overhead of Linux is less than 1% of a CPU core.) to max out my HD 7970 on Ubuntu is only 75% of one of my FX-8120 cores (might as well downclock to 1Ghz lol).
I initially thought so too looking just at the CPU load, but it's the entire Infrastructure around the CPU as well.
There combined, it seems the Socket FM2 is at a disadvantage over Socket AM3+ of about 15-20% alone.
Additionally, the A10 CPU core is a bit slower (IPC) than AMD latest AM3+ Vishera core and carries much less Cache, all of which reduces performance and increases I/O loads even further.
Therefor, I expect around 30% more performance just by switching the Socket from FM2 to AM3+ with a comparably clocked Vishera CPU...
FM2 Socket just seems to lack the I/O throughput capabilities if the AM3+, plus it doesn't offer alot of PCIe lanes even with the latest chipsets. AM3+ Chipsets aren't perfect either, but at least far better equipped than FM2 in those points.
Since all that is just based on my obvservation and analysis... I'll have to physically assemble the hardware next month and see actual results.
If you're buying a new AM3+ board anyway, would it make sense to go straight for the one model available with PCIe for AMD? Although the bus between chipset and CPU couldn't make use of this added bandwidth anyway.. it's probably much more useful for inter-GPU comunication (as needed in games).
On the other hand: if performance is limited by activity bursts on the PCIe bus, but with large pauses in between, the overall bandwidth between CPU and chipset might still be enough, while the faster PCIe could reduce the time it takes to finish individual bursts, so overall latency could be improved. That's really difficult to guess without further knowledge, though.
If you mean the one model with optimal PCIe configuration (in my case for 3 GPUs), yes of course. This is absolutely paramount to achieve desired performance and avoid severe PCIe bottlenecks (at least on Einstein).
This is what I did not consider, as I didn't double-check the board specifications directly on the manufacturer's websites.
Often the Boards are listed for example as having 3 PCIe 2.0 16x slots in hardware stores (technically correct, considering individual PCIe Slots), but they won't work at 16/16/16 when two or even all Slots are occupied.
Plus, I initially didn't think the PCIe bandwidth would have such a tremendous impact.
By migrating from FM2 to AM3+ I'll move from effectively 8/8/4 PCIe 2.0 (which seems the maximum for all FM2 Boards) to effectively 16/16/8 PCIe 2.0 of the AM3+ 990FX Chipset (if I remember correctly), even there I found only a few boards that offered so many functional lanes in total.
There are a few AM3+ boards with more slots and advertised as 16/16/16/8 PCIe 2.0, but the manufacturers specification websites always state that Slots 1/2 will throttle to 8/8 when Slots 3/4 are used due to shares lanes).
AM3+ Boards with even more PCIe slots (a board with 7 PCIe 2.0 exists) have them arranged in ways that only max. 4 are actually usable with typical Cards (just about all take 2 slots, even most smaller ones due to dimensions of the cooling package). Then above limitation strikes often again due to shared PCIe lanes, some even shut down some other slots to take up their lanes.
Basically it reminds me of good(bad) old PCI IRQ sharing times.
Quad CrossFireX for example sounds great, but often that ends up being 8/8/8/4 or even 8/8/4/4 config.
That's what got me in the end, hence another reconfiguration needed :p
This time I'll try Linux again first - last one failed on a RAID config I had setup - with Win7 as a backup option. I had a glance at intel platforms which offered great PCIe 3.0 configurations, but the price wasn't very friendly ;)
Plus, I see it as a challenge to build the fastest AMD platform on Einstein and later see how it does on other projects :)
I was talking about the only AMD mainboard offering PCIe 3 slots, probably using a PLX bridge chip (like on many Intels). But considering it's listed in a single shop and is not actually availalbe there.. you might as well forget about it :p
RE: My setup has changed
)
Are you sure you are limited by CPU? The total CPU consumption of Einstein+Linux (The overhead of Linux is less than 1% of a CPU core.) to max out my HD 7970 on Ubuntu is only 75% of one of my FX-8120 cores (might as well downclock to 1Ghz lol).
Woops, I meant to say HD
)
Woops, I meant to say HD 7950*; I don't have a HD 7970.
I initially thought so too
)
I initially thought so too looking just at the CPU load, but it's the entire Infrastructure around the CPU as well.
There combined, it seems the Socket FM2 is at a disadvantage over Socket AM3+ of about 15-20% alone.
Additionally, the A10 CPU core is a bit slower (IPC) than AMD latest AM3+ Vishera core and carries much less Cache, all of which reduces performance and increases I/O loads even further.
Therefor, I expect around 30% more performance just by switching the Socket from FM2 to AM3+ with a comparably clocked Vishera CPU...
FM2 Socket just seems to lack the I/O throughput capabilities if the AM3+, plus it doesn't offer alot of PCIe lanes even with the latest chipsets. AM3+ Chipsets aren't perfect either, but at least far better equipped than FM2 in those points.
Since all that is just based on my obvservation and analysis... I'll have to physically assemble the hardware next month and see actual results.
If you're buying a new AM3+
)
If you're buying a new AM3+ board anyway, would it make sense to go straight for the one model available with PCIe for AMD? Although the bus between chipset and CPU couldn't make use of this added bandwidth anyway.. it's probably much more useful for inter-GPU comunication (as needed in games).
On the other hand: if performance is limited by activity bursts on the PCIe bus, but with large pauses in between, the overall bandwidth between CPU and chipset might still be enough, while the faster PCIe could reduce the time it takes to finish individual bursts, so overall latency could be improved. That's really difficult to guess without further knowledge, though.
MrS
Scanning for our furry friends since Jan 2002
If you mean the one model
)
If you mean the one model with optimal PCIe configuration (in my case for 3 GPUs), yes of course. This is absolutely paramount to achieve desired performance and avoid severe PCIe bottlenecks (at least on Einstein).
This is what I did not consider, as I didn't double-check the board specifications directly on the manufacturer's websites.
Often the Boards are listed for example as having 3 PCIe 2.0 16x slots in hardware stores (technically correct, considering individual PCIe Slots), but they won't work at 16/16/16 when two or even all Slots are occupied.
Plus, I initially didn't think the PCIe bandwidth would have such a tremendous impact.
By migrating from FM2 to AM3+ I'll move from effectively 8/8/4 PCIe 2.0 (which seems the maximum for all FM2 Boards) to effectively 16/16/8 PCIe 2.0 of the AM3+ 990FX Chipset (if I remember correctly), even there I found only a few boards that offered so many functional lanes in total.
There are a few AM3+ boards with more slots and advertised as 16/16/16/8 PCIe 2.0, but the manufacturers specification websites always state that Slots 1/2 will throttle to 8/8 when Slots 3/4 are used due to shares lanes).
AM3+ Boards with even more PCIe slots (a board with 7 PCIe 2.0 exists) have them arranged in ways that only max. 4 are actually usable with typical Cards (just about all take 2 slots, even most smaller ones due to dimensions of the cooling package). Then above limitation strikes often again due to shared PCIe lanes, some even shut down some other slots to take up their lanes.
Basically it reminds me of good(bad) old PCI IRQ sharing times.
Quad CrossFireX for example sounds great, but often that ends up being 8/8/8/4 or even 8/8/4/4 config.
That's what got me in the end, hence another reconfiguration needed :p
This time I'll try Linux again first - last one failed on a RAID config I had setup - with Win7 as a backup option. I had a glance at intel platforms which offered great PCIe 3.0 configurations, but the price wasn't very friendly ;)
Plus, I see it as a challenge to build the fastest AMD platform on Einstein and later see how it does on other projects :)
I was talking about the only
)
I was talking about the only AMD mainboard offering PCIe 3 slots, probably using a PLX bridge chip (like on many Intels). But considering it's listed in a single shop and is not actually availalbe there.. you might as well forget about it :p
MrS
Scanning for our furry friends since Jan 2002
Ah, gotcha... I've also
)
Ah, gotcha...
I've also looked for it but it is indeed nowhere to be purchased in Europe, don't know what's the reason for that.