The memory interface of the new Navi cards can be problematic.
It will be interesting to see whether the evidence suggests that Einstein on Big Navi is well serviced by memory capabilities or a bit starved. Much depends on how well the Infinity cache serves in Einstein use.
As all three of the cards actually announced so far, 6800, 6800 XT, and 6900 XT apparently use exactly the same actual chip, with selective disabling used to lower the functional units offered and speed and power consumption allowed on the lower-rated cards, AMD had a choice available regarding memory.
The choice they appear to have made is to offer the full memory capability on all three cards. If indeed, the memory is strongly limiting for the Einstein case, this would push the optimum toward the lower-rated card.
My personal preference has shifted up one level, from the 6800 to the 6800 XT. It appears that in this case, AMD has plumped up the resources and limits quite a bit in going from the base to the XT. Unlike the 5700, I only see myself putting one in a box. One experience with super-fat high-power cards was enough for me.
By contrast, the additional increment up to the 6900 XT looks rather modest, the price increase considerable, and the availability yet more problematic
I think for Infinity cache to be utilized, the motherboard has to be the 500 series and the cpu needs to be the latest generation.
The new 6000 series gpus would most likely perform better on Einstein since it uses gddr6. By how much is the question. Not sure if the performance difference is worth the price difference between an 5700 xt and the 6800 xt.
I think for Infinity cache to be utilized, the motherboard has to be the 500 series and the cpu needs to be the latest generation.
Your comment implies that the Big Navi Infinity cache is not an onchip resource.
But the comments about it which I have seen, including marketing claims, are incoherent unless it is an onchip resource. Specifically it is claimed to allow suitable use of the abundant supply of compute resources despite a seemingly narrow 256 bit memory path even on top model 6900 XT.
Do you have a source you can give us for either the general prospect that this is an off-chip resource, or for the specific limitations as to motherboard and CPU?
If those are true, and if Infinity cache is important to Einstein Big Navi performance, I certainly should not purchase any Big Navi card at all.
On a quick review of things I see on the Internet, it appears to me that you are conflating AMD "Smart Memory Access" with the Navi2 Infinity cache, which is a whole different thing. SMA maps a bigger chunk of GPU RAM into the CPU memory address space at a given moment than previously.
Here at Einstein, the Gamma-Ray Pulsar GPU application would seem especially unlikely to be helped much by SMA. That appplication needs very little CPU support in general, and tolerates old slow PCI buses on generations old motherboards with very little harm.
On the other hand, the remarkable productivity of the VII card on Einstein GRP rather hints that on-card memory access is very important, as the most obvious difference in that card from others for which it has a surprising performance edge is in the memory system, which uses an uncommon and very expensive memory. So if Infinity cache works well or not would be a big swing for GRP performance.
So far as simple VRAM size is concerned, the Gravity Wave application here at Einstein seems likely to benefit from the extra-large (16 GB) VRAM size on the Big Navi cards which should easily allow 4X multiplicity. THAT combination probably would benefit from being paired with a modern and fast CPU.
With the 6900XT at default settings, I tried 1-3 WUs. Run times below were averages across some 10 WUs each without any CPU load, so not a large result base.
With the 6900XT at default settings, I tried 1-3 WUs. Run times below were averages across some 10 WUs each without any CPU load, so not a large result base.
1 WU = 240s 1247400c
2 WU = 420s 1425600c
3 WU = 630s 1425600c
The VII seems to remain the king,
Thanks for the report.
The first important point is that a Big Navi card actually was able to generate validating results on Einstein GRP work. Not a trivial point given the months-long problem with the first generation Navi cards.
I notice that in the stderr your card gets reported as gfx1030. I think this may refer to the parent chip, which I think at least initially is the same in all three of the announced levels (6800, 6800 Xt, and your 6900 XT).
I see that you are an experienced VII user. Can you give us any comparisons to VII for power consumption, fan noise, general stability, ...?
I'm actively watching for stock available to ship at retail for a particular 6800 XT card. Your report leaves me interested, but lowers the top end of my hopes.
I have a VII sitting in my garage. It was pretty troublesome for me in over a half year operation, especially compared the the pair of 570s I ran for a while after it, or the pair of 5700s I run in the same box currently.
If anyone here is interested in running an extra VII on Einstein, PM me. It was running on my primary daily use machine (the one I'm typing on now), so the problems I had eventually made me give up. I'd offer a very good deal to an Einstein user. (US shipping only--I'm allergic to filling out the export forms)
ExtraTerrestrial Apes
)
It will be interesting to see whether the evidence suggests that Einstein on Big Navi is well serviced by memory capabilities or a bit starved. Much depends on how well the Infinity cache serves in Einstein use.
As all three of the cards actually announced so far, 6800, 6800 XT, and 6900 XT apparently use exactly the same actual chip, with selective disabling used to lower the functional units offered and speed and power consumption allowed on the lower-rated cards, AMD had a choice available regarding memory.
The choice they appear to have made is to offer the full memory capability on all three cards. If indeed, the memory is strongly limiting for the Einstein case, this would push the optimum toward the lower-rated card.
My personal preference has shifted up one level, from the 6800 to the 6800 XT. It appears that in this case, AMD has plumped up the resources and limits quite a bit in going from the base to the XT. Unlike the 5700, I only see myself putting one in a box. One experience with super-fat high-power cards was enough for me.
By contrast, the additional increment up to the 6900 XT looks rather modest, the price increase considerable, and the availability yet more problematic
I think for Infinity cache to
)
I think for Infinity cache to be utilized, the motherboard has to be the 500 series and the cpu needs to be the latest generation.
The new 6000 series gpus would most likely perform better on Einstein since it uses gddr6. By how much is the question. Not sure if the performance difference is worth the price difference between an 5700 xt and the 6800 xt.
4BVCtZgPKCf3ewT7e8rLjpevGEJz
)
Your comment implies that the Big Navi Infinity cache is not an onchip resource.
But the comments about it which I have seen, including marketing claims, are incoherent unless it is an onchip resource. Specifically it is claimed to allow suitable use of the abundant supply of compute resources despite a seemingly narrow 256 bit memory path even on top model 6900 XT.
Do you have a source you can give us for either the general prospect that this is an off-chip resource, or for the specific limitations as to motherboard and CPU?
If those are true, and if Infinity cache is important to Einstein Big Navi performance, I certainly should not purchase any Big Navi card at all.
On a quick review of things I see on the Internet, it appears to me that you are conflating AMD "Smart Memory Access" with the Navi2 Infinity cache, which is a whole different thing. SMA maps a bigger chunk of GPU RAM into the CPU memory address space at a given moment than previously.
Yeah, SMA only made small
)
Yeah, SMA only made small differences on some gaming benchmarks and NVIDIA say their cards will be able to use this feature shortly.
Here at Einstein, the
)
Here at Einstein, the Gamma-Ray Pulsar GPU application would seem especially unlikely to be helped much by SMA. That appplication needs very little CPU support in general, and tolerates old slow PCI buses on generations old motherboards with very little harm.
On the other hand, the remarkable productivity of the VII card on Einstein GRP rather hints that on-card memory access is very important, as the most obvious difference in that card from others for which it has a surprising performance edge is in the memory system, which uses an uncommon and very expensive memory. So if Infinity cache works well or not would be a big swing for GRP performance.
So far as simple VRAM size is concerned, the Gravity Wave application here at Einstein seems likely to benefit from the extra-large (16 GB) VRAM size on the Big Navi cards which should easily allow 4X multiplicity. THAT combination probably would benefit from being paired with a modern and fast CPU.
Seems like it’s software, but
)
Seems like it’s software, but it needs the hardware to support it.
https://www.somagnews.com/amd-smart-access-memory-feature-to-intel-processors/#:~:text=One%20of%20them%20is%20the%20AMD%20Smart%20Access,all%20of%20the%20memory%20on%20the%20graphics%20card.
sorry i am mistaken. i am
)
sorry i am mistaken. i am thinking about SAM SMA and incorrectly referring to Infinity Cache.
With the 6900XT at default
)
With the 6900XT at default settings, I tried 1-3 WUs. Run times below were averages across some 10 WUs each without any CPU load, so not a large result base.
The VII seems to remain the king,
P3d-cluster wrote: With the
)
Thanks for the report.
The first important point is that a Big Navi card actually was able to generate validating results on Einstein GRP work. Not a trivial point given the months-long problem with the first generation Navi cards.
I notice that in the stderr your card gets reported as gfx1030. I think this may refer to the parent chip, which I think at least initially is the same in all three of the announced levels (6800, 6800 Xt, and your 6900 XT).
I see that you are an experienced VII user. Can you give us any comparisons to VII for power consumption, fan noise, general stability, ...?
I'm actively watching for stock available to ship at retail for a particular 6800 XT card. Your report leaves me interested, but lowers the top end of my hopes.
I have a VII sitting in my garage. It was pretty troublesome for me in over a half year operation, especially compared the the pair of 570s I ran for a while after it, or the pair of 5700s I run in the same box currently.
If anyone here is interested in running an extra VII on Einstein, PM me. It was running on my primary daily use machine (the one I'm typing on now), so the problems I had eventually made me give up. I'd offer a very good deal to an Einstein user. (US shipping only--I'm allergic to filling out the export forms)
Thanks for the report
)
Thanks for the report P3D_Cluster.
Huge fan of your hardware btw! LOL. :D
Look forward to more news.