Message boards :
Number crunching :
RX 480 - A serious contender?
Message board moderation
Author | Message |
---|---|
Michamus Send message Joined: 31 Aug 16 Posts: 13 Credit: 2,502,762 RAC: 0 |
Bear in mind on the following, I am looking at completion times and not really concerned with credits. I have been perusing machine lists and haven't been able to find anyone using the new RX 480. I did some searches and that also yielded little of value. Has anyone been using it and how does it fare? The numbers seem to give it the appearance of being equivalent to a GTX 1070 on W/TFLOPS and seems to be pretty cheap on the $/TFLOPS ratio. GPU Model - Dollar/TFLOPS - TFLOPS/KwH GTX 1080 - $78.99/TFLOPS - 45.71 R9 Fury X - $46.50/TFLOPS - 31.28 980 TI     - $73.69/TFLOPS - 22.53 RX 480 8G - $51.42/TFLOPS - 38.89 RX 480 4G - $44.57/TFLOPS - 38.89 This is obviously based on the theoretical limits of the GPU, which doesn't seem to correlate directly with S@H performance. For instance, these are my findings on my two main GPUs based on a random sampling of validated SoG tasks: 980 TI     - 786.53s - 250W - 5.632 TFLOPS R9 Fury X - 732.33s - 275W - 8.602 TFLOPS So, the 980 TI seems to be completing SoG tasks at 107.40% the time required for the R9 Fury X. However, it also seems to be eating 10% more power to get that performance. This gives me the impression that the apparent 52.73% increase in TFLOPS doesn't exactly correlate to that much faster task completion. So, you can imagine my hesitation. I'm wondering if there's some kind of architectural aspect at play here. Much like an older model CPU having a similar clock speed, but the newer architecture allows for more instructions per clock. My suspicion is due to the fact that the R9 Fury X had a greater range in the sample than the 980 TI. 980 TI    - low 646.91s - high 1066.58s - mode 690s R9 Fury X - low 196.96s - high  870.10s - mode 840s This greater consistency on the 980 TI leads me to wonder what could be going on. The R9 Fury X seems to be completing some tasks at 1/3rd the fastest time for the 980 TI, but the mode for the R9 seems to be 21.74% above the 980 TI. So strange. I wonder if the number of shading units are contributing to this? With this in mind, would the newer architecture RX 480 provide even more performance than the 1070, given its greater number of shading units? TL;DR: If you have an RX 480 that's a dedicated cruncher, I would love for you to link me your machine so I can see how fast it's working.[img][/img] |
Keith Myers Send message Joined: 29 Apr 01 Posts: 13164 Credit: 1,160,866,277 RAC: 1,873 |
It's too bad the RX480 isn't showing up in Shaggie76 tables yet to see what the relative performance on SETI tasks is. There has been some good discussion over at Einstein regarding the relative performance of the Pascal and Polaris architectures, but that has been specifically targeted at Einstein work which unfortunately has little comparison to SETI work. I have seen a pretty good writeup on the relative OpenCL performance comparisons between Pascal and Polaris architectures over at Phoronix. That has relevance since the greater mix of current and future work here at SETI is going to be OpenCL based. It looks like the RX480 is coming up short when compared to the GTX 1070 which you wanted a direct comparison to in your post. Seti@Home classic workunits:20,676 CPU time:74,226 hours A proud member of the OFA (Old Farts Association) |
Shaggie76 Send message Joined: 9 Oct 09 Posts: 282 Credit: 271,858,118 RAC: 196 |
Actually the RX480 is in my graphs under "Ellesmere" -- it's infuriating but it seems like the CUDA device identifier insists on naming the family not the model which tragically will mean this includes RX470's etc -- however if you go back in that thread you'll find my earliest data for Ellesmere was consistently mediocre even with the RX480 was the only part in circulation. |
Shaggie76 Send message Joined: 9 Oct 09 Posts: 282 Credit: 271,858,118 RAC: 196 |
If you want to find data here's a bunch of host IDs with an Ellesmere card in them: 5042538, 5118968, 5540457, 6316223, 6426017, 6790643, 6796592, 6823513, 6842350, 6858793, 6991626, 7068732, 7266357, 7274252, 7307367, 7311721, 7422707, 7433506, 7435521, 7444081, 7454760, 7454913, 7492259, 7626971, 7735231, 7821951, 7838567, 7839205, 7853018, 7854233, 7876296, 7921259, 7936757, 7938323, 7959852, 7983958, 8003454, 8007708, 8012031, 8020084, 8032704, 8035960, 8040474, 8043091, 8045906, 8054008, 8062297, 8064103, 8066283, 8068039, 8068803, 8073564, 8074922, 8074946, 8077581, 8078707, 8081369, 8082476, 8084874, 8085722, 8086878, 8087588, 8087974, 8089012, 8089252, 8092494, 8093154, 8093457, 8094333, 8094761. BTW I started writing a bunch of scripts to try to answer the same question you're thinking of. The source is all there if you want to modify it to suit whatever metric you'd prefer. FWIW when I was done I went out and bought 3 GTX 1070's instead, though. |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
Actually the RX480 is in my graphs under "Ellesmere" -- it's infuriating but it seems like the CUDA device identifier insists on naming the family not the model which tragically will mean this includes RX470's etc -- however if you go back in that thread you'll find my earliest data for Ellesmere was consistently mediocre even with the RX480 was the only part in circulation. BOINC 7.6.23+ uses a new name detection method for Radeon GPUs so that the BOINC developers don't have to hard code the GPU names. I don't know if you are taking that into account when searching for GPUs to chart. This is how my R9 390X is displayed with the new code vs the old. BOINC 7.6.23+: AMD AMD Radeon (TM) R9 390 Series (8192MB) OpenCL: 2.0 BOINC 7.6.22: AMD Grenada (8192MB) OpenCL: 2.0 SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Keith Myers Send message Joined: 29 Apr 01 Posts: 13164 Credit: 1,160,866,277 RAC: 1,873 |
I don't see 7.6.23 on the download site. Is your post a typo and you meant the 7.6.33 Beta? Seti@Home classic workunits:20,676 CPU time:74,226 hours A proud member of the OFA (Old Farts Association) |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14650 Credit: 200,643,578 RAC: 874 |
I don't see 7.6.23 on the download site. Is your post a typo and you meant the 7.6.33 Beta? He said 7.6.23+ That's shorthand for "7.6.23 and anything later". Some intermediate builds can be found, but in general BOINC only lists one 'recommended' and one 'development - use for testing only' version at a time. |
Shaggie76 Send message Joined: 9 Oct 09 Posts: 282 Credit: 271,858,118 RAC: 196 |
This is super exciting! Can't wait for it to be officially released. |
Keith Myers Send message Joined: 29 Apr 01 Posts: 13164 Credit: 1,160,866,277 RAC: 1,873 |
I don't see 7.6.23 on the download site. Is your post a typo and you meant the 7.6.33 Beta? Thanks for the clarity, Richard. I was hoping there might be an official 7.6.23 release I somehow missed. I tried 7.6.33 beta for a few weeks and had troubles with it and reverted back to stable 7.6.22. Seti@Home classic workunits:20,676 CPU time:74,226 hours A proud member of the OFA (Old Farts Association) |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14650 Credit: 200,643,578 RAC: 874 |
I don't see 7.6.23 on the download site. Is your post a typo and you meant the 7.6.33 Beta? "Official releases" are few and far between. If you could identify and report what your 'troubles' were with 7.6.33, maybe we could get that one out into public use - it has important fixes we first identified here. You can always find "invitations to test" intermediate versions here. |
Keith Myers Send message Joined: 29 Apr 01 Posts: 13164 Credit: 1,160,866,277 RAC: 1,873 |
I don't see 7.6.23 on the download site. Is your post a typo and you meant the 7.6.33 Beta? I'm pretty sure I DID report my experience with 7.6.33 beta on the alpha site. Or maybe in the Beta Questions and Answers forum. Upshot was if you are having issues, revert back to public release which I did and problems gone. Seti@Home classic workunits:20,676 CPU time:74,226 hours A proud member of the OFA (Old Farts Association) |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
Actually the RX480 is in my graphs under "Ellesmere" -- it's infuriating but it seems like the CUDA device identifier insists on naming the family not the model which tragically will mean this includes RX470's etc -- however if you go back in that thread you'll find my earliest data for Ellesmere was consistently mediocre even with the RX480 was the only part in circulation. For clarity I should have stated that the new detection method is only for non-CAL GPU/drivers. Older drivers that still report a CAL version will continue to display the hard coded names from BOINC like: AMD AMD Radeon HD 7870/7950/7970/R9 280/R9 280X series (Tahiti) (3072MB) driver: 1.4.1848 OpenCL: 1.2 Also it has been found, at least in Linux, that no board name is displayed for RX 400 GPUs. Do we have any Windows users with a RX 400's using a BOINC 7.6.23 or newer? SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Michamus Send message Joined: 31 Aug 16 Posts: 13 Credit: 2,502,762 RAC: 0 |
If you want to find data here's a bunch of host IDs with an Ellesmere card in them: 5042538, 5118968, 5540457, 6316223, 6426017, 6790643, 6796592, 6823513, 6842350, 6858793, 6991626, 7068732, 7266357, 7274252, 7307367, 7311721, 7422707, 7433506, 7435521, 7444081, 7454760, 7454913, 7492259, 7626971, 7735231, 7821951, 7838567, 7839205, 7853018, 7854233, 7876296, 7921259, 7936757, 7938323, 7959852, 7983958, 8003454, 8007708, 8012031, 8020084, 8032704, 8035960, 8040474, 8043091, 8045906, 8054008, 8062297, 8064103, 8066283, 8068039, 8068803, 8073564, 8074922, 8074946, 8077581, 8078707, 8081369, 8082476, 8084874, 8085722, 8086878, 8087588, 8087974, 8089012, 8089252, 8092494, 8093154, 8093457, 8094333, 8094761. I really appreciate the wealth of information that you've provided. The RX 480 performance does seem pretty disappointing when compared with the GTX 1070. I do find it interesting that the 750 TI has finally been displaced as the top Credit/WH by the GTX 1060. My graphs had placed the GTX 1060 as a pretty strong contender on the W/TFLOPS ratio, but the results seemed to point at the 1070 as the top contender on the current gen tech. I guess that just further shows that calculated TFLOPS don't translate well to S@H work. |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.