AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion
I think you're not trying very hard.
Conservative rasterization, something Pascal also has. Tier 3 seems to add precision, a few more modes, and options to handle geometry. Does that sound like your ticket to the missing 169mm2?
Various additional options to tiled resources. That's essentially memory management stuff. Not a likely candidate either.
Same think for resource binding tiers: more management stuff.
What kind of other features are we talking about that's new compared to Polaris?
It's funny, after AMD has gone through "something is clearly not right" convulsive revival, then finally lunched the card, layed out the "official" performance guidelines vis-a-vis competition, etc, people are still grasping at "wait for magic drivers" straws. I guess if you repeat a mantra long enough, it becomes ingrained.
The smoking fun that something unexpected went wrong is not perf/W or perf/mm but gaming perf/TFLOPS.
AMD was always lacking a bit in this department, but Vega is off the charts bad. 13 TFLOPS to compete against 9 TFLOPS.
DavidGraham
Veteran
Which did almost nothing to their rated TDP, AMD rates the air cooled @295W (despite lower clocks than Vega FE), Water cooled is rated @345W. Still incredibly high.
We're well past hanging to false hopes now, several outlets do warn not to expect miraculous gains, besides AMD's numbers are logically with these features enabled .
http://www.anandtech.com/show/11680...-vega-64-399-rx-vega-56-launching-in-august/3
http://www.tomshardware.com/news/amd-radeon-rx-vega-64-specs-availability,35112.html
Last edited:
The 1080 Ti FE tops out at 250 watts at stock settings. The 1080 FE is at 180 watts.
DavidGraham
Veteran
This was one of the greatest disappointments I've ever had in B3D, I couldn't believe how some of the best hardcore and technically oriented people repeat and cling to such an obvious fantasy after seeing the results of Vega FE.
gamervivek
Regular
The cut down version does decently in the power department with 210W for a boost clock of 1471. AMD still aren't out of the woods in the clockspeed department requiring 1.2V for 1.6Ghz while Pascal can touch 2Ghz under 1V.
My guess is that Vega is likely bandwidth bound as the tiled rasterizer is not enabled. Vega has slightly less bandwidth than Fiji and significantly more computational (and geometry) throughput. Also I would guess that Vega L2 cache is trashing a bit more in graphics workloads, since Vega brings ROP caches under L2. Tiled rasterizer should significantly help with rasterizer cache trashing (bin triangles first to tiles -> much better area (cache) locality when rasterizing). Vega RX reviews (today) also reveal that L2 cache has been doubled from 2 MB to 4 MB. This means that L2 cache flushes now cost more. They really need new driver code to avoid bottlenecks and take advantage of the new hardware.
I would assume that the final RX drivers are released on 14th. AMD recently released a tool that allows you to record frames and inspect all the barriers (cache flushes and stall types) of their GPUs. When there is an official developer driver supporting Vega, we could do some captures and inspect how the behavior has changed since Fiji/Polaris.
Link to this awesome tool (I have waited for long time to get tools like this on PC):
http://gpuopen.com/gaming-product/radeon-gpu-profiler-rgp/
I would assume that the final RX drivers are released on 14th. AMD recently released a tool that allows you to record frames and inspect all the barriers (cache flushes and stall types) of their GPUs. When there is an official developer driver supporting Vega, we could do some captures and inspect how the behavior has changed since Fiji/Polaris.
Link to this awesome tool (I have waited for long time to get tools like this on PC):
http://gpuopen.com/gaming-product/radeon-gpu-profiler-rgp/
I'd say a 30% increase in memory bandwidth is exactly what Vega needs. It is clearly held back by lower than Fiji memory bandwidth while 50% or more higher core clocks.
Yet you continue to post in this thread? Why not just wait for the official reviews?
Exactly. If memory bandwidth was always going to be a core issue they would have gone with 4 stacks (maybe 3?) again of cheaper (4x4 or 4x2) and slower memory and still had way more bandwidth than Fiji. They designed Vega to have better memory bandwidth utilization and its clear its not working properly right now. The engineers saying that its disabled just confirms it.
They should have done it in the last 12 months. It sounds irrational that you cant write release ready drivers in this timeframe...
DavidGraham
Veteran
Which means AMD numbers are from the DSBR enabled driver. They didn't measure those data with it enabled only to release "final" performance estimates with it disabled!
Ever heard of best case scenario?
GTX 1060 was beating RX 480 by 12% in TechSpot launch review. They ran benchmarks again with newest driver (at RX 580 launch). The difference had dropped to 1%. That's a 11% increase. And that's been AMDs track record for ages. There's plenty of reviews stating the same thing for different AMD GPUs.
Link: https://www.techspot.com/review/1393-radeon-rx-580-vs-geforce-gtx-1060/page8.html
Vega is the biggest change to GCN architecture since GCN launch. Would it be wrong to assume that AMD needs more time than Polaris required to get the drivers up to peak performance? AMD today officially told us that the tiled rasterizer has been disabled in the current drivers. Vega FE is likely bandwidth starved. Expecting 10%+ performance gains during the first year of driver upgrades is plausible. They managed to do it with Polaris which was a much smaller upgrade. Now they have a completely new rasterizer and new ROP cache hierarchy.
Obviously it's silly if someone expects a 30%+ perf bump over Vega FE in 2 weeks
Have you experienced anything similar to the weird texture fillrate results and unusually low effective memory bandwidth in your work with Vega?
Those are all very old titles, who knows when that was tested, maybe it was tested a year ago and something broke and caused them to have to rework it. We don't know
Yep, also once games start using the new features like rapid packed math that can give huge gains as well. Futuremark found over 15% gain in the benchmark they are creating that tests it.