AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

Ignoring clocks again?
I am afraid Vega might have even more changes muddling the comparison.

 

I'm not sure if the above or the following link http://www.fudzilla.com/news/graphics/43044-volta-expected-to-outperform-amd-s-radeon-rx-vega is more entertaining after all. Either way thank you both for the good laugh

 

I hear Ryzen 2 is going to trounce Hubbalake!

 

I will keep that in mind, but at home I have „only“ (I realize that this is quite a luxury in itself) an RX 480 and a Fury Nano. Hopefully I can fullfil the request over the weekend - at work things are quite busy atm, so no time there probably.

 

Some Vega 10 OpenCL benchmarks:
https://videocardz.com/67242/amd-vega-with-64-compute-units-spotted

 

4096 shaders with 1 200Mhz boost clock? How does that work out with the supposed 12.5 TFlops?

 

It's holding up pretty well against the 1080Ti especially clocked so low...if those result are real..

 

AMD had two Polaris 10 chips clocked at 800Mhz and 1266Mhz when the rumors with code names were being released last year. Not necessarily sandbagging.

Or they got ahead of themselves and have to downclock it in order to release it quickly which I think is unlikely.

 

Let's hope not.

 

Same here, if the final version is ~1600Mhz, it'd be slightly faster than those 1080Ti results which is promising.

 

1 600 Mhz? Let's not get carried away here

 

Why not, when the professional card is clocked around 1.5Ghz?

 

My Guess is that Vega 10 will between 1080 and 1080 TI in typical game scenarios and priced at 599. IOW, similar what FuryX was to 980 and 980 Ti.
But, please RTG, surprise us this time!

 

That's boost clock. I don't expect that the card will be able to stay at the boost clock during load. Some professional Polaris-based parts have higher boost than desktop models.

 

Not really they have a similar delta between the gtx 1060 to Polaris and a gtx 1080 to this card. Open CL it should perform better than nV's counterparts then in real world in other API's it ends up the same.

 

GTX 1060 and GTX 1080 are based on the same architecture. RX 480 and Vega aren't.

 

How much higher? I doubt Polaris 10 based pro cards are clocked higher than custom desktop cards or that AMD would go this unconventional route for Vega.

 

Then they shouldn't have a similar disparity to nV counter parts, should they? As much as AMD wants to say its way different than the rx480, I don't expect it to be that much different.

not only they don't have much disparity, Vega's tessellation performance is less than the 1080 by a factor of 4in the Catmull-Clark Subdivision Level 3 .... which Polaris doesn't show that its actually pretty close to the 1060 in that benchmark suite.

End results in real world testing we see Polaris is still behind in the 1060 with tessellation, and we can see by all the other test results, synthetics if equal, AMD's cards in real world are lower over all.

 

E. g. P11-based E9260 should boost around 1,4 GHz ("14 CUs, up to 2,5 TFLOPS"), while desktop cards boost up-to 1,2 GHz.

 

687F:C1 vs Fiji

So it performs roughly on par with Fiji allowing for 10-20% higher clocks. Better at small compute tasks, worse at more complex ones?
Nearly half the Catmull3 rate of Fiji allowing for those clocks.
Added FP16 support.
Lacking atomics and a few other extensions. Drivers?

Apparently unlimited constant table.
CL_DEVICE_MAX_MEM_ALLOC_SIZE 4244635648 (687F:C1)
CL_DEVICE_MAX_MEM_ALLOC_SIZE 3018326016 / 3019898880 (Fiji)
CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE 4244635648 (687F:C1)
CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE 65536 (Fiji)