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

It is not lardass, depends how you compare to Nvidia GPUs and views on some of the features implemented within Vega.
In some ways it is closer to the GP102 with regards to compute, other functionality also makes comparisons less clear with say GP104/gaming; so that is 470mm2 for GP102 compared to 486mm2 for Vega.

So sure if you compare it to the 1080 but that is not really an apples-to-apples comparison apart from pure gaming, which then ignores other functionality/compute use.
 
Last edited:
Is the nature of GF deal such that die size is irrelevant?
Not irrelevant, but you can't compare it apples-to-apples. AMD is definitely paying less per 14FF waffer than nvidia is paying per 16FF+ waffer.
That would be especially true now, with Bitmain taking a sizeable chunk of TSMC's 16FF+ production (larger than nvidia, apparently).
 
Not irrelevant, but you can't compare it apples-to-apples. AMD is definitely paying less per 14FF waffer than nvidia is paying per 16FF+ waffer.
That would be especially true now, with Bitmain taking a sizeable chunk of TSMC's 16FF+ production (larger than nvidia, apparently).

I'd say this is only speculation. We don't know the wafer costs for each party at each of their respective foundries
 
Makes you wonder if the PS4 Pro, now that we know it has 64 ROPs, isn't actually above Polaris 10 cards in real-world scenarios and how close it is to the Xbone X.


Maybe bandwith is limiting the gpu too much for the 64rops te be very efficient ?
 
I'd say this is only speculation.
What is speculation?
That AMD pays less per waffer or that Bitmain is taking more waffers at TSMC than nvidia?

Maybe bandwith is limiting the gpu too much for the 64rops te be very efficient ?
Yes, digitalfoundry stated the Neo's documentation itself says it's theoretically impossible to make full use of the 64 ROPs through the console's bandwidth.
However, it does have twice the number of ROPs that are present in Polaris 10 and Scorpio, and we don't know the full effect of that in practical performance.

Vega M carrying such a large number of ROPs is certainly an odd move here, IMO.
Though it could be that the ROP number in both Vega M and Neo is there to increase VR performance first in foremost (and why Microsoft became so damn quiet about VR in XBone X while the Windows Mixed Reality headsets keep coming up).
 
One also needs to weigh up product margin for TSMC and preferential customer/reservations; TSM has a gross margin around 50%, and it is fair to say Nvidia can support the TSMC margin better than Bitmain due to Quadro and Tesla prices (especially HPC-cloud-analytics-data center servers where the industry is finally seeing healthy growth generally including last quarter) .
The bigger headache for Nvidia would be Apple IMO in context of waifer supply disruption; not necessarily now but in the past and future.
 
Last edited:
I tried comparing the Korean sites results to RX 580 and RX 570 reviews from last year, and nothing seemed out of the ordinary.

So it really has 64 ROPs then? There's not enough compute to really make that useful except maybe for oodles of monitors? Edit: saw the mention of VR - makes sense, sorta.

I also wonder if the GH is the mother chip of both products, or if it's possible both the GH and GL are binned versions of an even better GPU that either Intel or AMD hasn't announced, hence why so many ROPS. There is only a 4 CU difference between the two, and not worth creating two distinct GPUs. I highly doubt AMD made these GPUs just for Intel, and surely they have plans for them in future graphics cards, clocked much higher, and in an unbinned form if GH itself is a binned variant.

Actual Vega M GH - 24 CU / 1536 SP @ 1300 MHz OC = 3.993 TFLOPS
Hypothetical "true" Vega M GH variant - 28 CU / 1792 SP @ 1300 MHz = 4.659 TFLOPS

Would make a good next gen RX 560 replacement if HBM2 prices could be wrestled down.
 
Last edited:
Here is the next part of Primitive Shader Saga:

btw what happened to Primitive Shaders? There's barely any info other than marketing stuff. I was looking forward to it
Timothy Lottes (AMD) said:
CS index culling + draw merging is a portable solution. Given how many devs have yet to adopt a portable solution, it is hard to justify exposing an AMD specific solution when we have a lot of other higher priority Vulkan tasks pending.
 
During anandtech's tour to globalfoundries , it was hinted that 14LPP is 10-15% behind TSMC's 16nm and 7LPP 1-2% behind TSMC's 7nm

It was? In this article? I can't find anything about that. Perhaps they mentioned it in the podcast?

Regardless, looking at measurements made from @Nebuchadnezzar from different chips it looks like the difference between GF's 14LPP and 16FF+ should be substantially larger, since that's the difference between Samsung's 14LPP and 16FF+ and GF largs further behind.
 
It was? In this article? I can't find anything about that. Perhaps they mentioned it in the podcast?

Regardless, looking at measurements made from @Nebuchadnezzar from different chips it looks like the difference between GF's 14LPP and 16FF+ should be substantially larger, since that's the difference between Samsung's 14LPP and 16FF+ and GF largs further behind.


if "extremely competitive" is 1-2% then i guess "very competitive" is around 10%

Q22: Do you/GlobalFoundries care what other semiconductor manufacturers call their processes?

GP: Not really. Our customers know the difference - they get our design kit, they can layout circuits, they know how dense we are. There's no confusion about what our node is. Our 14nm is very competitive with other industry 14nm and 16nm foundry offerings. Our 7nm, which we know from benchmarking, is extremely competitive with other foundries offerings. We are not talking like the differences are 10%; we are talking about 1 or 2 point differences. Our 12nm, we believe, is very competitive with other 12nm offerings. We think that we are pretty well aligned with the other foundries and we have tried to be consistent to avoid confusion, but you know at the end of the day, our platform customers know what our density is.

https://www.anandtech.com/show/1243...ew-with-dr-gary-patton-cto-of-globalfoundries
 
It was? In this article? I can't find anything about that. Perhaps they mentioned it in the podcast?

Regardless, looking at measurements made from @Nebuchadnezzar from different chips it looks like the difference between GF's 14LPP and 16FF+ should be substantially larger, since that's the difference between Samsung's 14LPP and 16FF+ and GF largs further behind.

If you mean the measurements regarding the Apple chips, that was 14 LPE vs 16 FF+. 14 LPP ist said to have 15% better power characteristics. If you look at GP107 at Samsung vs GP106 at TSMC, then there's basically no difference. Samsungs process is making a bit lower clocks (1,9ghz vs 2ghz at tsmc), but that's it. Perf/W is equivalent:
https://www.techpowerup.com/reviews/EVGA/GTX_1070_Ti_FTW2/29.html
 
It was? In this article? I can't find anything about that. Perhaps they mentioned it in the podcast?

Regardless, looking at measurements made from @Nebuchadnezzar from different chips it looks like the difference between GF's 14LPP and 16FF+ should be substantially larger, since that's the difference between Samsung's 14LPP and 16FF+ and GF largs further behind.
GF's 14LPP and Samsung's 14LPP are one and same thing, how could GF "lag further behind"?
 
If you mean the measurements regarding the Apple chips, that was 14 LPE vs 16 FF+. 14 LPP ist said to have 15% better power characteristics.
It depends on the clock zones AFAIR, but in general I think you got the order wrong.,

GF's 14LPP and Samsung's 14LPP are one and same thing, how could GF "lag further behind"?
Apparently they did.
16FF+ is going to hit higher performance at same power towards the end of the perf/power curve compared to 14LPP while being very close to each other at lower frequencies, we saw this in the dual-sourced A9 SoCs. The difference in power should be 10% against Samsung's and larger against GF's process as they are slightly worse (The PDK is the same, but there are manufacturing differences, some say 10% worse than SS).

(It's right here in this thread)
 
after all AMD has produced some of their chips at Samsung 14nm too
Which ones?

Are you perhaps mixing that with the fact that AMD dual-sourced their Vega HBM2 stacks from SK Hynix and Samsung?
 
Back
Top