Right, my point is he should clarify his statements in regards to what line he is talking about, or he ends up making blanket statements that are not totally true. Marketing doublespeak is not how you want to answer a botched launch.
Exactly what I said.
RX vs GP104
WX vs GP102
MI vs GP100
I bought a rx64 for well below 1080 and today just grabbed a rx 56 for just below cheapest 1070 (actually a 1070fe) at komplett that sells in scandinavia. The rx56 arives 6 september.When I shop around, Vega 56 is way more than a 1070, if you can find a Vega 56.
What's more interesting is Raja's exactly saying "look at the bigger picture while judging Vega", and people still cannot understand
Exactly what I said.
RX vs GP104
WX vs GP102
MI vs GP100
As a purchaser of a graphics card why should I care about what Raja thinks the bigger picture is? What is the tangible benefit me?
Exactly, i care about one thing and one thing only, real world gaming performance. I could care less what it does in compute workloads, if this is a compute only card and thats where AMD thinks the big picture is then they should have never released it in the RX line.
AMD's problem is their market share in workstations and servers is not enough to even pay back a fraction of their R&D costs. AMD needs to sell this Vega GPU across gaming(RX Vega), workstation(WX 9100 and Radeon Pro SSG) and deep learning (Radeon MI25) to atleast recover the R&D costs leave alone make profits. AMD will try and sell as much Vega as they can in as many markets as they can.
how much system RAM did you allocate to HBCC? does Windows show your system having less RAM then?Quick update, I tried Metro Last Light Redux at 4k again before work and HBCC definitely makes a huge difference in this use case both in average and especially minimum fps. It's a little strange as afterburner reports less than 2 GB of VRAM used so either afterburner is reporting memory usage wrong or the game/driver is streaming resources inefficiently and HBCC alleviates the bottleneck.
Which is fine, thats their choice. But they shouldn't be surprised when people call them on their lack of performance/efficiency vs their only competitor in gaming, when they themselves release/market it as a gaming card.
And they shouldn't use marketing doublespeak to talk around RX's lack of performance, by trying to push the compute metrics, which are irreverent to the RX line.
Basically what im saying is from a consumer standpoint they are handling the RX vega launch very badly, almost like they dont care about the RX version of vega, and want people to buy vega for compute tasks not gaming.
how much system RAM did you allocate to HBCC? does Windows show your system having less RAM then?
Vega 10 does not support 1/2 rate FP64 like GP100 does. I think what we are seeing here is excuses for a product which fails to compete in perf/watt and perf/sq mm. Hawaii at 438 sq mm competed with GK110 at 550 sq mm. Hawaii had 1/2 rate FP64 compared to 1/3 rate FP64 in GK110. Hawaii was a FP64 monster with 2.6 TFLOPS FP64 performance .
http://www.anandtech.com/show/7901/amd-announces-firepro-w9100
https://www.amd.com/Documents/FirePro_W9100_Data_Sheet.pdf
Hawaii (R9 390/R9 390X) was even able compete with GM104 in gaming performance but was far behind in perf/watt. But with Vega AMD has delivered a real dud. We can draw a direct correlation between AMD marketing and hype vs the real world performance and efficiency it delivers (higher the hype the harder the actual product fails) . With AMD in the past few years we see too much talk but very less actual product competitiveness. I attribute that to the RTG leadership and Raja Koduri. Vega regresses over even Fiji in terms of perf/watt and perf/sq mm. The last great GPU from AMD was Hawaii. I don't think thats going to change for quite sometime.
I find it really hard to believe that Vega is worse than Fiji in terms of performance per watt. Got anything to back that up?
If it were, then Fiji should outperform even a Vega 56, which as far as I know, it does not.
When I say Vega regressed from Fiji in terms of perf/watt its when you normalize for process node. 14LPP brings 50-60% lower power at same performance or > 40% increase in performance at same power. If AMD could have even gone 1.5x Fiji at same clocks they would have got a higher perf improvement at lesser power draw. But we know GCN has a hard limit of 4 shader engines (even though they claim it has no such limit).
64 ROPS since 2014. It just doesn't make any sense. Even if there is a new and better way of getting more geometry output, at least put in some more ROPs just to satisfy the large body of legacy games that wont respond well to anything short of more resources
You look at Hawaii which was a knockout GPU, way more ROPs than we would have expected at the time, and it's held up fantastically due to other aspects of the GPU in combination with solid fixed function output to appease the current API regimes
Be careful about using those marketing claims for process nodes, though. Those numbers do not apply across the board and are probably cherry picked figures of merit at some unspecified voltage/leakage level.
Aren't the ROPs at the back of the rendering pipeline? How can you effectively assess how much of a bottleneck having 64 ROPs is or isn't while the front end is still bottlenecked by being limited to 4 polygons per clock? Note, I'm not saying that 64 ROPs aren't a bottleneck, just that I don't know how you could effectively assess the impact while other bottlenecks remain earlier in the rendering pipeline?64 ROPS since 2014. It just doesn't make any sense. Even if there is a new and better way of getting more geometry output, at least put in some more ROPs just to satisfy the large body of legacy games that wont respond well to anything short of more resources
You look at Hawaii which was a knockout GPU, way more ROPs than we would have expected at the time, and it's held up fantastically due to other aspects of the GPU in combination with solid fixed function output to appease the current API regimes
Very interesting observation. Any other games coming?Quick update, I tried Metro Last Light Redux at 4k again before work and HBCC definitely makes a huge difference in this use case both in average and especially minimum fps. It's a little strange as afterburner reports less than 2 GB of VRAM used so either afterburner is reporting memory usage wrong or the game/driver is streaming resources inefficiently and HBCC alleviates the bottleneck.
Aren't the ROPs at the back of the rendering pipeline? How can you effectively assess how much of a bottleneck having 64 ROPs is or isn't while the front end is still bottlenecked by being limited to 4 polygons per clock? Note, I'm not saying that 64 ROPs aren't a bottleneck, just that I don't know how you could effectively assess the impact while other bottlenecks remain earlier in the rendering pipeline?
This thread might shed some light on ROPs mattering for GCN.
sebbi explains it. https://forum.beyond3d.com/posts/1822194/
Very interesting observation. Any other games coming?
Yes, they did that.If AMD have shipped vega with features turned off then they have shot themselves in the foot.