AMD: Navi Speculation, Rumours and Discussion [2019-2020]

Status
Not open for further replies.
no-X said:
Radeon VII and GeForce RTX 2080 have almost the same transistor budget and Radeon VII is expected to perform just a bit worse than GeForce RTX 2080. That doesn't indicate any architectural problem at all. The only problem is lower energy efficiency. Performace per transisor at the same clock is very close.
Click to expand...
Vega needs twice the bandwidth, ~1.35 raw FLOPS, a full process node advantage and that ~1.4 higher TDP to perform "a bit worse than RTX 2080". Sure, the performance per transistor is a nice metric but what about the others?
 
no-X said:
Radeon VII and GeForce RTX 2080 have almost the same transistor budget and Radeon VII is expected to perform just a bit worse
Click to expand...
RTX2080 still has 1/4 less CUs and TFLOPS, and the same transistor count and die clocks are implemented on a 14 nm process node vs 7 nm.

BTW did AMD ever publish INT8/INT4 throughput figures for Vega 20?
 
Last edited:
vipa899 said:
I seriously doubt that.
Click to expand...
I just think that if nvidia would have managed to increase efficiency (perf/w and perf/transistor) on the same process like they did with kepler -> maxwell that would have been a much, much worse problem for AMD. As it is, Turing did nothing (for gaming, excluding the new stuff) for perf/w and is worse for perf/transistor - yes it's got new features, but they came at a cost. Nvidia also increased perf / raw flop, but that did not increase perf/transistor neither (as those new SMs definitely didn't come for free).
I mean, if you compare gk104 (gtx 680) with gm206 (gtx 960), that was quite the efficiency improvement there - the latter has just over half the memory bandwidth, 75% the chip size, 2/3 the power draw, half the TMUs, ... yet in the end the latter is pretty much as fast, and it also had more features (sure they might not have been as flashy as RTX, but kepler actually had quite some limitations with its 11_0 feature set vs. 12_1).
None of that is true for Pascal->Turing, except is has new features. The chips are way bigger for the same performance (plus they get more memory bandwidth too), perf/w is largely the same.
So if AMD feared nvidia would pull off another maxwell, they certainly would have to be relieved. (That said, the writing was on the wall back then for maxwell, since at least some of kepler's inefficiencies, like the too high TMU count and the inefficient SMs where it was mostly impossible to get high alu utilization, were obvious. Since Paxwell didn't show any such obvious flaws, it would have been a miracle if nvidia could have improved on it in a similar spectacular fashion.)
 
no-X said:
Radeon VII and GeForce RTX 2080 have almost the same transistor budget and Radeon VII is expected to perform just a bit worse than GeForce RTX 2080. That doesn't indicate any architectural problem at all. The only problem is lower energy efficiency. Performace per transisor at the same clock is very close.
Click to expand...
Except NVIDIA has a bunch of transistors spent on tensor cores and RT cores that Vega 20 doesn't.
 
And AMD has a bunch of transistors spent on a 4096 Bit memory interface and half-rate FP64.

I don't think it's as easy as pointing out where each has spent this and that many transistors. They both made design choices for various, varying reasons and taking different gambles and bets on whether or not, when and where they would pay off.
 
Cat Merc said:
Except NVIDIA has a bunch of transistors spent on tensor cores and RT cores that Vega 20 doesn't.
Click to expand...
Well, and Vega 20 has a bunch of transistors spent on features, which are not supported on TU104 (e.g. native 16:1 Int4, 8:1 Int8, 1:2 FP64, ECC cache, full HW support for virtualisation, PCIe 4.0, HBCC etc.)
 
yuri said:
Vega needs twice the bandwidth, ~1.35 raw FLOPS, a full process node advantage and that ~1.4 higher TDP to perform "a bit worse than RTX 2080". Sure, the performance per transistor is a nice metric but what about the others?
Click to expand...
I mentioned that energy efficiency isn't good. But why should I care for TFLOPS or memory controller? It's within the limit of the same transistor budget.
 
Because transistor budget doesnt matter. GV100 has 60% more transistors than Vega20 but delivers 3x more training performance in Resnet-50.

Most of the budget is used for features which doesnt relate into direct performance increase. Build a game around Turing's features and a RTX2080 will perform much better.
 
You can make the argument for Vega too, or every gpu, build a game around Vega's feature/arch and it will perform much better... Welcome to the pc space.
 
Last edited:
Turing supports a lot more features: RT? Around 50%+ faster. DL? Around 50%+ faster. Mesh-Shading? VR?

And dont forget: AMD is using 7nm. The process allows for 25% higher clocks. Instead of using more transistors AMD uses the process to increase performance.
 
no-X said:
Well, and Vega 20 has a bunch of transistors spent on features, which are not supported on TU104 (e.g. native 8:1 Int4, 4:1 Int8, 1:2 FP64, ECC cache, full HW support for virtualisation, PCIe 4.0, HBCC etc.)
Click to expand...
Turing still supports IDP4A (4x Int8 dot product-accumulate as well as 2x, 4x and 8x int8, int4 and binary/boolean ops on TCs - http://on-demand.gputechconf.com/gtc-kr/2018/pdf/HPC_Minseok_Lee_NVIDIA.pdf ) and I am not sure whether HBCC is any different from Volta/Turings' virtual memory implementation, neither have I seen any "full HW support for virtualisation" / "Volta MPS" comparisons.
 
1-2% of gamers are buying $600+ GPUs. And 80-90% of them go for nVidia cards. Regardless of unit and production cost, AMD will not make huge profit in that market (probably never)
 
Status
Not open for further replies.
Back
Top