Why is the Radeon VII slower with similar harware specs (fill rate / FLOPS / etc.) to the 2080TI?

TheHolyLancer

Member
Aug 10, 2005
46
0
66
So I have been wondering about this,

what is making the VII slower than a 2080 TI and is priced at 2080 levels, when the thing according to wikipedia have very similar raw hardware numbers to the 2080 TI?

Numbers are from https://en.wikipedia.org/wiki/GeForce_20_series and https://en.wikipedia.org/wiki/AMD_RX_Vega_series#Radeon_VII_branded_discrete_graphics

Single precision processing power:
2080 Ti: 11750 GFLOPS
2080: 8920 GFLOPS
VII: 11136 GFLOPS

FIll Rate Pixel:
2080 TI: 118.8
2080: 96.96
Vii: 89.6

Fill Rate Texture:
2080 Ti: 367.2
2080: 278.76
VII: 336

Like, from all these specs and others, it should at least be sitting in 2080 + but then 2080 ti - area, but as it stands it is on par with the 2080 and falls far below the 2080 TI, what is going on??

From someone out of the hardware industry, it seems baffling to me why the raw numbers vs performance has such a big discrepancy. Can some one more in the know explain??
 

Markfw

CPU Moderator, VC&G Moderator, Elite Member
Super Moderator
May 16, 2002
18,345
5,672
136
It has like a 20% lower score on the fill rates (approx)
 

Bouowmx

Senior member
Nov 13, 2016
896
334
116
The maximum theoretical numbers you listed are calculated using base frequencies. Those graphics cards never run base frequencies, not even at the stated boost frequencies, but above.
 

JDG1980

Golden Member
Jul 18, 2013
1,658
556
136
Nvidia architectures from Maxwell onward provide better gaming performance per teraflop than AMD's GCN-based architectures, especially on DX11 and OGL. This is for a variety of reasons, related to both hardware and drivers. In particular, Nvidia's big chips have a wider front end that can render up to 6 triangles per clock while AMD's big chips can only render 4.
 

IntelUser2000

Elite Member
Oct 14, 2003
6,269
905
126
From someone out of the hardware industry, it seems baffling to me why the raw numbers vs performance has such a big discrepancy. Can some one more in the know explain??
Those are higher level architectural decisions. There are low level changes that we don't know about. That's what makes all the difference.

A complicated project like that takes thousands, or even tens of thousands of people to get it from concept to a product in your hands. Outsiders can't know every single detail of what went into its production. That's what makes the difference at the end of the day.
 

TheHolyLancer

Member
Aug 10, 2005
46
0
66
I guess the whole 4 vs 6 thing makes sense, its like IPC and how AMD's A64 was doing better but clocked way lower than P4, and then the reverse come Bulldozer and The Bridges.

I guess that's all I can get at short of digging really deep... and likely will only know after a few years before someone who is really in the know release some docs about it like the P4 debacle.
 
Mar 11, 2004
19,083
1,670
126
There's a variety of reasons, but most of them are because AMD hasn't put forth the resources to get their software backend to be competitive with Nvidia's. I don't think even the 4 vs 6 geometry issue would be a big deal were it not for that. But Nvidia has much better geometry culling on top of that (which is why AMD takes a big hit from Tessellation), which means they're cutting out more geometry (that isn't visible so it serves no purpose at the time of it being rendered) before it even hits their geometry processors (so AMD is bogged down even more than the simple 4 vs 6 would indicate). And the ROP situation is an issue for AMD, they routinely have like 2/3 the number of ROPs as the Nvidia cards they're competing with (or hinders them so they end up competing with similar ROP count, see Vega 64 vs GTX1080 where the former has substantial advantages in a lot of areas, but ends up competing with the 1080; they have same ROP count). And then depending on how those are clocked it could be an even bigger difference (Nvidia has had a relative clock speed advantage for awhile now, on top of having a wider geometry and higher ROP count).

AMD has claimed that ROPs aren't a bottleneck for them, but few people are buying it. They also claimed that there isn't a hard limit to the ratio with GCN, but then admitted it would take a lot of engineering to change from the ratio they had settled on. And big surprise, it kept being an issue once Nvidia started regularly offering more. They were allegedly working on things that would get around some of their limitations (were working on an advanced culling system called the NGG Fastpath, was supposed to show up with Vega, but got punted to Navi at the earliest; which Navi also likely has basic changes to the geometry - there's patents that indicate them going from 4 base to 6 for instance, so per SP they should be getting 50% more geometry throughput, we'll see if they can improve things beyond that).

Other areas that Nvidia objectively has advantage is in color compression. I'd guess texture compression as well, but not sure on that. And while AMD allegedly has "Draw Stream Binning Rasterizer" and other features that they claim are already working on Vega, I'm skeptical. If they are, then its another area where AMD's backend is behind Nvidia's. There's been some debate about tile based rendering, with Nvidia likely implementing it with Maxwell (which is why it saw big gains even thought it was on the same process as Kepler). AMD has claimed they support it but it doesn't seem to be nearly as good as Nvidia's implementation.

I think their hardware would've been ok, but AMD struggled to get developers to properly utilize their hardware outside of consoles. And that rarely translated to the PC space for whatever reason (although we did get some glimpses here and there, look at Doom Vulkan for instance). And Vega was I believe extra compute heavy, so it consumed more power for graphics use than it would likely ever be able to realize, and it was going up against an exceptional rasterization architecture from Nvidia (Pascal).

Supposedly AMD recognizes (or recognized) this issue and have been pumping resources into it. It was also why AMD has pivoted towards an open source focused software development situation. That's good and bad. Good in that it opens things up and is what developers want, but bad in that it takes time for that development to provide tangible benefits (its a long term thing). But AMD has been working to bolster it some themselves. Their Linux (open source) support seems to have massively improved starting with Radeon VII/Vega 20.

Raja, the former head of AMD's graphics division was pushing for overhauling the software backend and working to improve it (sounds like he was trying to get AMD more like Nvidia, who has had robust software support), but then Vega came out and underperformed, and then he left. Which AMD had been deliberately limiting and focusing resources of their GPU division for years as they were trying to get their CPU division back on track. So the few resources the GPU division had often got pushed to their semi-custom business (that was doing the game consoles), and allegedly they chose to push resources towards developing Navi instead of getting Vega working like it was supposed to. I personally think that was a good idea, as Vega was already out and so it'd be better to try and come out swinging harder with your next stuff than to try and fix older stuff (the instances where AMD got improvements out of their past hardware, sometimes quite significant, didn't really help them in sales, and Vega was considered a failure by most people at least in dGPU gaming space, so it'd be better to develop things and try to make future products look better from the outset than to try and salvage Vega). We'll see how that turns out. I'm expecting improvements, but not expecting miracles. AMD needs sustained development, and need to execute products better (even when they have good products they often do something to shoot themselves in the foot, like the PCI-e issue with Polaris at launch).

Which Vega was a good chip for some markets (it really isn't a horrible gaming chip, just hot and power hungry). Its an absolute monster for some compute tasks. But I think its inherently flawed. And then was made worse because it was so tied to the expensive HBM technology that meant that it was limited in its ability to be priced more competitively. And some other blunders (AMD has a serious issue in that they seem to do little to no binning/testing of their GPUs, so they tend to run far too high of voltages, which causes extra power draw and extra waste heat produced for no reason, and it even actually causes lower performance - especially in sustained situations; so they end up making their products use more power, put out more heat which causes them to be noisier or have to resort to all-in-one liquid coolers which add even more cost, and it causes their performance to be worse than it otherwise should be; you often can lower your voltages by 100mV and get significantly less power draw, lower waste heat, and performance especially in sustained situations ends up better as the GPU doesn't throttle as much).

So simplest answer: AMD, largely due to the disparity in software/backend, doesn't utilize their hardware as well as Nvidia.
 

GodisanAtheist

Golden Member
Nov 16, 2006
1,601
287
126
Just to nef a bit and say the same thing that was said better by other people:

AMD has a bigger gap between their theoretical throughput vs their actual than NV due to a variety of reasons: Old unbalanced arch, fewer "tricks" to reduce workload and increase occupancy, larger & hotter chips (a function of other problems) resulting in a maximum cap on performance, etc.

GPU's are a bit too complicated to start with a clean sheet design, but AMD does need to start pumping resources into fundamental arch modifications and "Great for today, OK for tomorrow" designs (as opposed to designing for a tomorrow that may never come).
 

IntelUser2000

Elite Member
Oct 14, 2003
6,269
905
126
Someone else summarized AMD using HBM quite well. HBM uses much less power per bandwidth than GDDR technologies, so AMD used HBM to counter their inefficient GPU.

The problem with HBM though is the silicon interposer and the HBM chip itself is a very expensive thing to do, so AMD ends up losing competitiveness, margins, or both.
 

AtenRa

Lifer
Feb 2, 2009
13,024
1,873
126
So I have been wondering about this,

what is making the VII slower than a 2080 TI and is priced at 2080 levels, when the thing according to wikipedia have very similar raw hardware numbers to the 2080 TI?
Two things,

1. Different architectures.
2. RTX 2080Ti = 18.6 Billion Transistors vs Radeon Vii = 13.2 Billion Transistors. That is 40% more Transistors.
 

ASK THE COMMUNITY