Why are AMD's Vega 56 and 64 performing so badly (please explain and not just guess)

[chugs]

Can someone explain why Vega 56 and 64 are so bad? 

 

Like on paper, even with the increase in price, Vega 64 & 56 should be absolutely smashing GTX 1080 Ti and the 1080. 

 

Like look at the Vega 64 vs the GTX 1080 Ti:

 

Quote

 

Better floating-point performance    12,665 GFLOPS    vs    11,340 GFLOPS    
Much wider memory bus    2,048 bit     vs    352 bit    More than 5.8x wider memory bus
Higher texture rate    395.8 GTexel/s    vs    354.4 GTexel/s    More than 10% higher texture rate
More shading units    4,096    vs    3,584    512 more shading units
More texture mapping units    256    vs    224

 

 

 

And at Vega 56 vs a GTX 1080:

 

Quote

$ Much higher memory bandwidth    409.6 GB/s    vs    224.4 GB/s  ~ Around 85% higher memory bandwidth
Better floating-point performance    10,544 GFLOPS    vs    8,228 GFLOPS   ~ Around 30% better floating-point performance
Higher texture rate    329.5 GTexel/s    vs    257.1 GTexel/s   ~ Around 30% higher texture rate
Much wider memory bus    2,048 bit    vs    256 bit ~   8x wider memory bus
More shading units    3,584    vs    2,560   ~ 1024 more shading units
More texture mapping units    224    vs    160 ~   64 more texture mapping units

 

 

Vega should be smashing the Nvida flagships, even with their increased price points and higher power usage they'd still be the logical choice. And yet with Doom running Vulkan (which is actually optimised for AMD for godsake) on Ultra 3840x2160 shows Vega 64 hitting 79 fps to the GTX 1080 Ti 98 FPS. 

 

What is going on? Is it optimisation? I don't think so because we're talking about Vulkan. Software specifically based on AMD's Mantle. 

 

Is the Stream Processors in Vega that are screwing up or is it HBM? Or is that too glib because from the looks of it there is something else going on because none of what we're seeing makes any sense. 

Hell even for mining the GTX 1070 is kicking Vega 64 and 56 when it comes to performance vs power usage.  

 

More benchmarks that underline what i'm saying make it clear that there is some deep rooted problem with Vega 64 that has not been illustrated yet. Some sort of bottleneck or problem because on paper this card should be creaming the GTX 1080 Ti, whilst Vega 56 should killing the GTX 1080. 

[AlTech]

6 minutes ago, chugs said:

Can someone explain why Vega 56 and 64 are so bad? 

 

Like on paper, even with the increase in price, Vega 64 & 56 should be absolutely smashing GTX 1080 Ti and the 1080. 

 

Like look at the Vega 64 vs the GTX 1080 Ti:

 

 

And at Vega 56 vs a GTX 1080:

 

 

 

Vega should be smashing the Nvida flagships, even with their increased price points and higher power usage they'd still be the logical choice. And yet with Doom running Vulkan (which is actually optimised for AMD for godsake) on Ultra 3840x2160 shows Vega 64 hitting 79 fps to the GTX 1080 Ti 98 FPS. 

 

What is going on? Is it optimisation? I don't think so because we're talking about Vulkan. Software specifically based on AMD's Mantle. 

 

Is the Stream Processors in Vega that are screwing up or is it HBM? Or is that too glib because from the looks of it there is something else going on because none of what we're seeing makes any sense. 

Hell even for mining the GTX 1070 is kicking Vega 64 and 56 when it comes to performance vs power usage.  

 

More benchmarks that underline what i'm saying make it clear that there is some deep rooted problem with Vega 64 that has not been illustrated yet. Some sort of bottleneck or problem because on paper this card should be creaming the GTX 1080 Ti, whilst Vega 56 should killing the GTX 1080. 

AMD Next Generation Stream processors introduced with VEGA are not comparable to GeForce Pascal Stream Processors.

[RadiatingLight]

maybe immature drivers, but most likely simply because AMD's "IPC" with vega somehow pulled a bulldozer and went backwards.

it's just a total clusterfuck of an architecture.

[RadiatingLight]

24 minutes ago, AluminiumTech said:

AMD Next Generation Stream processors introduced with VEGA are not comparable to GeForce Pascal Stream Processors.

+1

[chugs]

27 minutes ago, AluminiumTech said:

AMD Next Generation Stream processors introduced with VEGA are not comparable to GeForce Pascal Stream Processors.

ok that's great that the SP's on vega aren't comparable. 

 

But why not compare Vega with say the Fury X?

 

Look at the Vega 56 vs R9 Fury X:

More memory	8,192 MB	vs	4,096 MB	2x more memory
Better floating-point performance	10,544 GFLOPS	vs	8,602 GFLOPS	Around 25% better floating-point performance
Higher clock speed	1,156 MHz	vs	1,050 MHz	More than 10% higher clock speed
Higher pixel rate	94.14 GPixel/s	vs	67.2 GPixel/s	More than 40% higher pixel rate
Significantly higher turbo clock speed	1,471 MHz	vs	1,050 MHz	More than 40% higher turbo clock speed
Higher texture rate	329.5 GTexel/s	vs	268.8 GTexel/s	Around 25% higher texture rate
Slightly higher effective memory clock speed	1,600 MHz	vs	1,000 MHz	60% higher effective memory clock speed
Higher memory clock speed	800 MHz	vs	500 MHz	60% higher memory clock speed
Lower TDP	210W	vs	275W	Around 25% lower TDP

 

It absolutely dominates the Fury on paper and yet in the Vulkan Doom benchmarks the Fury X is 1 FPS off of beating the Vega 56. 

 

When AMD were developing Vega and you know doing tests and um benchmarks, didn't they like try to you know produce a product that beats the previous generation?  

 

That's why I want to understand is why they went ahead and launched it. Even in mining the Vega series is struggling to beat the GTX 1080.... 

 

I would love it if team did a video/investigation on this

[AlTech]

Just now, chugs said:

snip

Are you using GPUBoss by any chance?

It's really nasty

 

VEGA allegedly doesn't have all the architectural changes enabled in the drivers.

[mariushm]

immature drivers ,. some features are barely working, and were disabled in Vega FE. some features don't work right if you overclock too much

 

lack of HBM2 availability ... uncertainty over how much they could source... so my guess is they set default frequency low just in case they'll have a batch that can't reach high frequencies - overclocking memory gives you significant boost

 

dies use more power than expected -  maybe they set aside the top 5-10% performers for their high end (thousands of dollars) cards and these first cards are just more power hungry

 

the 14nm process is different and maybe not optimized as much as tsmc's 16nm process, bothhave pros and cons

 

maybe there were some designs tradeoffs , for example chip designed in such a way as to cut a chunk of it and copy-paste it in next AM4 processors as vega lite, instead of having smaller polaris 11 and bigger polaris 10

 

maybe vega chips have some features which are disabled in consumer boards like support for local storage (see radeon ssg, up to 2 TB of ssd storage on card, ecc on the memory etc)

 

rushed product ... reference $400 vega 56 uses same pcb and same expensive vrm and same cooler as the $1000 vega FE  etc

[chugs]

38 minutes ago, mariushm said:

immature drivers ,. some features are barely working, and were disabled in Vega FE. some features don't work right if you overclock too much

 

lack of HBM2 availability ... uncertainty over how much they could source... so my guess is they set default frequency low just in case they'll have a batch that can't reach high frequencies - overclocking memory gives you significant boost

 

dies use more power than expected -  maybe they set aside the top 5-10% performers for their high end (thousands of dollars) cards and these first cards are just more power hungry

 

the 14nm process is different and maybe not optimized as much as tsmc's 16nm process, bothhave pros and cons

 

maybe there were some designs tradeoffs , for example chip designed in such a way as to cut a chunk of it and copy-paste it in next AM4 processors as vega lite, instead of having smaller polaris 11 and bigger polaris 10

 

maybe vega chips have some features which are disabled in consumer boards like support for local storage (see radeon ssg, up to 2 TB of ssd storage on card, ecc on the memory etc)

 

rushed product ... reference $400 vega 56 uses same pcb and same expensive vrm and same cooler as the $1000 vega FE  etc

 

 

Ok so all guesses. We really don't know if its an immature product or what is technically causing it to be significantly slower in real world benchmarks then what its hardware should be allowing. 

 

Like even in general compute mining ethereum the Vega 64 is only doing 29 m/h compared to 28 m/h for the GTX 1070. 

 

Why didn't AMD's testing not reveal that its performance was significantly behind that of its main competitor when they were saying benchmarking Doom using Vulkan.

 

We didn't force AMD to ignore its internal benchmarks or testing. They knew where consumer expectation was especially when it came to power draw and yet despite all of the hype Vega is an absolute diasters. 

 

And it doesn't just stop with Vega. 

 

Ryzen is also a very poor performer when it comes to gaming. When AMD Ryzen should be absolutely smashing the competition it fails. For example the Ryzen 1700, overclocked mind you, still doesn't beat the i7-7700k. Hell even the 1800x doesn't either which $200 more. 

 

And again far greater power draw. 

 

Why isn't Linus and the tech community either finding out why, from testing, or demanding accountability from AMD to explain why their two product lines have huge question marks. 

 

They knew the tech community was going to test the hell out of these platforms. Now yes i'm talking about gaming but specifically even with AMD optimised API's like Vulkan. These platforms are still performing poorly. 

 

Hell at this rate it would be a better investment to buy a R9 Fury. 

 

[mariushm]

29 minutes ago, chugs said:

 

 

 

Like even in general compute mining the Vega 64 is only doing 29 m/h compared to 28 m/h for the GTX 1070. 

[...]

 

We didn't force AMD to ignore its internal benchmarks or testing. They knew where consumer expectation was especially when it came to power draw and yet despite all of the hype Vega is an absolute diasters. 

 

Ryzen is also a very poor performer when it comes to gaming. When AMD Ryzen should be absolutely smashing the competition it fails. For example the Ryzen 1700, overclocked mind you, still doesn't beat the i7-7700k. Hell even the 1800x doesn't either which $200 more. 

 

 

mining stuff .. when mining you access the video card memory A LOT. the mining software is optimized to work with 128bit and 256 bit memory buses, reading lots of tiny chunks of memory in parallel.

HBM memory is super wide, each chip has a 1024 bit connection to gpu chip, so vega cards are 2048 bit wide, but the mining software doesn't know how to take advantage of that. (think of it like using only 2nd and 3rd gear on a car that has 5-6 gears)

So you have HBM2 memory that works at ~  2x950 Mhz but it's 2048bit wide and then you have 2x3.5 Ghz GDDR5 like on RX 470/480 or 10-11 Ghz (sort of) with gddr5x but at 128 or 256bit.  

what's gonna win, 950 mhz or 3.5 ghz?

guess what's gonna be badly used if software is too dumb to go above 256 bit?

 

vega56 uses 210w , gtx1070 uses maybe 150w?  majority of people don't give a sh*t about the extra power.. you pay 10-20 us cents for 1kWh so vega56 would basically add a few dollars on your power bill each month. and most dont even pay bills, their parents do.

most people caring about power are reviewers and fanboys trying to find something to complain about (like "my card is the best because power consumption, amd sucks")

 

ryzen has slightly worse IPC compared to Intel but it's very competitive.

ryzen is slower compared to 7700k mostly because 7700k overclocks better - of course 7700k at 4.5-5ghz wil be faster than a 3.8-4 ghz chip.

 

7700k is a 91w tdp chip and probably uses 80-100w at stock, but overclocked at near 5ghz it will use ~150w

You can see here: http://www.anandtech.com/show/10968/the-intel-core-i7-7700k-91w-review-the-new-stock-performance-champion/11

They measured 122w at 4.8 ghz

 

ryzen 1700 is a 65w tdp and at stock will use 60-80w (for 4 more cores with their threads)

It's practically impossible to go over 150w with ryzen .. you get 4 ghz with 1.42v voltage and around 100A on vrm (150w) ..

 

again ryzen won't go easily above 4ghz , maybe it's just the 14nm process... inte's and tsmc (used by nvidia) processes are diifferent

 

so the point is that it's not right to say ryzen is poor performer .. as a cpu is excellent architecture, design... all that.Not his fault it can't do 5 Ghz easily.

 

do your benchmarks with the 7700k capped at 3.8 ghz on all cores against ryzen 1700 and you may see different results.

 

[Damascus]

 Vega (like zen and polaris) is EXTREMELY ram dependant.  

 

There were 2 reasons the Vega launch got pushed back

1. Lack of HBM2 availability 
2. The majority of HBM2 actually produced only performed at 30-60% of the promised spec, tanking the performance on production models of vega, leading to AMD trying to find a solution while pushing the date farther and farther out.

[xsimplyjosh]

Can't just look at numbers, remember Nvidia optimizes their shit all the way which is why they worked so well imho and why people are still springing for it, amd just somewhat went full out. Sure the amd cards have "more" compute units, more teraflops of performance... but it doesnt matter if it isn't well optimized.

[HKZeroFive]

So what I'm hearing is that you're asking why AMD is falling behind, even though on paper it looks to be a much more powerful card in terms of compute.

 

Let's keep it simple because there's a lot of stuff to explain; how a GPU renders an image is quite complex. There are so many factors to consider; how to handle shaders... or what to render... how is the memory compressed and cached to make use of the available bandwidth. NVIDIA does those things better than AMD.

 

The huge difference here is software; AMD has never utilised GCN-based hardware to its full extent. We saw with Polaris (and now with Vega) where there was an inherent lack of bandwidth to fill their cores. Those wasted resources take up a large amount of power.

 

I think Vega also lacks something on a hardware-level too. The GTX 1080Ti has more ROPs and SMs. That could play a part on why the GTX 1080Ti absolutely decimates the RX Vega 64.

 

But anyway, you shouldn't be comparing two different architectures side-by-side as a general rule.

[chugs]

18 hours ago, Damascus said:

11 hours ago, HKZeroFive said:

So what I'm hearing is that you're asking why AMD is falling behind, even though on paper it looks to be a much more powerful card in terms of compute.

 

Let's keep it simple because there's a lot of stuff to explain; how a GPU renders an image is quite complex. There are so many factors to consider; how to handle shaders... or what to render... how is the memory compressed and cached to make use of the available bandwidth. NVIDIA does those things better than AMD.

 

The huge difference here is software; AMD has never utilised GCN-based hardware to its full extent. We saw with Polaris (and now with Vega) where there was an inherent lack of bandwidth to fill their cores. Those wasted resources take up a large amount of power.

 

I think Vega also lacks something on a hardware-level too. The GTX 1080Ti has more ROPs and SMs. That could play a part on why the GTX 1080Ti absolutely decimates the RX Vega 64.

 

But anyway, you shouldn't be comparing two different architectures side-by-side as a general rule.. 

 

Ok so we've established fundamental flaws.  

 

You're saying that AMD doesn't have enough bandwidth (memory bandwidth?) to fill their cores, which Damascus above indicated is due to HBM running significantly below spec. That's a pretty big problem.

 

Don't they run constant tests on their production units? Did they see that these cards were failing to meet their specs? 

 

And re your software remark I think your wrong. How can you argue that when the benchmark's that i've provided are for Doom running on an AMD API, Vulkan. 

 

Unless your saying that AMD are so incomptemptly bad that their competitors, using far less (in hardware terms) can get a far better result. 

 

Like wtf.

 

I feel like i'm eating crazy pills and no gives a crap

[Damascus]

1 minute ago, chugs said:

You're saying that AMD doesn't have enough bandwidth (memory bandwidth?) to fill their cores, which Damascus above indicated is due to HBM running significantly below spec. That's a pretty big problem.

 

Frankly I don't pay too much attention, my understanding is that on early Vega models they used up to spec MB2 and it kicked ass but when they used lower speed stuff the chips massively underperformed

[Damascus]

6 minutes ago, chugs said:

Don't they run constant tests on their production units? Did they see that these cards were failing to meet their specs? 

 

Yes, there was apparently a possibility of a lawsuit against micron

[System Error Message]

clocks and rop.

Lets divide graphics performance in modern GPUs into 3 catogaries. Shaders, ROPs and TMUs. So TMUs handle textures and AF. ROPs mainly deal with stuff like AA (except FXAA) and other final render stuff like bloom.. Shaders do the math and most of the other works like effects, fxaa, shadows. If you tweak your game settings based on how many of these a card has you will get the best performance for a game.

 

Than there are also other dedicated units too. Nvidia has SFU (special function unit) which is very much like a bunch of shaders but different than a shader in architecture, this is what gives nvidia more horsepower despite having less shaders and numbers. You also have dedicated units for things like tessellation from both brands. So if you put all the architectural elements a GPU has together you could figure out the best settings for a game.

 

The other thing are clocks too. Despite having ROPs, TMUs and other units, they are all tied to the core clock. So nvidia's higher clocks gives its fewer units a better advantage especially the shaders too.

[HKZeroFive]

Okay, sorry for the late reply. I only just saw your response.

On 8/21/2017 at 10:45 AM, chugs said:

And re your software remark I think your wrong. How can you argue that when the benchmark's that i've provided are for Doom running on an AMD API, Vulkan. 

Because Radeon RX Vega 64 itself is architecturally inferior to the GTX 1080Ti - they're on two separate "tiers". The NVIDIA card has more memory bandwidth, more ROPs and SMs and higher clocked cores. One also has to consider that the Red Team card is memory bandwidth starved as we've previously mentioned before.

 

There's a reason why I think the best way for AMD to make a competitive card is to drop GCN and start anew. GPUs revolve around the concept of single input, multiple data (otherwise known as SIMD). To do that, you need the hardware to back it up; with SIMD, you need data not only for a thread, but for an entire "group" (AMD calls it wavefronts and NVIDIA calls it warps). NVIDIA might lack the "brute force" but they focus on things such as more and faster cache or a smaller warp size or dedicated units for data... the bottom line is that they do things more efficiently and therefore it usually means they also do it better. And this is just for general compute. It goes way beyond that.

 

To further add on my point about hardware and software, you see AMD absolutely dominating in several synthetic benchmarks where compute performance is fully taken advantage of. But in the real world, NVIDIA has put tons of money towards building a software ecosystem... machine learning and CUDA are some of the big ones. So even though AMD on paper looks to be significantly more powerful, in the end, it's more or less going to be the NVIDIA cards outperforming their AMD counterparts in real world applications. With Volta, that gap is only going to grow even further.

On 8/21/2017 at 10:45 AM, chugs said:

Unless your saying that AMD are so incomptemptly bad that their competitors, using far less (in hardware terms) can get a far better result. 

Yes. That's why hardware and software are equally important.

 

I don't want to say that they're "incompetent" but AMD is at a huge disadvantage in comparison to NVIDIA when it comes to R&D. Getting rid of GCN for future iterations would not be possible.