AMD Vega (FE and RX) Benchmarks [Updated Aug 10 - RX Vega 64 Unboxing]

[tential]

I also find it mind boggling that people will accept slightly better than 1080 performance at extreme consumption levels.

Amd marketed this card vs Volta when it said poor Volta. It can't even keep up with pascal. That's not acceptable to me.
Many people rightfully feel misled during this launch.

 

[DeathReborn]

I also find it mind boggling that people will accept slightly better than 1080 performance at extreme consumption levels.

Amd marketed this card vs Volta when it said poor Volta. It can't even keep up with pascal. That's not acceptable to me.
Many people rightfully feel misled during this launch.

I do wonder what those people thought of Fermi for all it's extra power for slightly better performance. This could be AMD's Fermi moment (albeit not besting the 1080Ti).

 

[Malogeek]

This is entirely an outlandish idea that I was entertaining today, what if Vega 10 is an HPC compute card and Vega 11 is the gaming variant without the extra wasted die space that will be leaner and more efficient?

There's zero indication they have more than 1 die at the moment, other than a possible binning variant with less CUs. If they do then it's best kept secret in years.

 

[stockolicious]

If this Rx Vega launches with perf that is 15% above a gtx 1080 and consumes 300w it will still be ok, so long as it's priced right 449$ would make this a decent card, assuming it can get another 8-10℅ overclock.
As it a new uarch drivers will improve massively over the next year as AMD will look to implement it's fine wine technology™, playing into the perception that AMD architectures improve more than nvidia over time. (Even though tech nerds know that performance should have been there on day one)

If we get that scenario then it won't be too bad for amd, not great but not a disaster imo as they will be able to compete in a market that has been given to nvidia on a plate the last 18 months.

Gtx1080ti will still be the performance and efficiency king, but
Honestly if you had a 450-500$ budget then Rx Vega is the way to go until poor Volta arrives to deliver a smack down Q2 2018.
Rx Vega would likely vastly outpace gtx 1080 over the next 6 months even if it consumes way more power.

I dont think Vega needs to be the best to sell a ton of cards - they are going to have a high attach rate in the Gaming world with Ryzen and also a high attach rate in the datacenter with EPYC. AMD has built this GPU to work well with their CPU's both have infinity fabric charicteristics. So, a gaming PC, if its all AMD, will have a MUCH lower price of the Intel/Nvidia combo and you will see the same thing on those DC builds. The price / Perfomance will be completely on AMD's side.

 

[french toast]

I also find it mind boggling that people will accept slightly better than 1080 performance at extreme consumption levels.

Amd marketed this card vs Volta when it said poor Volta. It can't even keep up with pascal. That's not acceptable to me.
Many people rightfully feel misled during this launch.

Are you coming from a geek/tech forum perspective or a consumer perspective?

From a geek perspective or even an economic perspective Vega looks to be horrendous for amd, at least until they can get advanced driver and or games patched for its features (6-12 months)

But depending on pricing from a consumer perspective Vega still has the potential to be a better buy than its competitors, by being potentially both faster and cheaper than gtx 1080.
Nvidia has essentially still got its 2016 lineup through 2017 and so this gives amd some wriggle room by chopping it's margins right down.

Joe bloggs doesn't read tech forums he just wants the fastest card for his money, if that money is 450-500$ then he is better served with a Vega, at the downside of the electricity of a light bulb and the upside of faster performance out the box, overclocking on top of that, with much faster drivers coming shortly after.

Its a horrendous technical failure that's almost certainly due to a razor thin budget and loads of concurrent projects, but due to market conditions there is a solid entrance point for Rx vega.

 

[swilli89]

Per game driver optimizations are a huge part of today's video card performance landscape. I know AMD has dedicated software engineers at all of the major game studios (just like nvidia), and it is entirely possible some Vega-specific optimizations aren't enabled/implemented in the current drivers for Vega FE. This is basically all I can think of right now.

Thankfully AMD's outlook has changed drastically with the highly successful launch of Ryzen. A year ago a worst case release like Vega FE performing as it does (in RX form) would be very scary indeed. However now.. if they can add tons of revenue from Ryzen and derivatives they can at least utilize Vega cores in upcoming APUs to offer the world something no one else can. I think with decently clocked DDR4 (and HBM2 in 2018) that these APUs will be rockstars for SFF/laptop gaming. Something like 1024 Vega cores clocked for efficiency and 4 Ryzen cores.

Highly anticipating RX Vega benchmarks. I will concede the outlook is dim, that is for sure. However Lisa Su personally locked down most of the Ryzen teams when they discovered a near critical bug. I have to imagine something similar is happening right this minute from Raja and team. There is a very good reason they pushed RX 6+ weeks out from Vega FE.. not to mention there is also probably a very good reason Vega FE is doing so well in the pro benchmarks.

 

[crisium]

I do wonder what those people thought of Fermi for all it's extra power for slightly better performance. This could be AMD's Fermi moment (albeit not besting the 1080Ti).

As you say with not besting 1080 Ti, GTX 580 still had the performance crown despite being massively larger and using more power.

Huge difference as the flagship effect is real. GTX 280/285/480/580 leading performance charts gets people to overpay for GTX 260/560, etc. If AMD made 400-500mm^2 GPUs during those generations then they would have been where Nvidia is today.

Such a wasted opportunity AMD had. I understand 3870 being small, but once they realized how much advantage they had they should have pushed for massive Nv sized chips. But AMD was scared from 2900XT or just inexperienced with large chips. Nvidia, however, is a wrecking machine and will not sit idly by when they have the die size and wattage advantage.

 

[Malogeek]

But depending on pricing from a consumer perspective Vega still has the potential to be a better buy than its competitors, by being potentially both faster and cheaper than gtx 1080.
Nvidia has essentially still got its 2016 lineup through 2017 and so this gives amd some wriggle room by chopping it's margins right down.

That "850w PSU required" stamp on the box might be a big issue for consumer market though. Sure it won't really need that but currently that's the official requirement for Vega FE. The RX variant will likely require at least the same or more than FE.

 

[tential]

Are you coming from a geek/tech forum perspective or a consumer perspective?

From a geek perspective or even an economic perspective Vega looks to be horrendous for amd, at least until they can get advanced driver and or games patched for its features (6-12 months)

But depending on pricing from a consumer perspective Vega still has the potential to be a better buy than its competitors, by being potentially both faster and cheaper than gtx 1080.
Nvidia has essentially still got its 2016 lineup through 2017 and so this gives amd some wriggle room by chopping it's margins right down.

Joe bloggs doesn't read tech forums he just wants the fastest card for his money, if that money is 450-500$ then he is better served with a Vega, at the downside of the electricity of a light bulb and the upside of faster performance out the box, overclocking on top of that, with much faster drivers coming shortly after.

Its a horrendous technical failure that's almost certainly due to a razor thin budget and loads of concurrent projects, but due to market conditions there is a solid entrance point for Rx vega.

Sorry this simply doesn't hold water. The average
consumer doesn't only look at performance or fury x would have flew off shelves when it was selling at ~$280.

People don't only look at performance per dollar.
If you're going to say the average consumer doesn't know much about this you can't claim they also know perf per dollar.

Being $50 cheaper 1 full year later means nothing. You could have matched that price with a sale.
. This product isn't good from any standpoint.
Vega has way too many things negative about it vs the gtx 1080 that isn't all washed away by a simple 15% performance advantage(if lucky) and a 10% lower price tag.

Only forum nerds will look at those stats.

Consumers will look at performance, power consumption, and the average perception of the card among other factors. This will be the worst selling amd card in recent amd history

 

[Glo.]

From a graphics perspective, It has the exact same Shader/Render/Texture unit count as Fiji, and has slightly less memory bandwidth than Fiji, and that is exactly what it performs like.

For compute loads, FP16 was improved to double performance, but that has ZERO effect on graphics.

The Missing part for graphics, is that there is supposed to be some kind of Tile Based Rendering capability added to improve efficiency and performance, that is clearly doing nothing.

Given the exact same unit count as Fiji, it is almost like AMD placed all their bets on performance increase coming form the Tile Based rendering that is still MIA.

Without that, it is just a clock speed boosted Fiji. The delay for Vega RX is all about getting TBR working right.

Explain this, then. Vega has higher throughput per clock than Fiji. Yet in games sometimes we see decrease in performance per clock, vs Fiji.

It doesn't matter in the end how many cores do you have, if the amount of work they can do in each cycle is higher, than your previous version of GPU architecture.

We observe this with Vega vs Fiji. We observed the same jump in performance with Maxwell vs Kepler. Something is bottlenecking the GPU on graphics side. Is it software? Who knows.

From architecture perspective - there is no bottleneck in GCN hardware level right now. So why does it perform like it performs?

 

[Glo.]

This is entirely an outlandish idea that I was entertaining today, what if Vega 10 is an HPC compute card and Vega 11 is the gaming variant without the extra wasted die space that will be leaner and more efficient?

Like @Glo. mentions the talk of Vega 11 has completely disappeared. Is the reason the RX Vega will be much better than the FE because the silicon is indeed actually different? I'm not really sure any of whats in the FE can be cut out to improve efficiency because it doesn't have DP performance correct?

There is no wasted space in Vega...

 

[Erithan13]

So this just turned up via reddit:

Further deactivated features in the current Vega driver

Binning Rasterizer
(Draw Stream Binning Rasterizer) Apparently inactive Enables the recognition of non-visible parts of the image during the screening, by filtering out these image parts, can be enormously saved by unnecessary work, which considerably speeds the processing of useful work (special advantages are the better utilization of memory bandwidth)
HBCC
(High-Bandwith Cache Controller) In the driver by default disabled Use of the main PC memory as an extended graphics card memory, the actual graphics card memory is used as a cache in this model - can significantly reduce the typical effects of too little graphics card memory in low-memory graphics cards, even dramatically increase the performance ( source )
AVFS
(Adaptive Voltage & Frequency Scaling) In the driver is not running properly Reduces the previous safety margin at the chip voltage by 2-3%, thus resulting in a minimally lower consumption and thus the possibility of minimally higher clock rates to the same power consumption ( source )
ACG
(Advanced Clock Gating) In the driver Exact mode of operation so far still unknown, a current spotting function should be equivalent to AVFS, by means of which the chip clock can increase (slightly) in the end

Grain of salt as usual, but if these features are enabled in RX Vega then a sizeable performance increase doesn't seem so unreasonable. One still has to wonder why AMD wouldn't just say or at least give a decent hint of it though.

 

[Glo.]

That would actually tell why Vega performs just like Fiji, or worse, despite having higher throughput per clock, than mentioned GPU...

 

[PeterScott]

Explain this, then. Vega has higher throughput per clock than Fiji. Yet in games sometimes we see decrease in performance per clock, vs Fiji.

It doesn't matter in the end how many cores do you have, if the amount of work they can do in each cycle is higher, than your previous version of GPU architecture.

We observe this with Vega vs Fiji. We observed the same jump in performance with Maxwell vs Kepler. Something is bottlenecking the GPU on graphics side. Is it software? Who knows.

From architecture perspective - there is no bottleneck in GCN hardware level right now. So why does it perform like it performs?

How much of the Compute loads are FP16? That is known to be twice as fast for Vega, and is NOT used in graphics.

 

[Glo.]

How much of the Compute loads are FP16? That is known to be twice as fast for Vega, and is NOT used in graphics.

All of those are FP32, apart from SNX-02.

 

[beginner99]

When I brought up that Vega 11 stuff no one wanted to talk a out it. It's too far away I was told and I was accused with all types of doll and gloom Now a few weeks later it's OK? Yes I've been saying this as well about Vega 11.

However, recall Polaris. The 11 chip in Polaris brought the efficiency improvements. So Vega 11 really depends on how much of an efficiency improvement it brings.
. But yes, Vega 11 is in a by far worse state as it will come out so close to Volta that I'm wondering if it will come out at all. I wonder if there would be any timetable gains from skipping production of vega 11 and moving to the next move.

What I found more puzzling and gives me a creepy 1984 feeling is the fact that maybe 1-1.5 years ago it was clear that in contrast to Polaris, vega 11 is the big die and Vega 10 the small die. It was said back then, by AMD, that naming is not by size but by design order. Meaning anything based in Vega 10 is the small Vega die. I don't remember how and when but at some point after Polaris launch and vega rumors starting this facts got lost.

So question is, is this really Vega 10 chip? If yes, then AMD actually has/had a bigger, not smaller design up it's sleeve. or this is actually vega 11 chip. Bigger die possible if die size of 3FE is really 484mm2. More problematic is power use. Could be they ditched small vega because not much better than polaris or ditched big vega because too power hungry.

 

[french toast]

Sorry this simply doesn't hold water. The average
consumer doesn't only look at performance or fury x would have flew off shelves when it was selling at ~$280.

People don't only look at performance per dollar.
If you're going to say the average consumer doesn't know much about this you can't claim they also know perf per dollar.

Being $50 cheaper 1 full year later means nothing. You could have matched that price with a sale.
. This product isn't good from any standpoint.
Vega has way too many things negative about it vs the gtx 1080 that isn't all washed away by a simple 15% performance advantage(if lucky) and a 10% lower price tag.

Only forum nerds will look at those stats.

Consumers will look at performance, power consumption, and the average perception of the card among other factors. This will be the worst selling amd card in recent amd history

I think you are not taking into account mindshare and public perception.
Back in 2015 AMD had horrible mindshare and respect amongst consumers or even tech sites, people just bought Intel or nvidia no matter what, Amd has seen a massive turn around in that regard and this alone alongside competitive prices, bundles and freesync2 will attract more sales imo.

Fury was also hampered by only 4gb ram, that didn't look great at the time compared to 980ti having 50℅ more, not to mention slightly better performance.

Your correct consumers do look at those things, but those consumers that do will be the same type who know this is a new architecture, that amd has a very positive and traceable history of improving drivers over time, so they will factor in all factors price/performance/power/future credentials and compare against competitor, I think Rx Vega would look pretty decent vs.1080@450$.

But that's the techy peeps, the casual will mostly but a gaming PC for their budget and whilst I totally accept the vast majority will pick nvidia regardless due to brand value, a larger chunk than in 2015 will plump for AMD imo.

Again this depnds on prices and final performance being at the expected 10-15℅ above 1080.

Edit: I forgot to say that first appearances and reviews count when assessing gpu lives, fury x got average reviews at the time and this stuck for the lifetime of the club together with AMD brand value and limited 4gb ram issue, had fury x had 6gb ram it would have sold much much more imo.
If Vega is priced correctly and is marketed right it could possibly get a better set of reviews on launch which will tide it over on every gpu purchase.
Just my 2 pence worth.

 

[Elixer]

This is entirely an outlandish idea that I was entertaining today, what if Vega 10 is an HPC compute card and Vega 11 is the gaming variant without the extra wasted die space that will be leaner and more efficient?

More efficient in what way, heat? power? MHz?

Like @Glo. mentions the talk of Vega 11 has completely disappeared. Is the reason the RX Vega will be much better than the FE because the silicon is indeed actually different? I'm not really sure any of whats in the FE can be cut out to improve efficiency because it doesn't have DP performance correct?

IIRC, Vega 11 was supposed to be mainstream, not high-end.

What AMD is doing is pushing their architecture much harder than they anticipated.
This is why, when they bump up the speed, they need to push more voltage, and that increases the heat which leads to the FE thermal throttling since it can't maintain the 1600MHz target boost longer than a few secs.
In other words, they never expected nvidia's card(s) to be clocked that much higher, and still have lots of room for O/C'ing.

For those saying they will just wait for Volta, if AMD can't get some decent competition here now, then Volta hardware will most likely be even more expensive than what the 1080ti is at now, and they can ride that until AMD's next chip is ready.

 

[Magic Hate Ball]

So this just turned up via reddit:

Further deactivated features in the current Vega driver



Grain of salt as usual, but if these features are enabled in RX Vega then a sizeable performance increase doesn't seem so unreasonable. One still has to wonder why AMD wouldn't just say or at least give a decent hint of it though.

If clock-gating isn't working right now, along with that minimal voltage lowering feature... that would explain the high energy usage.

 

[CatMerc]

From a graphics perspective, It has the exact same Shader/Render/Texture unit count as Fiji, and has slightly less memory bandwidth than Fiji, and that is exactly what it performs like.

For compute loads, FP16 was improved to double performance, but that has ZERO effect on graphics.

The Missing part for graphics, is that there is supposed to be some kind of Tile Based Rendering capability added to improve efficiency and performance, that is clearly doing nothing.

Given the exact same unit count as Fiji, it is almost like AMD placed all their bets on performance increase coming form the Tile Based rendering that is still MIA.

Without that, it is just a clock speed boosted Fiji. The delay for Vega RX is all about getting TBR working right.

There was more than just TBR that was supposed to uplift Vega. Improved scheduling and utilization, ROP's clients of L2, to name another couple of big ones.

 

[tential]

If you believe that amd had mind share problems beforehand wait until after rx Vega only competes with a 1 year old midrange card.

 

[french toast]

If you believe that amd had mind share problems beforehand wait until after rx Vega only competes with a 1 year old midrange card.

Since when has gtx 1080 been midrange?? It's still nvidia 2017 high end gaming card, not it's top enthusiast but it's high end premium.
And by competing it should be beating it, outside of power consumption.

 

[crisium]

Dunno about the semantics of GTX 1080 being midrange or not, but Vega risks competing with GV106 / GTX 2060 in a worst case scenario.

I guess realistic optimistic scenario for AMD is a repeat Tonga vs Maxwell? 380X and even cutdown 380 > 960 (but used more power and both were outsold by Nvidia anyway). In that case Vega will still be faster than 2060, but get smoked by the 2070.

I hope there's some driver magic to let them compete with 2070 (best case scenario). That would likely mean it needs to at least match 1080 Ti though, based on history.

 

[Elixer]

So this just turned up via reddit:
Further deactivated features in the current Vega driver

Grain of salt as usual, but if these features are enabled in RX Vega then a sizeable performance increase doesn't seem so unreasonable. One still has to wonder why AMD wouldn't just say or at least give a decent hint of it though.

This would make no sense at all that they would vastly gimp their $1000+ card just to spring a August surprise.
AMD can't be THAT incompetent to have pretty much all major improvements "disabled".

It also makes no sense at all that AMD hasn't released a new branch of the drivers for FE owners, which surely must be better than the launch drivers in some ways, unless there isn't any room for improvements that are really meaningful.
No hotfixes at all.
I bet they are working on game specific optimizations, and only for a handful of AAA games, to not make them look as bad as with the FE drivers.

 

[tential]

Dunno about the semantics of GTX 1080 being midrange or not, but Vega risks competing with GV106 / GTX 2060 in a worst case scenario.

I guess realistic optimistic scenario for AMD is a repeat Tonga vs Maxwell? 380X and even cutdown 380 > 960 (but used more power and both were outsold by Nvidia anyway). In that case Vega will still be faster than 2060, but get smoked by the 2070.

I hope there's some driver magic to let them compete with 2070 (best case scenario). That would likely mean it needs to at least match 1080 Ti though, based on history.

Exactly on point.
Basically, if fury x was 1 year later it would have done even worse. That's what amd did with Vega.