AMD Radeon VII Detailed Some More: Die-size, Secret-sauce, Ray-tracing, and More

[efikkan]

I wonder why AMD is stuck with maximum 4096 SP's ?
I mean.... Fury, Vega (1), Vega II ... they are almost identical.

Considering that the new chip is rather small at 331 mm2, what stopped them from making a 450 mm2 chip for example and fitting 72 CU's in it, or 96 !!
It would wipe the floor with 2080 Ti with 6144 SP's (let's say cut a few for being defective, even with 5760 SP's it would still crush it with raw computer power and that massive 1TBps bandwidth, WHILE BEING A SMALLER CHIP due to 7nm)

Instead, they just shrunk Fury, then shrunk it again without adding anything

FordGT90Concept said it's yields, and that's part of it, but the biggest reason is probably resource management. If AMD were to make a GPU with 50% more cores, it would need at least 50% scheduling resources. Resource management is already the main reason why GCN is inefficient compared to Nvidia, and the reason why RTX 2060 (1920 cores) manages to match Vega 64 (4096 cores). As we all know, AMD have plenty of theoretical performance that they simply can't utilize properly. Adding 50% more cores would require rebalancing of the entire design, otherwise they would risk getting even lower efficiency. Vega 20 is just a tweaked design with some professional features added.

 

[Manoa]

it also whay async help it significantly ?

 

[moproblems99]

whare do you see complaining ? and whare do you see don't exist ?

AMD shills ? I can respect that, you look like a fighter too. "fight for your right for gaming on AMD, kill anyone that looks like against AMD" ?
but you could use a brain: a gaming developer relationship program will benefit AMD more than a few more mhz and a few more GB/s memory banwith
you don't see the advantage of that ? for your own good ? what does that make you ?
how about async compute enabled on all games sounds to you ? should I mention how much faster doom 4 was with async enabled ? and that was just one game where is was used.....WITHOUT AMD's help...and that's just the beginning, are you able to imagine what it could mean if AMD was involved ? in all games ?
oh wait, you are a radeon expert, im sorry you must know more than I do

U mad bro?

Actually, I would prefer no developer relationships with either company because they are only good for one 'color'. Clearly, one of us needs a brain...

 

[FordGT90Concept]

it also whay async help it significantly ?

Yes, GCN is an async monster because of under utilization of its hardware resources.

 

[londiste]

it also whay async help it significantly ?

Yes, GCN is an async monster because of under utilization of its hardware resources.

Is it true any more? Latest testing on async seems to show both camps benefiting from it, AMD cards more than Nvidia but the difference is a couple percent at most.

 

[ssdpro]

I fail to see the missing technology. RTX is usable in one game...and the series is trash. DLSS looks like shit compared to the other available methods. I fail to see what benefits the 2080 has.

Missing tech is missing tech. It isn't in one game, it is in many more and many more coming. Real time ray tracing is in these games:

• Assetto Corsa Competizione
• Atomic Heart
• Battlefield V
• Control
• Enlisted
• Justice
• JX3
• MechWarrior 5: Mercenaries
• Metro Exodus
• ProjectDH
• Shadow of the Tomb Raider

As for DLSS, the list is longer:

• Ark: Survival Evolved
• Anthem
• Atomic Heart
• Battlefield V
• Dauntless
• Final Fantasy 15
• Fractured Lands
• Hitman 2
• Islands of Nyne
• Justice
• JX3
• Mechwarrior 5: Mercenaries
• PlayerUnknown’s Battlegrounds
• Remnant: From the Ashes
• Serious Sam 4: Planet Badass
• Shadow of the Tomb Raider
• The Forge Arena
• We Happy Few
• Darksiders III
• Deliver Us The Moon: Fortuna
• Fear the Wolves
• Hellblade: Senua’s Sacrifice
• KINETIK
• Outpost Zero
• Overkill’s The Walking Dead
• SCUM
• Stormdivers

Again, the Radeon VII price is just too high. You'll notice I never criticize the product - it uses great memory and plenty of it. I like that feature. What I don't like is the high price and missing features. At a proper market price of 499-549 it is a winner. To end with a tip, users appear more credible if you post your comments without profanity and with support for facts.

Data sources: https://www.digitaltrends.com/computing/games-support-nvidia-ray-tracing/ , https://www.kitguru.net/components/...rt-nvidias-ray-tracing-and-dlss-rtx-features/

 

[moproblems99]

Phew. Exactly 0 of those games are on my play list so not a huge loss. That said, thanks for throwing the lists up there. Now, what do we do about RTRT making everything look a mirror? or DLSS looking like a jaggy mess? I will say that I did go back and look at the comparisons and DLSS doesn't look as bad as I originally thought. I still wouldn't call either of these pieces of technology that are earth shattering. Down the line? Probably. Right now? Nothing special.

 

[ValenOne]

Which is why Vega 20 isn't bigger than Vega 10. I think Huang's explosion is because he realizes he made a "big" mistake with Turing. AMD is focusing on where the money is at, not winning performance crowns that mean little in the larger context of things. Turing is substantially larger (and more costly to produce) than even Vega 10 is.


On topic, Vega 20 doesn't really impress but it really wasn't intended to impress either. Vega 7nm w/ Fiji memory bandwidth.

VII is still Vega with the same 64 ROPS bottleneck. AMD should have scaled from Vega M GH with 64 ROPS and 24CU ratio.

Missing tech is missing tech. It isn't in one game, it is in many more and many more coming. Real time ray tracing is in these games:

• Assetto Corsa Competizione
• Atomic Heart
• Battlefield V
• Control
• Enlisted
• Justice
• JX3
• MechWarrior 5: Mercenaries
• Metro Exodus
• ProjectDH
• Shadow of the Tomb Raider

As for DLSS, the list is longer:

• Ark: Survival Evolved
• Anthem
• Atomic Heart
• Battlefield V
• Dauntless
• Final Fantasy 15
• Fractured Lands
• Hitman 2
• Islands of Nyne
• Justice
• JX3
• Mechwarrior 5: Mercenaries
• PlayerUnknown’s Battlegrounds
• Remnant: From the Ashes
• Serious Sam 4: Planet Badass
• Shadow of the Tomb Raider
• The Forge Arena
• We Happy Few
• Darksiders III
• Deliver Us The Moon: Fortuna
• Fear the Wolves
• Hellblade: Senua’s Sacrifice
• KINETIK
• Outpost Zero
• Overkill’s The Walking Dead
• SCUM
• Stormdivers

Again, the Radeon VII price is just too high. You'll notice I never criticize the product - it uses great memory and plenty of it. I like that feature. What I don't like is the high price and missing features. At a proper market price of 499-549 it is a winner. To end with a tip, users appear more credible if you post your comments without profanity and with support for facts.

Data sources: https://www.digitaltrends.com/computing/games-support-nvidia-ray-tracing/ , https://www.kitguru.net/components/...rt-nvidias-ray-tracing-and-dlss-rtx-features/

DLSS..promoting pixel reconstruction technique is like Sony's PS4 Pro's pixel reconstruction marketing. VII has higher memory bandwidth for MSAA.

Yes, GCN is an async monster because of under utilization of its hardware resources.

It's more ROPS read-write bottleneck with workaround fix like async compute/TMU read-write path.

 

[M2B]

VII is still Vega with the same 64 ROPS bottleneck. AMD should have scaled from Vega M GH with 64 ROPS and 24CU ratio.


DLSS..promoting pixel reconstruction technique is like Sony's PS4 Pro's pixel reconstruction marketing. VII has higher memory bandwidth for MSAA.


It's more ROPS read-write bottleneck with workaround fix like async compute/TMU read-write path.

You have no idea what you are talking about.
It has nothing to do with ROPs, having 64 ROPs is completely fine at up to 3840*2160p and won't cause any bottleneck.
Do you really think AMD is dumb enough to make GPUs that are bottlenecked by such a relatively simple thing?

 

[95Viper]

Stop the retaliatory comments, bickering, baiting, and insulting. You know who you are.
Keep the discussions civil and on topic.

Thank You

 

[FordGT90Concept]

VII is still Vega with the same 64 ROPS bottleneck. AMD should have scaled from Vega M GH with 64 ROPS and 24CU ratio.

AMD ROPs process 4 pixels per clock. In the case of Radeon VII: 64 ROPs * 4 pixels per clock * 1,800,000,000 clocks = 460,800,000,000 pixels processed per second. 4K is 8,294,400 pixels. Radeon VII has enough ROPs and clocks to handle 4K at 55,555 fps. ROPs aren't a problem.


GCNs problem is that the drivers (read: CPU) do more work than Maxwell+ do. This is why Vega and Fiji do not so great at low resolutions but well at high resolutions (CPU matters less). Vulkan and Direct3D 12 perform much better on GCN than Direct3D 11 because it naturally removes a lot of the underlying CPU burden.

I suspect Navi will take lessons learned from Xbox One and PlayStation 4 to produce hardware that needs minimal driver interference.

 

[efikkan]

It has nothing to do with ROPs, having 64 ROPs is completely fine at up to 3840*2160p and won't cause any bottleneck.
Do you really think AMD is dumb enough to make GPUs that are bottlenecked by such a relatively simple thing?

You're correct. If ROPs were a major bottleneck, AMD would have solved that by now and unleashed a massive performance gain, but they're not.
That doesn't mean you can't find an edge-case where more ROPs don't help, as with anything else, but that's beside the point.

GCNs problem is that the drivers (read: CPU) do more work than Maxwell+ do. This is why Vega and Fiji do not so great at low resolutions but well at high resolutions (CPU matters less). Vulkan and Direct3D 12 perform much better on GCN than Direct3D 11 because it naturally removes a lot of the underlying CPU burden.

Well, it's true that Direct3D 12 and Vulkan can offload a lot of the management done by the drivers, but not the truly low-level allocation and resource management, which happens inside the GPU, and this is where GCN struggles.

 

[M2B]

GCNs problem is that the drivers (read: CPU) do more work than Maxwell+ do. This is why Vega and Fiji do not so great at low resolutions but well at high resolutions (CPU matters less). Vulkan and Direct3D 12 perform much better on GCN than Direct3D 11 because it naturally removes a lot of the underlying CPU burden.

Driver overhead is only a small part of why GCN is behind Maxwell/Pascal/Turing.
Just like intel's advantage over AMD in gaming, my guess is that it has something to do with way lesser cache latencies on Nvidia GPUs.
Of course this isn't the full story.

 

[Mescalamba]

Think it would be maybe viable to have raytracing add-on card. Maybe something in SLi/CrossFire style. One regular, one for raytracing.

 

[moproblems99]

Do you really think AMD is dumb enough to make GPUs that are bottlenecked by such a relatively simple thing?

I mean in everyone else's defense, they aren't doing anything to change it so they were dumb enough to have a bottleneck in a place that isn't easily fixed.

 

[londiste]

Now, what do we do about RTRT making everything look a mirror?

Of course the demos and showcases for new features will over-emphasize the feature but once it makes it into actual games, it is generally much more toned down. Have you actually played BF5 with DXR?
Shadow of the Tomb Raider patch (shadows) and Metro Exodus (AO) should be a lot more interesting RTRT use cases once they show up.

 

[Wavetrex]

AMD ROPs process 4 pixels per clock. In the case of Radeon VII: 64 ROPs * 4 pixels per clock * 1,800,000,000 clocks = 460,800,000,000 pixels processed per second. 4K is 8,294,400 pixels. Radeon VII has enough ROPs and clocks to handle 4K at 55,555 fps. ROPs aren't a problem.

Those need to do both texture element reading and buffer output writing.

It's not just about the final frame, because they aren't drawing the same uniform color 55 thousand times per second.

Every texel that needs to be placed somewhere on a scene also needs to be accessed. The ratio of texels read to pixels written might vary widely, but considering that modern games have millions of triangles on the screen, that's a whole lot of textures that need to be read and their data calculated in order to obtain the final pixel.
Obviously there's CU caching involved as well but the ratio is still huge.

Let's assume some numbers: If Texel to Pixel ratio is 20:1, and due to overdrawing more pixels are written than actually seen the display ratio is "only" 5:1 (might be much higher in complex scenes with a lot of translucency, edge anti-aliasing and more), that is already an 100:1 vs your number.

55 thousand fps is now just 550.

Add multi-sampling (4X) and now back to a more realistic 550 : 4 = 137 fps, which seems to be what these cards can do...

 

[efikkan]

The myth of the ROP bottleneck for Vega must come to an end.
Just compare GTX 1080, RTX 2060 vs. Vega 64, three GPUs which perform similarly. While GTX 1080 had 102.8 - 110.9 GP/s, RTX 2060 reduced it to 65.52-80.54 GP/s (less than Vega 64's 79.8 - 98.9 GP/s), and still managed to maintain performance, even at 4K.
The issue with GCN is not the raw throughput of the various resources, it's the management of those resources.

 

[moproblems99]

Of course the demos and showcases for new features will over-emphasize the feature but once it makes it into actual games, it is generally much more toned down. Have you actually played BF5 with DXR?
Shadow of the Tomb Raider patch (shadows) and Metro Exodus (AO) should be a lot more interesting RTRT use cases once they show up.

I'll never play BF5 but Exodus is a maybe. All I have seen is it looks awful in the demos and videos so what is there to get excited about? I have nothing against the 20 series. I just don't understand what all the fuss is about. Are these the future? Probably, but it is not going to be until consoles take them on so there is no point to get all hot and bothered about it now. I hope NV get up to 1.21 jigga rays some day but until then, color me unimpressed. I am more impressed with extra performance they squeezed out with this gen. I didn't realize they could beat a horse that well.

 

[ValenOne]

AMD ROPs process 4 pixels per clock. In the case of Radeon VII: 64 ROPs * 4 pixels per clock * 1,800,000,000 clocks = 460,800,000,000 pixels processed per second. 4K is 8,294,400 pixels. Radeon VII has enough ROPs and clocks to handle 4K at 55,555 fps. ROPs aren't a problem.


GCNs problem is that the drivers (read: CPU) do more work than Maxwell+ do. This is why Vega and Fiji do not so great at low resolutions but well at high resolutions (CPU matters less). Vulkan and Direct3D 12 perform much better on GCN than Direct3D 11 because it naturally removes a lot of the underlying CPU burden.

I suspect Navi will take lessons learned from Xbox One and PlayStation 4 to produce hardware that needs minimal driver interference.

Wrong.

Vega 56 at 1710 Mhz with 12 TFLOPS beating Strix Vega 64 at 1590mhz with 13 TFLOPS. This shows higher clock speed improves raster hardware which enable TFLOPS to be exposed. TFLOPS is useless without ROPS read-write factors.

For absolute performance, lowering time to completion in all graphics pipeline stages should be a priority. Your argument doesn't lead to lowest latency operation.

Where did you obtain Vega ROPS unit has 4 color pixels per clock? Each RB unit has four color ROPS and 16 z-ROPS. For 64 ROPS, that's 16 RB units x 4 color ROPS = 64 color ROPS, hence it's 64 color pixels per clock.

IF AMD is pushing for compute/256 TMUs read-write path, why not increase ROPS count to match TMU read-write performance?

The myth of the ROP bottleneck for Vega must come to an end.
Just compare GTX 1080, RTX 2060 vs. Vega 64, three GPUs which perform similarly. While GTX 1080 had 102.8 - 110.9 GP/s, RTX 2060 reduced it to 65.52-80.54 GP/s (less than Vega 64's 79.8 - 98.9 GP/s), and still managed to maintain performance, even at 4K.
The issue with GCN is not the raw throughput of the various resources, it's the management of those resources.

RTX 2060 (30 SM) has 48 ROPS at 1900 Mhz stealth overclock. Full TU106 (36 SM) has 4 MB L2 cache. Full GP104 has 2MB L2 cache. This is important for NVIDIA's immediate mode tile cache render loop and improves to lower latency which improves reduces time to completion timings.

 

[efikkan]

Vega 56 at 1710 Mhz with 12 TFLOPS beating Strix Vega 64 at 1590mhz with 13 TFLOPS. This shows higher clock speed improves raster hardware which enable TFLOPS to be exposed. TFLOPS is useless without ROPS read-write factors.

This proves nothing in terms of claiming ROPs are the bottleneck. When using a die with fewer cores and compensating with higher clocks there are a lot of other things than just ROP per GFLOP that changes. Slightly cut down chips may have a different resource balance, both on the scheduling, but also cache and register files. All of these are impacting performance long before ROPs even come into play.

We see the same thing with GTX 970, which has higher performance per clock than it's big brother GTX 980. Why? Because it struck a sweetspot in various resources.

 

[ValenOne]

You're correct. If ROPs were a major bottleneck, AMD would have solved that by now and unleashed a massive performance gain, but they're not.
That doesn't mean you can't find an edge-case where more ROPs don't help, as with anything else, but that's beside the point.


Well, it's true that Direct3D 12 and Vulkan can offload a lot of the management done by the drivers, but not the truly low-level allocation and resource management, which happens inside the GPU, and this is where GCN struggles.

AMD is pushing compute shader -TMU read-write path.

Refer to Avalanche Studios lecture on TMU read-write workaround on ROPS bound situations.

This proves nothing in terms of claiming ROPs are the bottleneck. When using a die with fewer cores and compensating with higher clocks there are a lot of other things than just ROP per GFLOP that changes. Slightly cut down chips may have a different resource balance, both on the scheduling, but also cache and register files. All of these are impacting performance long before ROPs even come into play.

We see the same thing with GTX 970, which has higher performance per clock than it's big brother GTX 980. Why? Because it struck a sweetspot in various resources.

For AMD GCN, register files are with CU level. Each CU has it's own warp scheduling.

GTX 970 has 56 ROPS with less L2 cache.
GTX 980 has 64 ROPS

Vega 56 has full 64 ROPS and full 4MB L2 cache like Vega 64. VII's 60 CU still has full L2 cache and 64 ROPS access like MI60. Faster clock speed improves L2 cache and 64 ROPS i.e. lessen the time to completion.

 

[FordGT90Concept]

https://hothardware.com/news/amd-radeon-rx-vega-56-unlocked-vega-64-bios-flash

Vega 56 running Vega 64 BIOS is 2% slower than Vega 64. That likely has less to do ROPs and more to do with underutilization of shaders as I previous said. Vega, and Fiji before it, were designed for server farms running compute loads. They were never ideal for gaming. Polaris, on the other hand, is biased towards gaming. It has 32 ROPs to 36 CUs (8:9 compared to Vega 64 1:1 or Vega 56 8:7) . Again, if ROPs were really the bottleneck, AMD would have put more ROPs on it but they didn't.

Vega and Fiji do exceptionally well in async games like Ashes of the Singularity because all of those shaders aren't so underutilized.

 

[ValenOne]

https://hothardware.com/news/amd-radeon-rx-vega-56-unlocked-vega-64-bios-flash

Vega 56 running Vega 64 BIOS is 2% slower than Vega 64. That likely has less to do ROPs and more to do with underutilization of shaders as I previous said. Vega, and Fiji before it, were designed for server farms running compute loads. They were never ideal for gaming. Polaris, on the other hand, is biased towards gaming. It has 32 ROPs to 36 CUs (8:9 compared to Vega 64 1:1 or Vega 56 8:7) . Again, if ROPs were really the bottleneck, AMD would have put more ROPs on it but they didn't.

Vega and Fiji do exceptionally well in async games like Ashes of the Singularity because all of those shaders aren't so underutilized.

Async compute and Sync compute shaders has TMU read-write path software optimizations.
Again, read Avalanche Studios lecture on TMU read-write workaround on ROPS bound situations.

Btw: Vega M GH has 24 CU with 64 ROPS ratio.

For current Vega 64 type GPU, it's better to trade less CU count which reduce power consumption for higher clock speed which has higher ROPS/L2 cache/rasterization performance. Vega 56 at 1710 Mhz with 12 TFLOPS beating Strix Vega 64 OC at 1590Mhz with 13 TFLOPS.

Rasterization = floating point geometry to pixel integer mass conversion hardware.

Note the four Rasterzier hardware units, AMD increased this hardware during R9-290X's introduction before "Mr TFLOPS" joined AMD in 2013.

At higher clock speed, classic GPU hardware such as Rasterzier, Render Back End(ROPS) and L2 cache has higher performance i.e. lower time to completion.

At the same clock speed and L2 cache size, 88 ROPS has 88 pixels per clock has lower time completion when compared to 64 ROPS with 64 pixels per clock. AMD knows ROPS bound problem hence compute shader's read-write path workaround marketing push and VII's 1800Mhz clock speed with lower CU count.

NVIDIA's GPU designs has higher clock speeds to speed up classic GPU hardware.


Cryptocurrency uses TMU read-write path instead of ROPS read-write path.

AMD could have configured a GPU with 48 CU, 64 ROPS and 1900 Mhz

RX-580... AMD hasn't mastered 64 ROPS over 256 bit bus design, hence RX-580 is stuck at 32 ROPS with 256 bit bus. R9-290X has 64 ROPS with 512 bit bus which is 2X scale over R9-380X/RX-580's design.

 

[medi01]

WCC level of speculations in an article that in essence highlights what nVidia would like to highlight about VegaVII vs 2080.

How do things of that kind work, do authors of these texts simply root for nVidia, do they come from Huang's headquarters, are they somehow censored by NV's PR team?
Just curious.