AMD's RX Vega Low Key Budapest Event: Vega Pitted Against GTX 1080

[S@LEM!]

HBM2 should be lower power and higher performance than GDDR5X.

Thing is, we're just speculating. AMD has never said why they released a Frontier Edition ahead of the main product. They also never said why Vega keeps getting kicked down the road. I would love to be a fly on the wall in RTG meetings on Vega to hear their reasoning for these things.

1080ti messed them up

 

[FordGT90Concept]

HBM2 is a little more energy efficient than GDDR, but compared to the hot GPU it wouldn't matter much.
384-bit GDDR5X is faster than 2048-bit HBM2 BTW…

http://www.hotchips.org/wp-content/uploads/hc_archives/hc26/HC26-11-day1-epub/HC26.11-3-Technology-epub/HC26.11.310-HBM-Bandwidth-Kim-Hynix-Hot Chips HBM 2014 v7.pdf
http://www.anandtech.com/show/9883/gddr5x-standard-jedec-new-gpu-memory-14-gbps
Vega has 2 stacks of HBM2 at 512 GB/s (~7.3w). GTX 1080 Ti has 11 chips of GDDR5X at 484 GB/s (~27.5w). Titan Xp has 12 chips of GDDR5X at 547.7 GB/s (~30w).

Even thought Titan Xp has a slight edge over Vega in bandwidth, Vega has significantly lower latency and more flexibility in making memory requests. HBCC spawned from lessons learned in using HBM for Fiji.

1080ti messed them up

I wouldn't bet on that.

 

[cdawall]

http://www.hotchips.org/wp-content/uploads/hc_archives/hc26/HC26-11-day1-epub/HC26.11-3-Technology-epub/HC26.11.310-HBM-Bandwidth-Kim-Hynix-Hot Chips HBM 2014 v7.pdf
http://www.anandtech.com/show/9883/gddr5x-standard-jedec-new-gpu-memory-14-gbps
Vega has 2 stacks of HBM2 at 512 GB/s (~7.3w). GTX 1080 Ti has 11 chips of GDDR5X at 484 GB/s (~27.5w). Titan Xp has 12 chips of GDDR5X at 547.7 GB/s (~30w).

Even thought Titan Xp has a slight edge over Vega in bandwidth, Vega has significantly lower latency and more flexibility in making memory requests. HBCC spawned from lessons learned in using HBM for Fiji.


I wouldn't bet on that.

HBCC has shown no improvements with Vega as it sits.

 

[efikkan]

HBCC has shown no improvements with Vega as it sits.

HBC works kind of similar to the prefetcher in a CPU, it's only able to detect linear access patterns. It will probably work flawlessly for some compute workloads, but it wouldn't do anything for random access patterns, which are typical for games.

And BTW, caching is about hiding latency, not increasing performance.

 

[cdawall]

HBC works kind of similar to the prefetcher in a CPU, it's only able to detect linear access patterns. It will probably work flawlessly for some compute workloads, but it wouldn't do anything for random access patterns, which are typical for games.

And BTW, caching is about hiding latency, not increasing performance.

By decreasing latency it should increase performance. I mean that's how every single thing works. Maybe AMD just has that cool magic they keep hyping us about.

The latency in compute performance is also proving to be inconsistent as hell.

 

[FordGT90Concept]

And that could be the root of it. Perhaps there was a major hardware flaw in the HBCC that triggered erroneous cache misses. They got a pile of cards ready to ship so they sell them under a limited edition moniker. Fix the bug in another silicon revision and performance could improve immensely. We just don't know until the final product is available.

Decreasing latency mitigates how long it takes for the GPU to move forward when a cache miss occurs. This doesn't increase framerate (unless it is happening regularly) but it increases minimum FPS.

My point is that 13 TFLOPS of compute power shouldn't result in as low of a framerate as it gets. There's clearly something wrong in the Frontier Edition and AMD knows it.

 

[efikkan]

By decreasing latency it should increase performance.

You didn't get it. The point of caching is to utilize a small fast storage pool as if it was a larger fast storage pool, by using a small fast pool as a buffer for the large one.

So by Vega having 8 GB HBM with caching, it will only act as if the memory pool was larger, caching will never give you more performance than having the larger pool.

And that could be the root of it. Perhaps there was a major hardware flaw in the HBCC that triggered erroneous cache misses. They got a pile of cards ready to ship so they sell them under a limited edition moniker. Fix the bug in another silicon revision and performance could improve immensely. We just don't know until the final product is available.

Decreasing latency mitigates how long it takes for the GPU to move forward when a cache miss occurs. This doesn't increase framerate (unless it is happening regularly) but it increases minimum FPS.

I seriously doubt it. A cache miss from GPU memory over the PCIe bus to system memory will get close to a millisecond, while a cache miss for CPU to its memory is a little over 50 ns (~200-250 clocks wasted for Kaby Lake). Such cache misses for HBC would not only result in stutter, but a completely unplayable game.

My point is that 13 TFLOPS of compute power shouldn't result in as low of a framerate as it gets. There's clearly something wrong in the Frontier Edition and AMD knows it.

That's nothing new. Fury X had like ~53% more Flop/s than GTX 980 Ti, so things have been "wrong" for a while.

And BTW, back in the days we measured Flop/s for base clock, then for typical boost, and now AMD operates with max boost. So you wouldn't even hit 13.1 TFlop/s unless you increase the power limit to make it stay at 1600 MHz. We should really call it a 11.3 TFlop/s card, based on rated typical boost clock.

 

[FordGT90Concept]

That's nothing new. Fury X had like ~53% more Flop/s than GTX 980 Ti, so things have been "wrong" for a while.

Sad but true. Fiji had seriously underutilization problems, even at 4K.

 

[cdawall]

You didn't get it. The point of caching is to utilize a small fast storage pool as if it was a larger fast storage pool, by using a small fast pool as a buffer for the large one.

So by Vega having 8 GB HBM with caching, it will only act as if the memory pool was larger, caching will never give you more performance than having the larger pool

Bullshit lower latency memory access will always give better performance that is the entire point

 

[RejZoR]

What precisely are you talking about here? Is it the tiled rasterization again? Do you even know what it is? No such feature is implemented in a driver; it's a hardware scheduling feature.


They've had working hardware to test since last November, even demonstrated it working in December. That's about nine months of polishing the driver, which is more than they usually need. And it's not even a new architecture.


Hopefully they will, because going with HBM has been their greatest mistake with Vega. It will help with supply and cost, but it wouldn't help with consumption, performance, etc.

Just because something is done in hardware, it doesn't mean it just magically works. The driver needs to be aware of the thing. And the "it's not even a new architecture" is getting very old. Everyone parroting how it's not a new architecture and yet they changed nearly everything in the core. Just because the fundamentals are not changed much, that doesn't mean you can just slam a Fury X driver on it and it'll just magically work. AMD didn't use tiled rasterizing ever before. And like it was already said, the feature, while hardware implemented needs driver awareness. Pixel shaders have also always been hardware feature, but they wouldn't work if they weren't exposed in drivers correctly. And there are probably many things AMD hasn't even mentioned that works underneath. With R9 200 and R9 300 series they could literally just slam a renamed driver on top because it's what they already developed drivers for in the past. RX Vega with all the changes doesn't have that luxury. And 9 months is really not a lot of time. It may sound a lot, but it's not when RTG is most likely understaffed/underfunded. That's another AMD problem, but that's the reality we are aware of and is no secret. That's something neither me or you can change, that's entirely up to AMD.

 

[notb]

Yeah, well, it's kinda very hard being a f**king AMD fanboy if you don't own a single fucking thing from them, don't you think?

There's no contradiction.

It seems a lot of your brain is in "AMD rulez" mode.
But then you go to a shop, you reach for your wallet... and you buy NVIDIA/Intel. That's where the (usually shouted down) sensible part of your brain decides on financial decisions. And this is a good sign.

I'm a Mazda MX-5 fanboy and I don't own one. I might never do.
I bought a new Toyota in January and for more or less the same money I could have bought a used MX-5 from few years back.
And I have to admit: every time I see an MX-5 I wish I had one. But I made a sensible choice. And I'm really glad I did.

It's like they don't get it that while core is the same, Vega FE was released with different tasks in mind, meaning drivers could in fact be far more primitive and "half baked" and still work for what it was meant. As it's evident from tests where they actually tested shit that aren't games.

Problem is: the "different tasks in mind" is a theory you are popularizing. AMD said this card is - among other things - designed for creating, testing and optimizing games.
If this card is aimed at game developers, shouldn't it be the fastest Vega available? How will a Vega FE user be able to test a game that maxes RX Vega out, if he can't run it?

Good f**king question, don't you think? It's you people who accuse me of being a massive AMD fanboy. Hard to hold that narrative against someone with Intel CPU and NVIDIA GPU, isn't it?

Not at all. Still, this isn't an answer to my question.
Why don't you buy some AMD gear if you like it so much?

 

[efikkan]

Bullshit lower latency memory access will always give better performance that is the entire point

No, you still don't get what caching does.
If you have two comparable GPUs, GPU A have 12 GB, and GPU B have 8 GB + caching, the caching will try to weigh up for the missing memory in GPU B. Whenever you need less than 8 GB, there will be no difference, and when you need more GPU B will perform up to the level of GPU A, never above it. Your confusion is what to compare it to. HBC will not have lower latency than other GPU memory, only lower latency than falling back to system memory.

Just because something is done in hardware, it doesn't mean it just magically works. The driver needs to be aware of the thing.

The driver is aware of the hardware capabilities, but it does not micro-manage low-level scheduling inside the GPU, that is controlled on the GPU side. Tiled rasterization is not a new unit with a new feature set to expose through an API, it's a reordering of operations inside the GPU.

And the "it's not even a new architecture" is getting very old. Everyone parroting how it's not a new architecture and yet they changed nearly everything in the core. … And 9 months is really not a lot of time.

It has been enough in the past, and it's not like they start from scratch when the working chips arrive. Remember, they did demo it working in late December. Well at least this time with all the delays, the driver should be ~2.5 months more mature than the drivers of Polaris and Fiji at their respective releases.

But this boils down to what we've heard for every single generation from AMD the last five years; at release AMD fans say we can't judge it, because the driver are immature. Yet, they somehow "know" it will improve, we only need to give it more time, but no substantial improvement ever materializes.

 

[rtwjunkie]

So which game features have AMD intensionally disabled on the card they call optimized for gaming?

Ok, now you are becoming as much Court Jester as he is. They never said Frontier Edition is optimized for gaming. You are reading into it. They say it is optimized for every stage of the game production process.

As I've pointed out before, and as you have ignored, game designers don't have to be able to play a game at top-level quality to give you that to play.

 

[RejZoR]

@effikan
Sure, then explain to me what they are doing this whole extra month? Drinking booze and laughing? If they knew this is what they have, they'd just release it. It would be of less of an embarrassment than delaying it for whole month and then releasing the exact same thing as we've already seen with Vega FE, just with half the memory. AMD made some questionable decisions in the past, but they aren't that dumb, you can be assured of that. If I'm aware of those things, someone paid several times as much as I am sure as hell knows that. But entire computer world seems to be entirely oblivious to those tiny facts. If everything was where it should have been, they'd release entire Vega range back then and call it a day. Even if availability would actually come later if HBM2 production is the real issue. But whatever, apparently thinking logical isn't what people are expected to do over here anymore... You can do all the math and power draw and whatever, but tell me, this aspect doesn't strike you as very odd?

People were wondering what's up when RX480 was the fastest thing they offered and was really just a mid range. It didn't really bother people that there isn't any top end. The user base simply adapted to the offerings. If AMD pulled the same thing with RX Vega, release it as GTX 1080 competitor with engaging pricing scheme, people would be all over it even if it wasn't king of the hill. And yet they aren't doing that either. So, clearly something is going on. because otherwise, they could've done all of it long ago.

 

[efikkan]

Ok, now you are becoming as much Court Jester as he is. They never said Frontier Edition is optimized for gaming. You are reading into it. They say it is optimized for every stage of the game production process.

I'm just going to refer to AMD once again, it's even designed for "playtesting" and "performance optimization".

It can't get any clearer than that. Anyone failing to understand that are having trouble with fundamental logic.

Sure, then explain to me what they are doing this whole extra month? Drinking booze and laughing?

Please try to stay serious.
They are stockpiling cards for the launch.

 

[RejZoR]

Because paper launches haven't been invented yet apparently...

 

[rtwjunkie]

I'm just going to refer to AMD once again, it's even designed for "playtesting" and "performance optimization".

As I said, logical thought is something you need to strive for. Design team members play testing is NOT gaming. Playtesting is checking bugs and implementation of new things into the program as they go along in production.

It is not actual gaming as either us as consumers will do, or quality control testers will do when production is near finished.

 

[cdawall]

No, you still don't get what caching does.
If you have two comparable GPUs, GPU A have 12 GB, and GPU B have 8 GB + caching, the caching will try to weigh up for the missing memory in GPU B. Whenever you need less than 8 GB, there will be no difference, and when you need more GPU B will perform up to the level of GPU A, never above it. Your confusion is what to compare it to. HBC will not have lower latency than other GPU memory, only lower latency than falling back to system memory.

So it performs up to the same as a 12gb gpu while only having 8gb that would be a performance increase over a standard 8gb card no? Hence the whole reason amd has done it.

 

[efikkan]

As I said, logical thought is something you need to strive for. Design team members play testing is NOT gaming. Playtesting is checking bugs and implementation of new things into the program as they go along in production.

It is not actual gaming as either us as consumers will do, or quality control testers will do when production is near finished.

It's amazing when people fail to understand plain text. This has got to be the most ridiculous things I've heard this year.

Anyone familiar to development knows you do performance optimization on hardware representative of what the end user will run. If Vega FE lacks gaming features of RX Vega, then it's completely useless for performance optimizations, which AMD claims is it's intended use.

 

[londiste]

hbcc seems utterly useless in gaming use cases. caching is all nice and dandy but in case of vega, that cache is vram. it will not help performance if there is enough and there will still be latency issues if there is not enough. optimizing for cleaning up allocated but not used swaths of memory will undoubtedly cause its own fair share of issues in addition to this needing to be written for on software side (not very likely with this being a feature only in high end even on amd side of things).

however, there are excellent use cases for hbcc when it comes to compute, especially with the memory access/addressing improvements of last few generations that will be able to make use of system (and other bits of) memory in a single pool.

 

[R0H1T]

hbcc seems utterly useless in gaming use cases. caching is all nice and dandy but in case of vega, that cache is vram. it will not help performance if there is enough and there will still be latency issues if there is not enough. optimizing for cleaning up allocated but not used swaths of memory will undoubtedly cause its own fair share of issues in addition to this needing to be written for on software side (not very likely with this being a feature only in high end even on amd side of things).

however, there are excellent use cases for hbcc when it comes to compute, especially with the memory access/addressing improvements of last few generations that will be able to make use of system (and other bits of) memory in a single pool.

HBCC is mostly an enterprise feature, I would've thought that AMD might've included SLC cache or something for Vega to make use of HBCC.
It's that or I'm reading HBCC wrong /:

 

[notb]

@effikan
Sure, then explain to me what they are doing this whole extra month? Drinking booze and laughing?

Let's hope they are designing something good.
As for the launch... in the optimistic variant: building inventory, shipping, preparing benchmarks for the launch even. In pessimistic one: waiting for HBM2 supply...

If they knew this is what they have, they'd just release it.

Maybe they're hoping for a miracle?
They must have already had, since they've continued developing this card. They must have noticed months ago how will the power draw look for the performance they aimed at.
As it's been told already: a dual RX480 could be better. AMD's board or shareholders wanted a Vega release in time (to show this architecture actually works) and a launch of gaming model with solid inventory for preorders (like they did with Ryzen).

If everything was where it should have been, they'd release entire Vega range back then and call it a day. Even if availability would actually come later if HBM2 production is the real issue. But whatever, apparently thinking logical isn't what people are expected to do over here anymore... You can do all the math and power draw and whatever, but tell me, this aspect doesn't strike you as very odd?

That could be just about accomplishing targets.

People were wondering what's up when RX480 was the fastest thing they offered and was really just a mid range. It didn't really bother people that there isn't any top end. The user base simply adapted to the offerings. If AMD pulled the same thing with RX Vega, release it as GTX 1080 competitor with engaging pricing scheme, people would be all over it even if it wasn't king of the hill. And yet they aren't doing that either. So, clearly something is going on. because otherwise, they could've done all of it long ago.[/QUOTE]

HBCC is mostly an enterprise feature, I would've thought that AMD might've included SLC cache or something for Vega to make use of HBCC.
It's that or I'm reading HBCC wrong /:

Vega FE is not exactly and enterprise-grade product. Sure, it'll be used in some workstations, but the general corporate audience would prefer something more FirePro-ish.
This would mean that Vega FE is in fact not a product. It's just a showcase of technologies that AMD has and can use in future products. I wouldn't be shocked - this kind of launches happen quite often.

 

[londiste]

HBCC is mostly an enterprise feature, I would've thought that AMD might've included SLC cache or something for Vega to make use of HBCC.
It's that or I'm reading HBCC wrong /:

slc cache would not be helpful. in this case where vram itself acts as cache, next level of memory is system ram over pci-e. this is much faster than slc cache.

 

[R0H1T]

slc cache would not be helpful. in this case where vram itself acts as cache, next level of memory is system ram over pci-e. this is much faster than slc cache.

Yes it is but unless Vega reserves a part of System RAM for cache, like primocache, I don;t see how HBCC would work more efficiently with it than say a dedicated pool of memory or storage for caching the program (or games) & the various chunks of data it'll work on, like content creation or editing. Also I haven't seen HBCC in action, or a thorough review, so I;m waiting to see how it actually works.

 

[londiste]

Yes it is but unless Vega reserves a part of System RAM for cache, like primocache, I don;t see how HBCC would work more efficiently with it than say a dedicated pool of memory or storage for caching the program (or games) & the various chunks of data it'll work on, like content creation or editing. Also I haven't seen HBCC in action, or a thorough review, so I;m waiting to see how it actually works.

gpus have been able to access system ram for a while now. as far as i understand the idea behind hbcc is not that system ram is used as a cache but vram is used as cache for everything further away - system ram, vram on other vega cards (either of them could be over pci-e or infinity fabric), storage of any kind etc. that has definite benefits for compute purposes - for example storage arrays being used for data and only necessary bits taken into vram at a time. data streaming has been doable for a while but this should make it much more seamless.