Return to Level1Techs.com

The VEGA 56 / 64 Cards Thread! General Discussion

amd
vega

#191

I never said it was. Just that it is more efficient.

I forgot about that. Sure doesn't sound like they are focusing on neural networks then either.


#192

In its current form is is also not a card sold for gaming either.

HBM may use less power but placing it on the die concentrates the area of the card that is generating heat. Spreading the heat generation over a wider area will reduce the peak thermal loads that the cooler needs to deal with over a wider area, making it much easier to cool

If you re-read what I said, you will note that I did not say that they would not use HBM at all. I speculated that they may use 4GB of HBM on die and use other memory as a near to die cache to feed the 4gb of HBM and the GPUs immediate needs. A GPU cannot use all 4GB of data at once, let alone 8gb of data, so there is no reason why a 2 tiered frame buffer with intelligent switching could not work.

They have already shown that they have the technology to support that in the polaris workstation cards that they launched with the built in NAND flash. There is nothing that says that a frame buffer has to be populated with memory that is all the same speed

They made the mistake of selling Fury X cards with 4GB frame buffers when the market had already moved to 8GB. It is hurting those cards in 4K performance now so I doubt that AMD will make the same mistake twice


#193

The last time workstation cards came out first... It was:

....
Noone would freak out either lol, its normal for the workstation cards to be significantly lower clocked and lower performance than the gaming ones.
The Pro cards are tuned for power efficiency and pro uses not gaming so a 20-30% improvement for a gaming card on the same core wouldnt surprise me at all.


#194

We just have to wait for its official launch,
And then we will see benchmarks from more reputable sources to compair.


#195

Something interesting that has come out of this was some testing I saw on the AMD Subreddit that makes it appear as though Vega isn't using tile based rendering/rasterizing as they said it was going to....

Hmm. Will need more testing to see.

But honestly now this isn't impressive. This is worse than Fiji at the same clocks. Something isn't right


#196

I already explained why this does not make sense.


:open_mouth: If that is true then there is still hope. That would make a huge difference.


#197

Meh. Tile based rendering is not as exciting as Raja is trying to make you believe. Games are going to great lengths to ensure hidden objects are not rendered, so there's not too much the GPU can do here. Besides the main improvent over immediate rasterization is reduced bandwidth, which is probably not overly helpful on a card with HBM2.


#198

If they fucked up the driver we should see a fix for that or at least hear about it from AMD in a couple days or maybe a week. Because that is something game devs would probably be pissed about too.


#199

Umm it actually makes a pretty big difference...

That is where Maxwell get's alot of its huge efficiency jump over Kepler....

Bandwidth yes but also less calls to memory generally which saves power


#200

But maxwell cards have comparatively little memory bandwidth. Nvidia relies more on efficiency, while AMD appears to just try to fix it with faster memory.

I haven't seen any benchmarks specifically testing the impact of tile based rasterization so my knowledge here is purely theoretical. Maybe I'm underestimating the impact, but I don't think so.


#201

I noticed that but still not holding my breath for much more pref then 1080 and probably will cost more unless AMD is cool with slim margins


#202

I'm not holding my breath for more either lol

But I think they're alright with the margins. Epyc and Ryzen have such good yields and such high margins they can skimp a bit here. They just need Vega to tide them over until they can get IF on the GPU with Navi.

EDIT:

Apparently all previous die size estimations were really wrong. PcPer is reporting a 564mm die size... Holy hell this is bad lol

https://www.pcper.com/news/Graphics-Cards/Radeon-Vega-Frontier-Edition-GPU-and-PCB-Exposed


#203

Was pretty quick on the article.

Also glad that back end of the PCB is super bare, was looking forward to the smaller card lenght


#204

Ryan from pcper in the Livestream confirmed that he had talked to AMD and that the current and only driver out for the launch of Vega frontier edition does in fact have all of the Vega gaming optimizations done up till this point. Vega frontier edition is not targeted at gamers, but it is the full fat Vega core running with gaming optimized drivers, so it is reasonable to extrapolate from Vega frontiers performance numbers as to what RX Vega performance numbers will look like.


To the point that "well the FirePro W9100 used more power than the R9 290x but was based on the same core and clocked lower," that is true, but the Frontier Edition is not 100% analogous to a FirePro card. The Radeon Pro line is the equivalent to the old FirePro line, and the frontier edition isn't in the Radeon Pro line. FirePro and Radeon Pro cards have ECC memory, the Frontier Edition does not. The Frontier Edition gets a gaming mode in its driver with a gaming optimizations. It's not a firepro equivalent, and the comparison isn't one to one.


#205

Dat VRM though. Overbuilt as fuck.

Taken from nVidia's white paper but this is what they say about HBM2:

Another benefit of HBM2 memory is native support for error correcting code (ECC) functionality. ECC
provides higher reliability for compute applications that are sensitive to data corruption. It is especially important in large-scale cluster computing environments where GPUs process very large datasets and/or run applications for extended periods.

ECC technology detects and corrects single-bit soft errors before they affect the system. In comparison, GDDR5 does not provide internal ECC protection of the contents of memory and is limited to error detection of the GDDR5 bus only. Errors in the memory controller or the DRAM itself are not detected.
~ https://images.nvidia.com/content/pdf/tesla/whitepaper/pascal-architecture-whitepaper.pdf

Did AMD offer any official word on HBM2 implementation on Vega FE?


#206

AFAIK no nothing like an official statement or a spec page.

However, in the AMA on Reddit, someone from AMD when asked if FE had ECC enabled, said it did not and then proceeded to say "Why would you want it?."

That makes me kinda feel like they have no clue wtf they are talking about but it could still also be true

Another weird thing was that PcPer said, apparently quoting from AMD, that changing the driver mode doesn't actually do anything besides changing the UI. Which seems weird as hell. Why even bother with the different modes then?


#207

Maybe, because its not a pro card as much as it's a gaming card that they're trying to spin for professional use?


#208

That is exactly what I think. It is effectively the top end RX-Vega with 10 bit color.


#209

I'm not going to say that person was lying but I would rather wait for a proper white paper release to know what features are actually included on Vega FE. Radeon Instinct MI25 is a very similar design and does have ECC enabled so perhaps it's all software layer. With the official classification on Vega FE being "HBC" they're marketting this as if the HBM is operating as an L4 passthrough cache level for system memory, kinda like eDRAM on Skylake. While I admit I don't know what workload that benefits it probably isn't gaming.


#210

Yeah I'm waiting as well. Whole thing has been a bit of a mess. No point in getting nuts now. Just wait and see.

Working with huge data sets and shit for rendering and scientific applications. But yeah not gaming