TR Forums

https://gpuopen.com/compute-product/amd ... hitecture/

Here's the RDNA architecture. There have been some changes to the GCN ISA (removal of CBranch_force / CBRANCH_Join, working 32-at-a-time for bpermute / permute / DPP instructions, etc. etc.) but its fundamentally still GCN assembly under it all. The architectural changes are quite severe, with Warp32, three-level cache (labeled L0, L1, and L2), 128kB LDS, and more.

Purported upcoming Asus and XFX aftermarket-cooled 5700 series cards.

On that note, what's the deal with non-reference graphics cards suddenly ALL being over-height?? My Silverstone GD-07B is somewhat height-constricted (CM 212-series HSFs do not fit), and it's looking like finding aftermarket-cooled GPUs is going to be a PITA moving forward.

The Egg wrote:

On that note, what's the deal with non-reference graphics cards suddenly ALL being over-height?? My Silverstone GD-07B is somewhat height-constricted (CM 212-series HSFs do not fit), and it's looking like finding aftermarket-cooled GPUs is going to be a PITA moving forward.

They're all going for larger fan diameter so they can run the fans at lower RPM while maintaining total airflow. Restricting the cooler height necessitates lower fan diameter, which means higher RPM for a given amount of cooling, which means higher noise.

K-L-Waster wrote:

The Egg wrote:

On that note, what's the deal with non-reference graphics cards suddenly ALL being over-height?? My Silverstone GD-07B is somewhat height-constricted (CM 212-series HSFs do not fit), and it's looking like finding aftermarket-cooled GPUs is going to be a PITA moving forward.

They're all going for larger fan diameter so they can run the fans at lower RPM while maintaining total airflow. Restricting the cooler height necessitates lower fan diameter, which means higher RPM for a given amount of cooling, which means higher noise.

Correct. When I bought an RTX 2060 for an office workstation a few weeks ago I was startled to find it was a triple height card, and it's because of this very thing. As for paying for a 2060, it wasn't my money and it does the CUDA job we needed 20 times faster than leaving it to the CPU in that employee's workstation...

K-L-Waster wrote:

The Egg wrote:

On that note, what's the deal with non-reference graphics cards suddenly ALL being over-height?? My Silverstone GD-07B is somewhat height-constricted (CM 212-series HSFs do not fit), and it's looking like finding aftermarket-cooled GPUs is going to be a PITA moving forward.

They're all going for larger fan diameter so they can run the fans at lower RPM while maintaining total airflow. Restricting the cooler height necessitates lower fan diameter, which means higher RPM for a given amount of cooling, which means higher noise.

I figured that was the case, but some are still more egregious with wasted height than others. The Asus in the link, for instance, has probably 2cm above and below the fan blades. It definitely wouldn't fit in my case, and I'm not sure why they would assume everyone has that much height to spare. My guess is they got overly excited about being able to use moar lights and shrouds.

The XFX is debatable on whether it would fit. Does anyone know the exact method they use to measure card height (i.e., to the bottom of the PCIe pins, just the body of the card, etc)?

The Egg wrote:

K-L-Waster wrote:

The Egg wrote:

On that note, what's the deal with non-reference graphics cards suddenly ALL being over-height?? My Silverstone GD-07B is somewhat height-constricted (CM 212-series HSFs do not fit), and it's looking like finding aftermarket-cooled GPUs is going to be a PITA moving forward.

They're all going for larger fan diameter so they can run the fans at lower RPM while maintaining total airflow. Restricting the cooler height necessitates lower fan diameter, which means higher RPM for a given amount of cooling, which means higher noise.

I figured that was the case, but some are still more egregious with wasted height than others. The Asus in the link, for instance, has probably 2cm above and below the fan blades. It definitely wouldn't fit in my case, and I'm not sure why they would assume everyone has that much height to spare. My guess is they got overly excited about being able to use moar lights and shrouds.

The XFX is debatable on whether it would fit. Does anyone know the exact method they use to measure card height (i.e., to the bottom of the PCIe pins, just the body of the card, etc)?

Wouldn't you want a reference cooler for small-form-factor anyway? Blower coolers IIRC are superior for tight quarters.

dragontamer5788 wrote:

Wouldn't you want a reference cooler for small-form-factor anyway? Blower coolers IIRC are superior for tight quarters.

The case is a full ATX. It's an HTPC desktop-style (as compared to tower) though, so it doesn't have gobs of extra height available for tall cards and HSFs. I'm currently on my 3rd GPU in there, and this is the first time I've had to take height into consideration.

The Egg wrote:

dragontamer5788 wrote:

Wouldn't you want a reference cooler for small-form-factor anyway? Blower coolers IIRC are superior for tight quarters.

The case is a full ATX. It's an HTPC desktop-style (as compared to tower) though, so it doesn't have gobs of extra height available for tall cards and HSFs. I'm currently on my 3rd GPU in there, and this is the first time I've had to take height into consideration.

I had a similar problem when getting a GPU for my GD09B. I ended up going with an EVGA as they were one of the few card makers that stuck to the height of the PCI tiedown. I don't think EVGA makes AMD based cards, though, so not much help in this case.

AMD released more info here: https://gpuopen.com/wp-content/uploads/ ... public.pdf

Page 15 is the big thing: it requires 1/4th the number of threads to have full utilization of the GPU. Since I'm targeting my Vega64, that means I have to figure out how to use (at minimum) 16384 SIMD-threads on my GPU before the resources are utilized (x4 for memory hiding, so 64k SIMD threads are targeted in my code)

A hypothetical RDNA-64 GPU would only need 4096 SIMD-threads to become fully utilized. Occupancy x4 would be 16384 threads, which would be easier. I guess in both cases, the number of threads fits inside of a 16-bit number, so its not a very big deal... but a lot of parallel-structures / parallel algorithms require communication. Cutting down the minimum number of threads to get things working will lead to various efficiency boosts: at a minimum, the RDNA-based code can potentially use 1/4th the per-thread memory.

This was a big advantage for NVidia devices, which have been Warp32 for a long time, as well as single-issue. So AMD can catch up to single-threaded performance / single-threaded latency finally. AMD managed to figure out an architecture that allows them to catchup, while largely retaining backwards compatibility as well, so its kind of impressive.

Alright my 5700 XT should be here Wednesday.

I forgot XFX was an American company, heh. Can't help but be skeptical about that but what the hell.

Heh. I liked my system before. But now it's damn near flawless. *grin*

I think I can now stop buying things for at least five minutes.

(I actually plan this build to be finished for a very long time, so we'll see how everything goes.)

Very nice. Obviously I can max everything without even thinking about it.

Still a few new gen glitches for them to work out, but eh.

Well I think I found the one game I play that really doesn't like this card much: Quantum Break is a stuttery mess on here at 1440p.

I'm assuming (and hoping) Control will have a much better engine.

Flashing a Radeon RX 5700 graphics card with just the right Radeon RX 5700XT BIOS can allow the adventuresome to overclock the heck out of it.
https://www.techpowerup.com/review/flas ... nce-guide/

Reminds me of my ATI 9700 flashed to 9800.

Radeon RX5500 looks very decent for the price and power consumption,
https://www.techpowerup.com/261371/amd- ... -as-rx-580

JustAnEngineer wrote:

Radeon RX5500 looks very decent for the price and power consumption,
https://www.techpowerup.com/261371/amd- ... -as-rx-580

About the same performance as the RX 580 8gb.

Aranarth wrote:

About the same performance as the RX 580 8gb.

For significantly less, if the expected prices are accurate.

Plus the not needing a bigger power supply and etc thing.

I love my 5700 XT so far. Lack of HDMI 2.1 is the only thing worth complaining about (or Quantum Break, but one game...heh.)

JustAnEngineer wrote:

Radeon RX5500 looks very decent for the price and power consumption,
https://www.techpowerup.com/261371/amd- ... -as-rx-580

Yeah, it seems to bounce between the 580 and occasionally give the 590 a run for its money within a tighter thermal envelope. If the prices were more in line with what it's replacing and it had day one Linux support without feeling like I'd be taking a gamble, I'd grab one for myself. As it stands, I'm tempted to stretch a little further, grab an RX 5700 for my main box, and thump the Vega into my Linux machine.

Aranarth wrote:

JustAnEngineer wrote:

Radeon RX5500 looks very decent for the price and power consumption,
https://www.techpowerup.com/261371/amd- ... -as-rx-580

About the same performance as the RX 580 8gb.

RX 580 8GB offers 20% better perf at 4K. Gotta love this card. Still going strong despite being pretty cheap.

Man. I like AMD and all but this whole scene feels like a crapshow right now. I hope the March drivers are much more solid.

I suspect almost all of the black screen issues stem from super-aggressive power management (set by default) not dancing with GDDR6 chips on some factory-overclock SKUs. The WDM on Windows 10 puts out a surprising amount of GPU utilization if it needs to pool enough windows and screens (a bit more than idle state). This is why multi-monitor users seem to be more problems with it. IIRC, Nvidia had its own share of aching pains (instability, artifacting)with GDDR6 on their first batch of Turing cards but quickly cleared up. Reference design cards don't seem to suffer from the issues. The symptoms are eerily similar to overclockers/undervolters who push their card too hard.

I'm running a driver from last year on my Radeon VII. The newer drivers (December, January) would cause random lockups under all sorts of varied load and idle conditions. Stock clocks, reference card under water. Perhaps the next driver set will be better.

Knock on wood - so far both my Vega 56 and the Radeon Fury I snagged for cheap a few weeks ago have been as solid as ever on Win10. I'd say it sounds like GDDR6 is the culprit, but Radeon VII users have had problems too, so I've honestly got no idea. So far I'm just glad it hasn't gotten me, too.

Concupiscence wrote:

Knock on wood - so far both my Vega 56 and the Radeon Fury I snagged for cheap a few weeks ago have been as solid as ever on Win10. I'd say it sounds like GDDR6 is the culprit, but Radeon VII users have had problems too, so I've honestly got no idea. So far I'm just glad it hasn't gotten me, too.

HBM2 is a bit touchy with undervolting/underclocking which is why there isn't really much of a clockspeed gradient between idle and loaded states and options are rather limited within Wattman.

Krogoth wrote:

Concupiscence wrote:

Knock on wood - so far both my Vega 56 and the Radeon Fury I snagged for cheap a few weeks ago have been as solid as ever on Win10. I'd say it sounds like GDDR6 is the culprit, but Radeon VII users have had problems too, so I've honestly got no idea. So far I'm just glad it hasn't gotten me, too.

HBM2 is a bit touchy with undervolting/underclocking which is why there isn't really much of a clockspeed gradient between idle and loaded states and options are rather limited within Wattman.

Ah. I believe that. If that's not true for HBM1, that'd explain why the guy who sold it to me warned about the current driver and recommended going back to last year's. I took a gamble and it's been okay, but others' mileage may vary.

I solved almost all my black screen issues by running games in a window.

Turning HDR off can also help, though it's not a cure.

This is not what I would've expected out of a GPU launch, nevermind past that.

Does Adrenalin 2020 Edition 20.1.3 have these black screen issues too? If so, how is it passing through the WHQL validation process?

Igor_Kavinski wrote:

Does Adrenalin 2020 Edition 20.1.3 have these black screen issues too? If so, how is it passing through the WHQL validation process?

Yes, though it didn't seem to be as bad.

I wish I knew what was happening. Or if we can expect a full fix in March. If not I might get annoyed.

LostCat wrote:

Igor_Kavinski wrote:

Does Adrenalin 2020 Edition 20.1.3 have these black screen issues too? If so, how is it passing through the WHQL validation process?

Yes, though it didn't seem to be as bad.

I wish I knew what was happening. Or if we can expect a full fix in March. If not I might get annoyed.

Could be related to the video memory usage. Have you tried double buffering instead of triple buffering? Or try triple buffering if your settings were set for double buffering.