12nm polaris 30 rumours

GTwannabe said:
Would not be competitive given the high cost of interposers and HBM2.
Click to expand...
They could have die shrink Vega to 12nm and couple it with GDDR5X or GDDR6 memory. 14Gbps GDDR6 on 384-bit bus yield 672GB/s bandwidth, higher than current 2048-bit bus 1.89Gbps that could only give 483GB/s bandwidth.

Not to mention most of Vega feature (NGG fastpath, DSBR) is disabled/malfunction on current variant.
 
I wish they'd just stop with the HBM crap and work with GDDR5X or GDDR6. They're both putting out more than enough bandwidth. HBM is too expensive and the memory bandwidth for the price has been .... disappointing, to say the least.
 
Nunz said:
I wish they'd just stop with the HBM crap and work with GDDR5X or GDDR6. They're both putting out more than enough bandwidth. HBM is too expensive and the memory bandwidth for the price has been .... disappointing, to say the least.
Click to expand...

but needed if they ever get MCM to work

like NV is working on RT AMD is working on HBM and may take a few generations
and it's only been two at least they didn't price them at 1200 bucks

and the memory bandwidth on the Vega 20 looks great
1TB/sec+ of memory bandwidth
Click to expand...

https://www.tweaktown.com/news/59079/vega-20-coming-2018-7nm-1tb-sec-hbm2-much-more/index.html

almost 2x of the 616gb/sec memory bandwidth RTX 2080 ti
 
Last edited:
The memory bandwidth for Vega20 at over 1TB/sec is great if that's the bottleneck on the card. On the Vega64 it doesn't seem to be the case. While Flyordie did show some interesting numbers regarding OCing the memory and the performance increases that come with, I'm still ... unconvinced, I guess, that memory bandwidth is the bottleneck. Maybe in synthetic benchmarking and the like, but in actual real-world gaming I still think there are other areas that need improving.

HBM is also making the cards less profitable because of the price involved. I mean they can work on HBM while still using GDDR5X/GDDR6 until it's ready. As it stands, HBM just isn't ready imo from a $/perf perspective, at least in terms of profit margin for AMD.
 
Nunz said:
The memory bandwidth for Vega20 at over 1TB/sec is great if that's the bottleneck on the card. On the Vega64 it doesn't seem to be the case. While Flyordie did show some interesting numbers regarding OCing the memory and the performance increases that come with, I'm still ... unconvinced, I guess, that memory bandwidth is the bottleneck. Maybe in synthetic benchmarking and the like, but in actual real-world gaming I still think there are other areas that need improving.

HBM is also making the cards less profitable because of the price involved. I mean they can work on HBM while still using GDDR5X/GDDR6 until it's ready. As it stands, HBM just isn't ready imo from a $/perf perspective, at least in terms of profit margin for AMD.
Click to expand...

either is RT

and would say HBM is more usable now than RT will be for a year or two

and maybe a AMD MCM in 2020 with RT if they can find the right glue
 
Nunz said:
The memory bandwidth for Vega20 at over 1TB/sec is great if that's the bottleneck on the card. On the Vega64 it doesn't seem to be the case. While Flyordie did show some interesting numbers regarding OCing the memory and the performance increases that come with, I'm still ... unconvinced, I guess, that memory bandwidth is the bottleneck. Maybe in synthetic benchmarking and the like, but in actual real-world gaming I still think there are other areas that need improving.

HBM is also making the cards less profitable because of the price involved. I mean they can work on HBM while still using GDDR5X/GDDR6 until it's ready. As it stands, HBM just isn't ready imo from a $/perf perspective, at least in terms of profit margin for AMD.
Click to expand...


I was merely showing that its compute engines are bandwidth starved. Gaming still suffers immensely because it lacks the pixel pushing power of the Nvidia cards.

Give Vega64 32 more ROPs and it'll mop the floor with anything Nvidia has. Problem is, its limited by 1. Silicon area. 2. Density. Maybe with 7nm they'll have what they need to get a 92 or w/e ROP Vega/Navi out. or maybe Navi is a gamer oriented card and forgoes some of the compute stuff to make room for the gamer equipment? Honestly, idk. Those are just rough guesses.

Basically, AMD went compute over pixel pushing power. Its shaders took up an immense amount of die space and they couldn't add more ROPs to it. As a result, Vega suffered on the gaming side.
 
Flyordie said:
I was merely showing that its compute engines are bandwidth starved. Gaming still suffers immensely because it lacks the pixel pushing power of the Nvidia cards.

Give Vega64 32 more ROPs and it'll mop the floor with anything Nvidia has. Problem is, its limited by 1. Silicon area. 2. Density. Maybe with 7nm they'll have what they need to get a 92 or w/e ROP Vega/Navi out. or maybe Navi is a gamer oriented card and forgoes some of the compute stuff to make room for the gamer equipment? Honestly, idk. Those are just rough guesses.

Basically, AMD went compute over pixel pushing power. Its shaders took up an immense amount of die space and they couldn't add more ROPs to it. As a result, Vega suffered on the gaming side.
Click to expand...

I dont know if the gcn arch can exceed 64 rops, i remember reading something a while back that 64 is the max. Makes sense since amd has been stuck on 64 rops since hawaii.
 
Higgy10 said:
I dont know if the gcn arch can exceed 64 rops, i remember reading something a while back that 64 is the max. Makes sense since amd has been stuck on 64 rops since hawaii.
Click to expand...
The ROPS are tied to Shader Engines, which is limited to 4 Shader Engines with each Shader Engines having 16 ROPS.

The Vega's NCU (upgraded GCN) is supposed to remove this 4 Shader Engine limitation, but their argument being wanted to offset it via other means like HBCC and rasterization rather than making more Shader Engines which IMO is silly. For compute yes it isn't ROP bound (mostly) but for games at high resolution its another story.
 
Pretty sure its not limited Id read its just a ratio issue which entails die space. They could do a 96 or 128 rop under GCN but at 128 they would have to double the shaders. While itd be cool to see 8000+ shaders not sure that even 7nm can accommodate that and still have space for Raytracing related hw engines.

Unless they intend to run RT on shaders...
 
Vega/GCN's biggest problem is it doesn't clock high enough. It runs at only 1500, maybe 1600 MHz maximum (rare). And even those clocks are pushing too much voltage resulting in poor efficiency (at lower frequency Vega is basically as efficient as Pascal).

Compare that to a 2080 or 2080 Ti running at 2100 and 2000 MHz respectively. 2080 Ti at the same die size is about 30% faster and not coincidentally it clocks around that same amount higher.

It was the same story with Fury vs 980 Ti. Clock for clock GCN does more work, but the clocks are too low. However, that could very well be an inherent limitation of the architecture, thus requiring a revamp from the ground up.

Still, I suspect that is why Vega didn't have more CUs, etc. It shouldn't have been necessary had the silicon clocked higher. I think that Vega just didn't run at the frequency they hoped it would. Granted, this is just speculation on my part.
 
Really hope we get to see that pro 7nm vega card hacked with gaming drivers. or at least tested with some load bearing testing app to see if it can clock up. Story was vega didnt clock well because they werent able to optimize it as they moved that team of engineers to work on ryzen.

Isnt it supposed to be out this month?
 
https://videocardz.com/78583/amd-radeon-rx-590-spotted-at-3dmark-database

What is Radeon RX 590? Well, no one really knows. Rumors about a ‘new’ Radeon card started appearing last week (actually they started appearing a few months ago, but it was hard to believe at the time — still is).

What is believed now, however, is that Radeon RX 590 would feature Polaris architecture, except on 12nm the fabrication node. This would improve performance and power efficiency, but it would certainly not provide a proper response to NVIDIA’s Turing.

The entry at 3DMark database indicates that RX 590 would feature 1545 MHz clock, that’s 205 MHz higher than RX 580 boost clock.

The memory clock of 2000 MHz suggest there is no change to the memory subsystem, it is likely still GDDR5-based.

Performance wise, we are looking at ~10% improvement compared to RX 480. Our factory-overclocked RX 580 PowerColor Red Devil scores 4399 Graphics Score (4639 while overclocked).
Click to expand...
 
Nagorak said:
Vega/GCN's biggest problem is it doesn't clock high enough. It runs at only 1500, maybe 1600 MHz maximum (rare). And even those clocks are pushing too much voltage resulting in poor efficiency (at lower frequency Vega is basically as efficient as Pascal).

Compare that to a 2080 or 2080 Ti running at 2100 and 2000 MHz respectively. 2080 Ti at the same die size is about 30% faster and not coincidentally it clocks around that same amount higher.

It was the same story with Fury vs 980 Ti. Clock for clock GCN does more work, but the clocks are too low. However, that could very well be an inherent limitation of the architecture, thus requiring a revamp from the ground up.

Still, I suspect that is why Vega didn't have more CUs, etc. It shouldn't have been necessary had the silicon clocked higher. I think that Vega just didn't run at the frequency they hoped it would. Granted, this is just speculation on my part.
Click to expand...

AMD needs to massively improve their geometry performance. Throwing clockspeed at the problem is a bandaid at best.
 
I was browsing my local retailer on the AMD GPU section and I see most GPUS have taken a pretty steep price cut. Vega 56 originally $1079 AUD now going for $599 plus 3 games. :nuts:


RX 580 OC 8GB, $660, now $399 AUD. WTF is going on?

That's crazy talk.
 
Megaman said:
I was browsing my local retailer on the AMD GPU section and I see most GPUS have taken a pretty steep price cut. Vega 56 originally $1079 AUD now going for $599 plus 3 games. :nuts:


RX 580 OC 8GB, $660, now $399 AUD. WTF is going on?

That's crazy talk.
Click to expand...
GPU mining died. Bitcoin never "recovered".
 
Megaman said:
I was browsing my local retailer on the AMD GPU section and I see most GPUS have taken a pretty steep price cut. Vega 56 originally $1079 AUD now going for $599 plus 3 games. :nuts:


RX 580 OC 8GB, $660, now $399 AUD. WTF is going on?

That's crazy talk.
Click to expand...

yeah prices are looking quite cheap if you can find them in stock

how about $499 AUD for xfx vega 56 (tax inclusive)

https://www.pccasegear.com/category/193_1900/graphics-cards/radeon-rx-vega

Can't seem to find vega 64's in stock anywhere
 
Last edited:
You must log in or register to reply here.
Back
Top