I think AMD will not sit idly by and let whatever negative impressions of vega stand without trying to fix or improve on it. Kind of like Nvidias situation with fermi, gtx480 > 580. If a refresh is out over next few weeks/months, what sort of improvements would we be looking at? Will it all come from improved binning, lower voltages or are there other areas that may also be improved/addressed?
-
We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.
Vega refresh - Expected? How might it look?
- Thread starter amenx
- Start date
HBM has been the thorn in AMD's side on this one. Not meeting spec, requiring overvolting to even come close to spec, etc. Yields must also blow because in true AMD form the core is overvolted as well.
Best move for AMD would be to release a nano that's clocked lower, binned, and has extremely low power consumption. It would compete with the 1070 but potentially could have extremely low power consumption right from the factory.
Best move for AMD would be to release a nano that's clocked lower, binned, and has extremely low power consumption. It would compete with the 1070 but potentially could have extremely low power consumption right from the factory.
Last edited:
I think this would be a horrible decision, whatever they do is going to take a few months, and in a few months they will have Volta to compete with not the current gen of Nvidia cards. Having a Vega thats competing with a 1070 is not going to be enough, this would leave Nvidia with the 1080/1080Ti/Titan, and then volta 2060/2070/2080 all ahead of them. You dont want to be in a situation where your only competitor has 6 cards faster than your fastest offering.
I think they should push out another respin as fast as possible and aim for 1080Ti level of performance or Volta is going to murder them. They need to get clock speed up and power usage down ASAP. Luckily for them 6 months should be enough time to get a respin out. They need at least one top tier card in their linup, having only a 1070 competitor will not be enough IMO.
They can release a Nano in a month or two with their current process and available HBM. It could actually be a competitive product.
Maybe they can get power usage down and clock speed up but that remains to be see. They couldn't do it with Vega despite a very mature node and yields must be horrible given AMDs long standing practice of over-volting. On top of that AMD is dealing with all the problems HBM brings to the table and nobody knows if even six months will be long enough for a refresh ready HBM to be available.
I have news for you, NVIDA is going to destroy AMD this generation and next no matter what straws AMD grasps at. Their only choice will be to sell at cost as soon as the mining craze dies down.
AMD have a lot of problems to fix with Vega
1. Sub par drivers. AMD have launched Vega with really poor drivers. Here is a solid example.
https://www.techpowerup.com/reviews/AMD/Radeon_RX_Vega_64/10.html
Primitive shaders which unlocks the improved geometry performance of Vega is not enabled in drivers. The features that are enabled such as Draw stream binning rasterizer are not well optimized. AMD needs to improve their drivers and extract the best possible performance out of Vega hardware.
2. HBM2 clocks - AMD's memory partners are struggling with HBM2 clocks/yields . Vega 10 was supposed to launch with 2 Gbps HBM2. Instead they have launched with 1.89 Gbps , which in fact was achieved through overvolting. By mid-2018 AMD should get its HBM2 vendors Samsung and Hynix to deliver 2 Gbps HBM2 stacks at good volume and healthy yields.
3. Process and clocks - Vega on 14LPP is a turd. The card draws too much power as it cannot hit its targetted clocks at reasonable voltage/power. AMD has a 14nm+ process (most likely GF 14HPP) for its 2018 products. That process needs to deliver good clocks at reasonable voltage/power.
If AMD can achieve all of these by mid-2018 they might be able to atleast try and compete with GV104. Otherwise its just pack up and go home. Nvidia will run away to > 80% market share.
1. Sub par drivers. AMD have launched Vega with really poor drivers. Here is a solid example.
https://www.techpowerup.com/reviews/AMD/Radeon_RX_Vega_64/10.html
Primitive shaders which unlocks the improved geometry performance of Vega is not enabled in drivers. The features that are enabled such as Draw stream binning rasterizer are not well optimized. AMD needs to improve their drivers and extract the best possible performance out of Vega hardware.
2. HBM2 clocks - AMD's memory partners are struggling with HBM2 clocks/yields . Vega 10 was supposed to launch with 2 Gbps HBM2. Instead they have launched with 1.89 Gbps , which in fact was achieved through overvolting. By mid-2018 AMD should get its HBM2 vendors Samsung and Hynix to deliver 2 Gbps HBM2 stacks at good volume and healthy yields.
3. Process and clocks - Vega on 14LPP is a turd. The card draws too much power as it cannot hit its targetted clocks at reasonable voltage/power. AMD has a 14nm+ process (most likely GF 14HPP) for its 2018 products. That process needs to deliver good clocks at reasonable voltage/power.
If AMD can achieve all of these by mid-2018 they might be able to atleast try and compete with GV104. Otherwise its just pack up and go home. Nvidia will run away to > 80% market share.
Last edited:
If they get Vega to at least 1080Ti levels of both performance and power usage they will have a chance, as that should be about 2080 level of performance, and the Volta Ti will not launch for at least a year. So if vega hits 1080Ti levels that will give them something to hold off volta, and they can then focus on navi. Bottom line is they need to get something to market in the next 6 months that can sit up in the top end with the nvidia cards, or their market share is going to tank hard.
Primitive shaders are enabled, but it is not useful without the NGG path. This will improve the performance and decrease the power consumption.
HBM2 is fine. But the pseudo channel mode is currently missing.
But to 1080ti perf/power from where they are now? Needs much more than a refresh! Navi might do it.
The 2070 will match the 1080ti, 2080 a bit faster etc. It's how NV do it.
Honestly they've had no top card at all for six+months & Fiji only half counted brute that. Vega crashing won't hurt their market share overly.
Their bigger problem in market share terms is going to be Polaris.....
Sent from my XT1635-02 using Tapatalk
I think they have the mid range pretty well covered now with vega 56 and polaris. After volta launches and the 2050/2060/2070 is released vega 56 should compete very well with the 2060, obviously AMD will have to drop its price down to $200-300 to reflect its performance vs the 2060. And RX580 or small vega should do ok vs 2050 and under cards, for under $200 price point.
They need a top tier card not more mid range cards, mid range i actually think they have well covered.
fleshconsumed
Diamond Member
Rofl, how sad it is that we're already talking about Vega respin just 9 days after it launched...
Given its lackluster performance, and the close proximity to the Volta rollout, im not surprised its come up already.
Zlatan. You need to tell us more 🙂
The ngg path is it in game engines or in the driver?
When can we expect to see it in action?
If HBM2 is fine why are clocks not 2 Ghz and why is pseudo channel missing.
William Gaatjes
Lifer
That is very interesting what you mention about pseudo channel. 🙂
It divides the hbm2 stack into smaller 64bit wide memory channels.
As if there where 16 sort of independent 64 bit channels from a single hbm2 stack to gpu for a total of 32 sort of independent 64bit channels for vega.
That should help with concurrent access, i think.
http://www.anandtech.com/show/9969/jedec-publishes-hbm2-specification
This is also an interesting text from extreme tech.
https://www.extremetech.com/extreme...w-details-new-gpu-architecture-codename-vega0
edit:
This last sentence got me thinking, Wolfenstein colossus may be a game(maybe the first) taking advantage of vega much more than just using the fp16 capabilities alone.
Last edited:
They can't and won't meet 1080TI on both performance and power usage with a refresh. That mark won't be hit until Navi by which time Volta will be significantly ahead in every metric sans of course price to the consumer.
Here is a HBM2 deep dive.
http://monitorinsider.com/HBM.html
Vega is not able to achieve as high a % of real bandwidth as peak theoretical bandwidth.
https://techgage.com/article/a-look-at-amds-radeon-rx-vega-64-workstation-compute-performance/5/
Nvidia Titan Xp and Quadro P6000 real world bandiwdth reach 78% and 82% of peak theoretical bandwidth respectively while Vega Rx 64 is just 67% even below Rx W7100 which is at 71%.
http://monitorinsider.com/HBM.html
Vega is not able to achieve as high a % of real bandwidth as peak theoretical bandwidth.
https://techgage.com/article/a-look-at-amds-radeon-rx-vega-64-workstation-compute-performance/5/
Nvidia Titan Xp and Quadro P6000 real world bandiwdth reach 78% and 82% of peak theoretical bandwidth respectively while Vega Rx 64 is just 67% even below Rx W7100 which is at 71%.
Last edited:
After the launch reviews i would agree with this. However some of these undervolting youtube vids and web reviews show some promise but only if they work with a majority of the chips. If AMD can respin and then bin the chips well enough maybe they can eek out 1080Ti performance over the next 6 months at least for one high end SKU getting all the top binned chips and getting special attention to voltage and clocks.
They can't bin the chips well enough now with a mature process. I seriously doubt we will see Vega on a high performance process unless it's easy/cheap to port to from 14LPP.
Moving to a high performance process from GF will help with clocks and may help power consumption but the yields will more than likely be worse than now, so we're right back to the overv-olting nonsense to keep profit margin high.
If the HBM problems get sorted, and they are on a new process, and they bin high and apply appropriate voltages and have phenomenal cooling, Vega may match 1080ti on performance and possibly even come close on power consumption. That said, I don't see AMD going through the trouble given that 1080ti performance on AIB cards will be had for under $700 at that point.
AMD realizes Vega is stillborn. I'm sure they are working hard to correct that with Navi but I wouldn't expect anything sooner. This is the 480 launch all over again with hopes that the 500 series would use faster memory and have improved power consumption and look how that panned out. Any refresh will basically be a re-brand.
Last edited:
I have a feeling you are right, but i hope for AMD's sake you are wrong. AMD is going to be in a world of hurt if they have nothing to counter volta till navi.
I'm an AMD fan and have owned far more AMD products than Intel/NVIDIA. I want to see them do well and to return to a competitive stance against NVIDIA as much as anyone but that's not going to happen any time soon. Given that Navi will be a first gen product I have fairly low hopes for it given AMD's performance in the GPU space over the last several years. I fully expect the gap for power efficiency and pure performance to increase between NVIDIA and AMD and for that trend to continue until at least Navi's successor. NVIDIA is just a well oiled machine with a much larger R&D budget.
No, this doesn't work. Some people hate this idea but the markets you're talking about here, even outside notebooks which is a big percentage, really are quite TDP sensitive, as has been proven a few times before. That's very definitely true with xx50 style chips - the 580 just won't fly in that market at all - but also holds through to the xx60 tier to a fair extent.
I also rather doubt if they can really sell vega 56 down at xx60 pricing levels.
Its not like they have a choice, im sure they dont want to sell the 56 for Nvidia xx60 pricing but if the 2060 is as fast or faster than vega 56 they wont have a choice in the matter, they will have to sell it at whatever price its relative performance justifies or no one will buy it.
The TDP is for sure an issue. Not sure how they are going to tackle that one.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 25K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 24K
-
-
