WYP
News Guru
A challenge to Nvidia in the low-end and High-end with RDNA.
Read more about AMD's rumoured Navi 12 and Navi 14 GPUs.
Read more about AMD's rumoured Navi 12 and Navi 14 GPUs.
Last edited:
Details of AMD's Navi 14 and Navi 12 chips leak - Big Navi and Small Navi
- Thread starter WYP
- Start date
Yeah think that's a small error in that table. Also the leaked 3 GB ‘7340:CF’ model (5600 non-X?) not mentioned here but found here (https://compubench.com/device.jsp?benchmark=compu20&D=AMD+7340:CF&testgroup=info) would be 96-bit(Assuming the full small chip is 128-bit as claimed in the leaks and as would be logical really), likely just cut down rather than a separate design.
Last edited:
WYP
News Guru
Yeah. A little mixed up there. Fixed now.
TBH, Does the 5800XT really need 12GB of VRAM? I have never heard anyone say their 5700XT benefits from GDDR6 overclocking. Not many games use more than 8GB of VRAM either, even at 4K.
With 16Gbps GDDR6 memory being available, I'd guess that a 256-bit memory bus is fine for big Navi. It's enough for the RTX 2080 Super.
Last edited:
Depends where it lands performance-wise. I'd like the AMD card to have a 'one-up' on there competitor card though. And Radeon has been sticking 8gb on cards with the 390, so why not lol.
What I would say is that the 2080S is getting close to it's memory limit, which can be seen in certain games. Presumably it can only get worse in the future.
Still no point competing if they can't compete on price, the cost and size(And complexity of design) of a hefty 384-bit bus probably doesn't warrant enough of a gain on a 64CU model now to be a worthwhile addition, 5700XT's success shows they just need to get close in performance across most use cases with a reasonable undercut. Still there will come a time when they want a true successor to the VII(probably with the seemingly slightly more compute orientated RDNA2 cards presumably from late next year), where they will probably need to go beyond 256-bit GDDR6 to satisfy all their target markets.
Lets not forget that AMD combated the GTX 980 with the 512 bit 390x, and the 1080 with the HBM2 V64. They have a history of over competing on the memory bandwidth front.
It would be shame to have memory holding back big navi
WYP
News Guru
A lot of that was due to Nvidia using their bandwidth better. Navi has a lot of architectural improvements that are designed to reduce the need for extreme levels of memory bandwidth through lossless compression techniques and other methods. Basically, the stuff that Nvidia has been able to do for years.
AMD legitimately needed more bandwidth than Nvidia. With Navi, AMD has levelled the playing field somewhat.
That would be a lot of effort to go to to satisfy a small niche market. Adding 4GB of VRAM and adding 50% to the GPU's memory bus size would have huge implications on the cost of their GPUs. AMD will have considered this and made what they think is the right call.
Yeah that was GCN though where they needed it, whereas RDNAs biggest gains are in memory efficiency, and RDNA 1 is a fairly pure gaming arch, 5700XT already seems to be over-provisioned with 448GB/s from benchmarks so far, seems it can be underclocked on memory a fair bit without losses.
Last edited:
Yeah according to the RDNA whitepaper they claim 2.5x more bytes per FLOP for 5700XT over Vega64 (So effectively x2.5 the performance in memory bound workloads).
A big part of this is down to optimising instructions, caches, and using (improved) compression across more of the execution units.
Last edited:
WYP
News Guru
TBH, AMD's memory issues are fixed twofold by GDDR6 offering a lot more bandwidth and AMD's architectural changes. It didn't get much press, as it ultimately doesn't matter that much with GDDR6.
This is why AMD can afford to use the same memory bus on a larger RDNA graphics card. Much like how the RTX 2080 Super and RTX 2060 Super use the same memory bus.
RDNA offers a lot of small architectural changes that all add up to offer some huge performance gains. It's a big leap in practically every respect.
Dicehunter
Resident Newb
If AMD could bring something out, Even a year and a half later, At the same performance level of a 2080 Ti then I'd be interested, If not then my money will stay with me for the time being.
AngryGoldfish
Old N Gold
I wonder whether this is going to be totally crushed by Ampere, now that it's announced. AMD are starting to make good cards, but they're coming out so late that Nvidia are only another 6-9 months away from beating them again. I know that's the way it goes, topping each other at every turn, but it does make it hard to settle on what the right product for you is and make a decision.
The industry will always move ahead, best to just buy what's best for you at the moment imo, GPUs don't usually depreciate in resale value too much during their lifetimes anyway, it's a very nice issue to have I guess, that we're getting quick meaningful iterations again.
Given Ampere are 7nm EUV parts I think it it's quite likely they will leapfrog in perf/watt and top end perf unless Samsungs node has issues with high clocks or something like that, but it sounds like AMD will be releasing their next iteration of RDNA on 7nm+(Also EUV) in H2 next year so things won't stay still for long.
Given Ampere are 7nm EUV parts I think it it's quite likely they will leapfrog in perf/watt and top end perf unless Samsungs node has issues with high clocks or something like that, but it sounds like AMD will be releasing their next iteration of RDNA on 7nm+(Also EUV) in H2 next year so things won't stay still for long.
