Discussion in 'Graphics Cards' started by subbytna, Aug 20, 2017.
Sneaky, looks like they've been taking lessons from the 'competition'
1) The cache changes were already done in Vega, what AMD talked about seems to be increasing cache sizes and potentially ration but the cache architecture is not changed.
2) If there is anyone trolling here it is you. AMD's own driver team consider Navi to be a GCN architecture. IF AMD don;t continue to invest in new architectures just because their console parts use GCN will simply mean AMD get left even futher behind as new processors and technology are developed that enable fundamentally new architecture approaches.
AMD have never been honest with architecture naming (nor has Nvidia for that matetr).
Remember, AMD claimed that the 390 was a brand new architecture when it was just a 290 with a new bios.
Fiji was a whole new architecture according to AMD< when the main change was HBM support
Polaris was a new architecture, when it included tweaks for delta color compression and geometry throughput.
Vega was touted as no longer having GCN compute units and AMd marketed Vega as being NCU rather than GCN. Vega actually had a signigicant number of changes, just that the real-wrold performance changes were pretty limited.
Everyone seems to be at that now to be fair. Intel brands each of it's new generation cpus a new architecture when it's the same. Nvidia have done it for a long time too with every single generation since Kepler I think.
New architectures have significantly higher costs and higher risks, they also tend to need a new smaller fabrication process due to the extra transistors required, and since new nodes are fewer and furhter apart and at higher costs it is cheaper and less risky to modify existing architectures.
Kepler to Maxwell to Pascal all had significant architectural changes though, big differences in cache design moving from texture cache to a unified cache and then better cache hierarchies. Scheduling changed significantly, and core configurations changed to reflect increases in compute workload and improve scaling. Turing i also nhas some major changes but likely smaller in effectiveness overall than previous generations.
Some of the Turing changes are listed here, which overall shows bigger architectural changes than vega to Navi GOING BY WHAT AMD HAVE STATED SO FAR. When reviewers get more details we may see more significant changes but as it stands Navi is to Vega what Turing is to Pascal, except Pascal added 2 completely new types of processing cores and technology (even if they aren't used much now)
Nvidia seem to want to wait until the 7nm+ EUV node before making a big architectural change and I suspect AMD has some of the same constraints. the 7nm+ is supposed to be much bigger change than you expect from the name, potentially bigger than going form the 16/14nm down to the current 7nm. That bigger improvement in transistor density, power and clock speeds is liekly deemed sufficient for big architectural changes requiring billions new transistors .
None of them are fundimentally changing anything though. They're optimising the architecture much more than they are changing it and that doesn't make it a new architecture.
The changes are much more significant than from Vega to Navi though. they are indeed evolution but at a significant enough difference that it would be fair to call them new architectures.
But TBH, this is going to come down to what you personally define as a new architecture. If the changes form Kepler through to pascal were not new architectures then Navi is certainly not. If Navi is a new architecture then all Nvidia recent GPUs are new architectures, including Turing.
And none of that changes the fact that Navi is GCN. Amd's last change of true architecture was moving from terrascale to GCN.
Not trying to defend AMD with it, the changes from Vega to Navi from what I understand would be similar to Nvidia going from Kepler to Maxwell which again I don't feel are really a proper architecture change.
Honestly Maxwell to Pascal seemed more about the die shrink and new memory technologies than any real architectural changes from my point of view, though I do admit I never looked into the changes in depth.
Jim @ AdoredTV is going a bit nuts on Twitter lol. His debrief video should be entertaining when it's out.
Not sure why some of you are still thinking Navi is GCN, Lisa spells it out for everyone here..
She says we'll see GCN in new versions of Vega/high compute. RDNA is "from scratch design" like Zen was, and "for the next decade".
Really cant see GCN going for another 10 years.
I'm sure Raja Korduri said he worked on Navi from the ground up before he left unlike polaris and also Vega and expects Navi to do better.
But then again they say all sorts and kinda hard to prove other wise.
Apparently just like Pascal was brand new and so was Turing lol.
I really do hope Navi pulls it out of the bag for price / performance.
Nothing is new according to some people.
Everything is new according the other people.
The truth is somewhere in between.
Navi looks like a power hog on 7NM! RTG doh!
AMD Radeon RX 5700 Navi series feature 225W and 180W SKUs
Navi at the top end looks like it's a hungry beast, doesn't mean Navi across the stack will be hungry. We've seen time and again the midrange to low cards on XYZ AMD arch have been pretty damn efficient, but that trails off into joke-levels as you push the arch up and out of its comfort zone.
The terminology can be confusing but Navi is certainly a next-generation chip once all the various flavours are rolled out, it will have ray-tracing in hardware (as confirmed by Sony) little features like that will be various in Navi.
Is it a 'all new, next-generation card' ? no, it does not need to be, in the same way that the 7nm die shink for ARM processors gave it such a kick up in performance that a ground-up new architecture was not really required.
Don't personally care if it's hungry if it does something with the power.
From what AMD have said, Kepler to Maxwell was a far greater change. remember, Nvidia introduced a deferred tile based render back-end that massively increased efficacy and is what really let Nvidia march ahead. Similarly, Maxwell to pascal was far mroe than simply a die shrink, there was a lot of redesign of the cores in enable to scale to higher clock, new scheduler that maintained efficiency at higher core counts and this was the first architecture that fully separated the compute from gaming GPUs. There were a lot of changes with the gaming focused compute engine in order to increase performance in DX12. Remember, Nvidia and AMD both had a die drink but the relative performance difference between AMD and Nvidia grew massively because of fundamental changes to nvidia's architecture.
Vega to Navi reads much the same as Pascal to Turing, except minus the RTX and tensor cores. where AMD really gain is by leveraging the 7nm process that straight up gives them the 50% efficiency bump.
The tile based renderer is (as far as I know) what is giving Navi the performance boost it's gaining over Vega so it would be very similar to that change. There will be other smaller things I am sure, but I am pretty certain that's a big part of the Navi architectural changes.
225 is entirely expected. Remember, AMD said 10% faster than a 2070 in a game highly favored by AMD. Once you translate that to real-world benchmark you get something around Vega64 to Vega64 + 10%, which is a little behind a 2070. AMD also said there is a 50% efficiency increase compared to vega64, that mostly coming form the 7nm process. Vega64 uses around 300w gaming load, a 50% 50% better efficiency means vega64 performance at 200w. If Navi 5700 end up more like 10% faster than a Vega64 you get soemthign along the lines of 220watts.
If you factor in that AMD's marketing numbers will be optimistic then you get something more like Vega64 at the stated 225w selling at the rumours $500 + VAT. This compares to a 2070 with a TDP of 175w on a larger 16/12nm process.
The Tile based renderer was introduced in vega but was apparently not very effective. AMD likely made improvements there for sure.
Separate names with a comma.