AMD Could Launch New Navi GPUs Soon

[Mephis]

Current consoles are powered by Jaguar cores. What developers have been able to squeeze out of that hardware is impressive, IMO. Any modern desktop CPU would run circles around the x86 cores found in Xbox and PS4. And we all know how nvidia has gotten developers in their camp—many PC games launch with an nvidia logo on the splash screen. I’m sure that isn’t because they are nvidia fanboys.

I agree completely. My point was that AMD having the hardware in the consoles hasn't given them any advantage over Nvidia yet, so there was no reason to think that having the next gen would give them an advantage when it comes to RTRT.

 

[efikkan]

I agree completely. My point was that AMD having the hardware in the consoles hasn't given them any advantage over Nvidia yet, so there was no reason to think that having the next gen would give them an advantage when it comes to RTRT.

AMD have surely gotten benefits from the console dominance, but most people misunderstand how it benefits them. When the code which is run is the same, the room for "optimizations", either intentional or unintentional, is very limited. So as long as their games use standard x86 instructions and Direct3D and similar APIs, the performance characteristics wouldn't be vastly different across different hardware. The differences that we see is mostly the result of resource "bottlenecks", rather than optimizations. There may of course be a measurable difference here and there, but I seriously doubt there are many cases where AMD gains >5% benefit from games being developed for their console hardware. With everything else being equal, this benefit might have mattered a tiny bit, but not as long as Nvidia have a vastly more efficient and scalable architecture.

There are other potential benefits from AMD too, hard to measure things like more useful feedback about hardware and drivers. But I don't see AMD exploiting the full potential here.

 

[medi01]

Turing already scales better than Navi, despite Navi having the advantage of a more advanced node

By which metric???
2070 Super is 13.6 billion transistors chip with die size of 545mm2 (12nm)
5700XT, which is quite close, is 10.3 billion transistors on 250mm2 (7nm)

Except that right now the major consoles run AMD APUs and pc games still run better on Nvidia hardware.

Games run pretty well on AMD GPUs, despite its market share being one third of NVidias, possibly even less.
It doesn't mean NV cannot bribe devs to do #tesselation, of course they can.

AMD hasn't been able to use their console dominance to push devs to their camp yet

The only reason Blizzard optimized for 6 threads was consoles.
Doom was pretty much focused on console performance first.

And console influence would only grow, as now both previous (150+ million units) and the next gen (20-30 million per year per console) will be there.

 

[bug]

I really wonder if we will see desktop SKUs of Ice Lake at all. The IPC gain is good, but the mobile SKUs seem to suggest there are issues getting the frequency up to 14+++ levels. Even the IGP isn’t seeing higher clocks, and it depends on faster RAM. Put mobile Ryzen on the same RAM as Ice Lake, and I don’t think it would be very close.

Don't get too hung up on frequencies. Remember P4 vs AthlonXP?
Though yes, desktop Ice Lake gets the crown for the weirdest launch story of all CPUs I remember.

By which metric???
2070 Super is 13.6 billion transistors chip with die size of 545mm2 (12nm)
5700XT, which is quite close, is 10.3 billion transistors on 250mm2 (7nm)

By the "read the graphs" metric: https://www.techpowerup.com/review/amd-radeon-rx-5700-xt/29.html
Turing offers almost the same perf/W all the way from 2060 to 2080 (including Supers); Navi takes a 12-15% nose dive in that department just going from 5700 to 5700XT.

 

[Xaled]

Don't get too hung up on frequencies. Remember P4 vs AthlonXP?
Though yes, desktop Ice Lake gets the crown for the weirdest launch story of all CPUs I remember.



By the "read the graphs" metric: https://www.techpowerup.com/review/amd-radeon-rx-5700-xt/29.html
Turing offers almost the same perf/W all the way from 2060 to 2080 (including Supers); Navi takes a 12-15% nose dive in that department just going from 5700 to 5700XT.

These results mean nothing because they are based on the Performance Summary and TPU's performance summary tables always give Nvidias cards higher results because small fps spikes are included too (other sites call less than %1 or something similar)
İf you can prove your claim with similar information from other sites then please share

 

[Darmok N Jalad]

Don't get too hung up on frequencies. Remember P4 vs AthlonXP?
Though yes, desktop Ice Lake gets the crown for the weirdest launch story of all CPUs I remember.



By the "read the graphs" metric: https://www.techpowerup.com/review/amd-radeon-rx-5700-xt/29.html
Turing offers almost the same perf/W all the way from 2060 to 2080 (including Supers); Navi takes a 12-15% nose dive in that department just going from 5700 to 5700XT.

The thing is, it’s fine if IPC goes up, but that ends up being a wash if your new chip surrenders the IPC gain on the frequency side. Ice Lake mobile caps out at 4.1 GHz. That is 600mhz less than the coffee lake i7-8569U.

 

[bug]

The thing is, it’s fine if IPC goes up, but that ends up being a wash if your new chip surrenders the IPC gain on the frequency side. Ice Lake mobile caps out at 4.1 GHz. That is 600mhz less than the coffee lake i7-8569U.

Let's put it this way: it's easier to fix an initial low frequency in time than it is to fix an initial low IPC.

These results mean nothing because they are based on the Performance Summary and TPU's performance summary tables always give Nvidias cards higher results because small fps spikes are included too (other sites call less than %1 or something similar)
İf you can prove your claim with similar information from other sites then please share

I would be happy to look up other graphs for you if you explain to me first how a few FPS spikes can mess a perf/W graph.

 

[FordGT90Concept]

By the "read the graphs" metric: https://www.techpowerup.com/review/amd-radeon-rx-5700-xt/29.html
Turing offers almost the same perf/W all the way from 2060 to 2080 (including Supers); Navi takes a 12-15% nose dive in that department just going from 5700 to 5700XT.

5700 is Navi's...normal. 5700 XT is effectively factory overclocked. Lower 5700 XT to 1465 MHz and it will have similar perf/watt. 5700 boost is literally 20 MHz above where 5700 XT starts (1605 MHz). In other words, 5700 XT is boosting all the time...with an even higher boost clock. 50th edition will boost beyond 2 GHz. The further a chip is pushed, the more power it is going to need to remain stable.

 

[lexluthermiester]

5700 is Navi's...normal. 5700 XT is effectively factory overclocked.

Incorrect. The differences between the 5700 and 5700XT are due to binning. The dies that ran at a minimum spec became 5700's, dies that ran best became 5700XT's. I predict that we will be seeing 5600/5600XT's as AMD's stock of dies that run fine at certain specs but not enough to be 5700's increases. This is business as usual.

 

[bug]

5700 is Navi's...normal. 5700 XT is effectively factory overclocked. Lower 5700 XT to 1465 MHz and it will have similar perf/watt. 5700 boost is literally 20 MHz above where 5700 XT starts (1605 MHz). In other words, 5700 XT is boosting all the time...with an even higher boost clock. 50th edition will boost beyond 2 GHz. The further a chip is pushed, the more power it is going to need to remain stable.

I'm talking about retail products, you're talking... not sure what about.
But you're essentially agreeing with me: 5700XT's extra 150MHz is enough to push Navi past its sweet spot. So it doesn't scale as well as Turing.

 

[efikkan]

5700XT's extra 150MHz is enough to push Navi past its sweet spot. So it doesn't scale as well as Turing.

You're right. Navi seems to have a very narrow "sweetspot", while Turing scales across a wide range, despite Turing being on an "inferior" node.

 

[bug]

You're right. Navi seems to have a very narrow "sweetspot", while Turing scales across a wide range, despite Turing being on an "inferior" node.

Navi will probably scale on a wider range, too. But that ranges is from 5700 down.
It's not that bad, the $200 and below segment is where it's at. It's just that AMD hasn't had a good top-tier product in ages and that erodes their image in the market. Not to mention, it leaves Nvidia free to charge $1,300 for 2080Ti

 

[efikkan]

Navi will probably scale on a wider range, too. But that ranges is from 5700 down.
It's not that bad, the $200 and below segment is where it's at.

I guess they can scale down…
TSMC's 7nm node should be better than their 16/12nm node.
But Navi starts to fall behind from ~1500 MHz and up, which is not a good sign. Nvidia is on an "inferior" node and scales better.

 

[FordGT90Concept]

I'm talking about retail products, you're talking... not sure what about.
But you're essentially agreeing with me: 5700XT's extra 150MHz is enough to push Navi past its sweet spot. So it doesn't scale as well as Turing.

Navi 10 has perf/watt parity with TU106. They overclocked/overvolted Navi 10 to push the binned chips to within striking distance of TU104 but at the cost of efficiency. NVIDIA has little in the way of competition so they have no reason to push their chips as AMD is doing. That is to say, NVIDIA has different chips for each market segment while AMD only has one.

 

[bug]

I guess they can scale down…
TSMC's 7nm node should be better than their 16/12nm node.
But Navi starts to fall behind from ~1500 MHz and up, which is not a good sign. Nvidia is on an "inferior" node and scales better.

Technically, AMD could build a bigger die, clock it at 1.3-1.5GHz and end up with a faster product. But as of now, that would be cost ineffective. That's why my money's on AMD going for 5600 and 5500 before going for the 5800 (probably sometime next year, when 7nm capacity won't cost as much).

 

[medi01]

Is there any reliable conclusion on what next "quarters" means?
Perhaps at Xmas, guaranteed late Q1 2020 or I am too optimistic?

You're right. Navi seems to have a very narrow "sweetspot", while Turing scales across a wide range, despite Turing being on an "inferior" node.

Thanks to NVs infinite greatness and all-encompassing wisdom of the Leather Man, amen.

 

[FordGT90Concept]

There won't be any unannounced Navi-based cards nor mobile Ryzens until at least 2020. In other words, they're working on stuff but haven't announced anything and aren't going to be announcing it until they're ready to do so.

 

[ZoneDymo]

"Perfectly fine" is a lot more inclusive than "excellent" and "outstanding piece of equipment". I don't have a decades worth of data in my head but the phrase "quiet cooling" as applied to AMD reference cards is not something I have often seen.

I just dislike loud cards because they clutter up the market. I really don't understand why that is so difficult to fathom, and also I don't understand why AMD has such a hard time designing properly good cooling. Sometimes it happens, but it seems to me they for quite a while now has used exactly the same design for many of their cards and all of those are loud.

Well nobody said anything about quiet, by perfectly find I mean, you put it in a system and it will work at stock speeds without overheating, aka it just works, and they have fitted that role.
But what Nvidia cards then have fit that bill according to you?
The 8800GTX? 8800GT? 9800GT? GTX280? GTX480? GTX580? you tell me which one of those had a stock cooler (all blowers) that were cooling so fantastic and did so while remaining noticebly more quiet then AMD cards of the time.

Also you dont dislike loud cards because they clutter up the market, that makes no sense, you hate them for being loud, not for there being a bunch of them.
I agree that the RX480 and some others had just simply cheaply done base cooling, I would imagine this was part of trying to sell the cards at that low low MSRP which sadly never came to pass.

Now with these new cards, is this not just something that happens every now and then? on BOTH sides?
A card gets released and is exclusively stock cooler for all manufactuers for a while, before they get/have the freedom to make some custom models?