[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[soresu]

You realize that I said, that CLOCK FREQUENCY got 10% uplift at the same power, versus Zen 1, despite TSMC claiming it would be 30%?

Did you really expect a 30% clock boost with a dramatically increased FPU/SIMD engine?

That theoretical gain would be 30% at iso design under ideal circumstances - Zen2 is not the same design as Zen1 by a long shot, do I really need to explain this?

 

[soresu]

It is just a show how good are some developers with optimizing the engine.

Agree to disagree, I have a vehement dislike of NPR stylisation - and NPR should be easier to optimise or design for in the first place, so it's hardly a big indication of talent on the part of the coders.

Unlike the people that made their pre-rendered cutscenes that is, now that is stellar work.

Similar performance between different levels has more to do with a similar degree of level optimisation and design complexity than the optimisation of the engine that they run on - ie it's down to the level designers and their targets/max for a given platform.

It won't matter how efficient your engine is if one map has 20000 lights with a big variance in highly detailed geometry, and another map has only 100 lights with some very basic geometry repeated/instanced often.

The performance delta between those 2 designs would be quite noticeable.

Add in a significant variance in shader complexity and counts and it would be even worse again.

 

[Glo.]

Did you really expect a 30% clock boost with a dramatically increased FPU/SIMD engine?

That theoretical gain would be 30% at iso design under ideal circumstances - Zen2 is not the same design as Zen1 by a long shot, do I really need to explain this?

Facepalm.Gif

Yes, go ahead and explain. I will have a lot of entertainment.

 

[soresu]

Dunno if it was mentioned earlier in the thread, but Samsung just announced they cracked 12 high stack TSV's, so the mythical 24 GB HBM2 stacks are now possible (if likely far off).

Link here.

 

[beginner99]

At the very least you have to acknowledge that their perf/mm2 has been on a downward slope since they moved away from VLIW, I'm truly a bit baffled as to the why of this to be honest.

I know that things like HMCC, primitive shaders and AI instructions have been added, but even so the area utilisation seems woeful for a 3+ year jump on a new process.

First you can't compare CU to CU of different designs and even less so with Navi since it has changed a lot.

But the main point you are missing is clock speed. Higher clocks simply need more space.

That is also why saying apple is better at CPU design than intel is a little stupid. Intel manages to clock pretty high with pretty small dies which isn't easy. Apple simply makes huge SOCs. Their effiency comes from the wide design, eg. large die size. Works well if your chips is only sold in >$700 phones. Not so well if you need to sell it in $300 craptops.

Here AMD wanted to up clocks and that simply means bigger die per same amount of CU even if the CUs are identical in other aspects.

 

[AtenRa]

I dont know about you but I dont get it how they got 1.6X better perf/watt over the RX480 with only 12% higher perf and 150W TDP.

 

[Glo.]

I dont know about you but I dont get it how they got 1.6X better perf/watt over the RX480 with only 12% higher perf and 150W TDP.

Maybe because they undersold their product?

 

[Veradun]

I dont know about you but I dont get it how they got 1.6X better perf/watt over the RX480 with only 12% higher perf and 150W TDP.

This slide make no sense. If I'm not missing something trivial I get an 1.16x uplift in perf/w going with the 30% and 12% stated there, not 1.6x

 

[maddie]

This slide make no sense. If I'm not missing something trivial I get an 1.16x uplift in perf/w going with the 30% and 12% stated there, not 1.6x

1.12/0.7 = 1.6

AtenRa is correct. TDP cannot be 150W for the rest to be true, unless the RX480 used 215W.

 

[Glo.]

1.12/0.7 = 1.6

AtenRa is correct. TDP cannot be 150W for the rest to be true, unless the RX480 used 215W.

Because it isnt true .

TDP for RX 480 was 150W. AMD claims that Navi 14 has 30% lower power. Which translates to 110W.

This GPU will be between 110 and 130W of power draw, depending on custom designs.

AMD is heavily underselling this product. Only they know why they do this.

 

[maddie]

Because it isnt true .

TDP for RX 480 was 150W. AMD claims that Navi 14 has 30% lower power. Which translates to 110W.

This GPU will be between 110 and 130W of power draw, depending on custom designs.

AMD is heavily underselling this product. Only they know why they do this.

If they are, then who exactly do they think they're fooling? Nvidia? Us? It's not like this is high level thinking stuff here.

 

[soresu]

1.12/0.7 = 1.6

AtenRa is correct. TDP cannot be 150W for the rest to be true, unless the RX480 used 215W.

Something weird going on there, Anandtech's article went back up to 150W after going down to 110W - makes me wonder if certain people are communicating in the same language when they talk about the specs.

 

[rainy]

TDP for RX 480 was 150W. AMD claims that Navi 14 has 30% lower power. Which translates to 110W.

To be exact it's 105W.
Quite obviously marketing department of RTG looks somehow even more confused after departure a few of their guys to Intel.

 

[lifeblood]

From Anandtech's 5500 launch review:

"...today is not a retail video card launch. In fact, AMD isn’t even announcing exactly when retail cards will show up, beyond a vague Q4."

So the RETAIL launch of 5500 will be sometime between now and 31 Dec. They could launch Dec 31 but it makes much more sense to do it before Black Friday so they get all those sales. Of course if the problem is insufficient chips from TSMC then they won't have much choice in the matter.

I wonder if they are supply constrained which is why this is only an OEM launch? TSMC 7nm is very popular so nobody (except maybe Apple) is getting all the chips they want. I bet AMD is prioritizing Ryzen first as it's doing so well, then 5700 chips as they are more expensive and therefore have a higher profit margin, then finally 5500 chips. Since the supply of 5500 chips is tight they fulfill their OEM obligation first, then whatever is left will go to retail.

That's all conjecture but it does fit the evidence. As it is nobody is even sure if there will be a 5500 XT. We only assume so as Nvidia is supposedly releasing a 1650 Ti and a 1660 Super, and they wouldn't do that without some reason.

 

[Mopetar]

I wonder if they are supply constrained which is why this is only an OEM launch?

That's almost assuredly true. The 16 core mainstream Ryzen CPU has been delayed so it's almost certainly a case of AMD not being able to keep up with demand. It does make a lot more sense for them to fulfill OEM contracts first and to focus on growing those. Brand loyalists will wait.

That's all conjecture but it does fit the evidence. As it is nobody is even sure if there will be a 5500 XT. We only assume so as Nvidia is supposedly releasing a 1650 Ti and a 1660 Super, and they wouldn't do that without some reason.

Even if there isn't, NVidia would have to assume that there is unless they're okay with potentially getting caught with their pants down. Unlike Intel, NVidia isn't going to underestimate AMD. Putting that aside, there's too much of a performance hole between the 5500 and 5700 for AMD not to have a card. If you look at it in terms of what such a card would cost, it's one of the biggest market segments and not having a product there would be foolish. I suspect that AMD doesn't want to announce anything until they've got supply issues under control and actually know when they can get the product to market in volume.

 

[Glo.]

I would be really happy if AMD would use Navi 10 dies, with 6 GB's of VRAM, 2048 CUs, but low clocked, so that GPU would fit Vega 56 performance profile perfectly, and offer very good efficiency like under 150W Total power draw.

 

[Yotsugi]

Even if there isn't, NVidia would have to assume that there is unless they're okay with potentially getting caught with their pants down

They're already sitting with their pants down, waiting for 7LPP to yield.

Unlike Intel, NVidia isn't going to underestimate AMD

They already did, both Intel and AMD alike.
Oh, and they've pissed off hyperscale too. Whoops.
Now let's just wait and witness nV magically evaporating from every laptop ever, top to bottom.

it's one of the biggest market segments and not having a product there would be foolish.

Hitting $300 pt is good if you're doing it in economical (i.e. not horrendous binning down) way.

I would be really happy if AMD would use Navi 10 dies, with 6 GB's of VRAM, 2048 CUs, but low clocked, so that GPU would fit Vega 56 performance profile perfectly, and offer very good efficiency like under 150W Total power draw.

Nuh-uh, who would ever underclock their GPUs outside of mobility SKUs?
Also 6GB (and 192b) sounds like another 5830-tier abomination.

 

[jpiniero]

They're already sitting with their pants down, waiting for 7LPP to yield.

At least the node is in mass production now (the Note 10 is using it partially). I think, especially given the Datacenter business slump you'll see Ampere Tesla sooner rather than later.

 

[Glo.]

Nuh-uh, who would ever underclock their GPUs outside of mobility SKUs?
Also 6GB (and 192b) sounds like another 5830-tier abomination.

Price/performance segmentation .

About VRAM capacity, of course, AMD can use low-clocked GDRR6, like 12 Gbps, on 256 bit bus, for 384 GB/s and call it a day.

 

[Yotsugi]

At least the node is in mass production now (the Note 10 is using it partially)

SS is (in)famous for eating the costs when things go wrong (see 7420).

I think, especially given the Datacenter business slump you'll see Ampere Tesla sooner rather than later.

It already slipped, can't be any worse than it is now.

Price/performance segmentation

That's not how it works.
You sell a given die for as high as you could.
So, more clocks, higher prices.
Yay.

About VRAM capacity, of course, AMD can use low-clocked GDRR6, like 12 Gbps, on 256 bit bus, for 384 GB/s and call it a day

DRAMs are DRAMs, they have fairly fixed pricing disregarding the bins.
Also, the AIBs buy the memory, not AMD.

 

[jpiniero]

SS is (in)famous for eating the costs when things go wrong (see 7420).

True, but I bet yields are much better than say Intel's 10 nm. And margins are so crazy high for Tesla that there's room for bad yields even with big dies. Some anyway.

 

[Yotsugi]

True, but I bet yields are much better than say Intel's 10 nm.

Not that hard to be better than the worst trainwreck in semis manufacturing history.

And margins are so crazy high for Tesla

Not with all the training/inferencing hardware currently on/coming to the market.
nV's gotta tread cautiously now.

 

[Mopetar]

They're already sitting with their pants down, waiting for 7LPP to yield.

It doesn't matter if they aren't on 7 nm yet when there's no one around to punish them for it. It's the same story with Intel where it wouldn't matter at all if they were stuck on yet another iteration of 14 nm if AMD hadn't come out with Zen.

Navi is a big step forward for AMD, but they still need to improve it further and hopefully now that they're doing well on the CPU side they'll have the resources to invest into their GPU division as well. It's good that AMD has a competitive product again, but keep in mind it's competing against an NVidia chip that's on an old node that spent a lot of transistor budget on what's kind of a gimmick when it could have invested those into even better performance.

If AMD were even more competitive, NVidia has the option of ditching RT for an even larger segment of their cards. They're probably hoping that it takes AMD long enough to get their new RDNA architecture polished that they won't have to do that, but it is an option for them if it really comes down to it.

 

[Yotsugi]

It doesn't matter if they aren't on 7 nm yet when there's no one around to punish them for it.

Good old OEM deals will punish them in ~6-8 months.
In 3-4 years nV will cease to exist in laptops period.

Navi is a big step forward for AMD, but they still need to improve it further and hopefully now that they're doing well on the CPU side they'll have the resources to invest into their GPU division as well

Oh noes, you don't understand (but neither does nV, so whatever).
Products in vacuum are worth nothing is this dreaded world of neat OEM partnerships and yadda yadda yadda.

NVidia has the option of ditching RT for an even larger segment of their cards

How would that stop x86 vendors from giving away their GPUs for free?
OEMs like free stuff more than anything.

 

[jpiniero]

If AMD were even more competitive, NVidia has the option of ditching RT

With the consoles having HW RT, there's no going back now. If anything I expect Ampere to double down on RT.