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

[crazzy.heartz]

Well its two step problem.
First, youre comparing clock problems of 14nm polaris cards to 7nm navi cards. Second, vega already bumped achievable clocks on the same node, third vega20 showed how much more it does just by the added process improvements.

In short you can a) expect clock boost going from gcn 4 to 6, b) you can expect clock boost from 7nm alone.I would say 1750mhz is rather conservative, vega 20 stock does 1800 while having 2 stacks of hbm2 more than vega 64 yet it consumes the same power(which means the chip alone consumes less).Also GDDR6 may consume less than GDDR5.

In other words, it seems pretty feasible, im not telling that it is doing that clock.But it should be able to be around that.
Also, i know that vega 20 draws close to 300W, but remember that this is going to have 3,2x less shaders , and thats a vega20, navi should be a bit more advanced.

Appreciate the info. I am trying not to expect a lot, so as to not feel disappointed in next few months. I just wish to purchase a chip that has the same uArch (Navi) as one that would be in PS5 / Next XBox. Otherwise, a 120 Watt 1660Ti or 165 Watt Sapphire Pulse 56 are always there

Why would 1.8 GHz 1280 GCN core chip be slower than GTX 1650, which would have lower clock speeds and less cores?

Lack of ROPs on AMDs 128bit cards. A key reason why lesser cards from Nvidia are able to perform perform better.. ( 16ROP RX560 vs 32ROP 1050/1050Ti )

Then again, they would just launch the 20CU card with a 6Pin (similar to bulk of 1650's that launched recently) and keep the 16CU variant for sub 75Watt models.

The thing's almost guaranteed to get better caches and waaaay better reg file.

Looking in "some" places, it appears that really is the case.

I hope there are more such confirmations in days to come. Am happy to end the day with some positive information on these chips.

 

[Yotsugi]

Looking in "some" places, it appears that really is the case.

Absolutely.
Do people think AMD people there at Markham, Austin or Orlando just, like, do nothing?

Lack of ROPs on AMDs 128bit cards

Not really.
Better SIMDs first!

Then again, they would just launch the 20CU card with a 6Pin (similar to bulk of 1650's that launched recently)

No, the full die would probably be mobile-only in the same vein 1650 is.

I hope there are more such confirmations in days to come

AMD Computex keynote is only ~21 days away depending on your timezone.

 

[NostaSeronx]

If it is Super-SIMD derived, but without the second VALU array per SIMD.

SIMD0/1 will have their own textures and L1 cache and SIMD2/3 will also have their own textures and L1 cache. Any amount of changes to the front-end can mean more RBEs.
(20 CUs of SIMD64 32-bit, or 40 CUs of SIMD64 16-bit *shrug*; https://rocm.github.io/GCN_Float16.html)

So, a 1280 ALU design could have a 128 TMU and 32 ROP allocation. However, HBM2e is a cheaper option at this point. Add with a more cheaper, R9 Nano/Vega Nano or even a single-slot low-profile PCB design.

128-bit @ 18 GHz(highest) => 288 GB/s at ~4x the power and significant area on PCB.
1024-bit @ 2.4 GHz(lowest) => 307.2 GB/s at ~1x the power and no area on PCB.

 

[IEC]

I suspect that there will be plenty of room to undervolt, yet again.

My Radeon VII has a stock voltage of 1084mV, but is able to go all the way down to 950mV and maintain the 1800MHz boost clock... the "auto undervolt" mode ranges anywhere from 975mV to 1024mV.

When undervolted the Radeon VII is a decent card, and quiet. At stock with stock voltage, it's noisy.

Extrapolating from Vega 20 to Navi... I think you'll see more of the same.

 

[Mockingbird]

Here's a classic clickbait.

First, release a video showing things too good to be true

Second, release a subsequent video that's all doom and gloom

Since you've covered every scenario, you couldn't possibly be wrong.

 

[Mockingbird]

His rumors now has Navi anywhere from worse than Vega to better than Turing.

Who would have thought that Navi would be somewhere in there? /sarcasm

 

[Yotsugi]

First, release a video showing things too good to be true

These weren't even THAT good, but whatever nets him his clicks.

 

[ozzy702]

I suspect that there will be plenty of room to undervolt, yet again.

My Radeon VII has a stock voltage of 1084mV, but is able to go all the way down to 950mV and maintain the 1800MHz boost clock... the "auto undervolt" mode ranges anywhere from 975mV to 1024mV.

When undervolted the Radeon VII is a decent card, and quiet. At stock with stock voltage, it's noisy.

Extrapolating from Vega 20 to Navi... I think you'll see more of the same.

Yup. This is AMD we're talking about, I suspect we'll see the same over-volting across the board approach to wring max performance out of as many chips as possible.

Hopefully we see a different approach and a new course charted with NAVI. We'll see.

 

[Head1985]

Absolutely.
Do people think AMD people there at Markham, Austin or Orlando just, like, do nothing?

Well last time they actually do something in the core of GCN was in 2013 when they doubled Shader engines on hawaii.So yes they dont do anything.Since hawaii they trying get cheap sulotions like primitive shaders which dont even work instead just increase Shader engines to 6x.With 6x Shader engines they will get instantly +50% geometry performance and they could use 96rops so +50% pixel fillrate aswell.Pretty much biggest botllenecks solved.
Another problem with GCN is bad delta color compression and it needs insane memory bandwidth.Last time they improved it was with polaris and it was still worse than maxwell at that time.
Rx480 have 256GB/s and its worse than 224GB/s GTX980

Since then Nvidia released pascal with +25% over maxwell and turing with another +25% over pascal with DCC.
Why they didint improved DCC with vega?So yes since 2013 they really didnt do anything atleast compared to NV.They are stuck in 2013 design and trying get cheap sulotions instead just change things in the core.

 

[Yotsugi]

Well last time they actually do something in the core of GCN was in 2013 when they doubled Shader engines on hawaii.

For one, the very concept of SEs was created in Sea Islands.
Also Tonga introduced DCC, then Polaris and Vega iterate on it.

With 6x Shader engines they will get instantly +50% geometry performance and they could use 96rops so +50% pixel fillrate aswell.Pretty much biggest botllenecks solved.

Games are rarely bound by small tris or pixelfill.
What they need is better SIMDs, which you're getting either way.

Since then Nvidia released pascal with +25% over maxwell and turing with another +25% over pascal with DCC.

Paslel barely improved DCC.
Most Turing changes (being client Volta) are SM-related.

 

[Dribble]

My bet is the same GCN, but with DDR6, 7nm, and some architecturally minor stuff - perhaps better color compression, hardware support for some new video standards, and latest HDMI/display port.
They'll be competitive with Nvidia as long as you pump more volts, have a little faster memory and don't mind a few missing features. AMD will also proclaim on power point a bunch of features that get much fanfare but in the end don't amount to much.

Bit depressing, but all I've said is AMD will do basically do the same sort of thing they have done every release for the last few years.

 

[ozzy702]

My bet is the same GCN, but with DDR6, 7nm, and some architecturally minor stuff - perhaps better color compression, hardware support for some new video standards, and latest HDMI/display port.
They'll be competitive with Nvidia as long as you pump more volts, have a little faster memory and don't mind a few missing features. AMD will also proclaim on power point a bunch of features that get much fanfare but in the end don't amount to much.

Bit depressing, but all I've said is AMD will do basically do the same sort of thing they have done every release for the last few years.

This is the safe bet, but I'm open to believe that AMD has changed course and actually makes meaningful gains on NVIDIA. Personally, I don't believe we'll see those gains with NAVI, but will down the road as AMD's cash influx and greater R&D play a part. We'll see, I hope I'm wrong and we'll know fairly soon.

 

[Glo.]

Lack of ROPs on AMDs 128bit cards. A key reason why lesser cards from Nvidia are able to perform perform better.. ( 16ROP RX560 vs 32ROP 1050/1050Ti )

Then again, they would just launch the 20CU card with a 6Pin (similar to bulk of 1650's that launched recently) and keep the 16CU variant for sub 75Watt models.

Has nothing to do with ROPs, but clock speeds of the GPUs. Me and my friend few years ago have tested how 768 GCN Polaris Chip behaves compared to equally clocked GTX 1050 Ti.

We bought MacBook Pro with Radeon Pro 455, cause he needed it for Video Production at the time, and we downclocked GTX 1050 Ti to 900 MHz, because thats all we could do. Radeon Pro's clock speed was 850 MHz(I think...). Both GPUs have had the same memory bandwidth(GTX 1050 Ti also have had GDDR5 speeds declocked).

The difference in best case scenario in games was 3 FPS more for GTX 1050 Ti in Overwatch. In AMD optimized games Polaris was faster. Average difference across 10 games tested was... 0.5 FPS.

Difference in performance has zero to do With ROPs, unless you have to be genuinely ROP bound(you are never ROP bound in today's games). Two things affect gaming performance most. Clock speeds and Geometry performance. And those two things are the sole reason why Nvidia GPUs APPEAR to be faster.

Also, if you want more proof for this. Check out Radeon Pro Vega 20 performance compared to GTX 1060 Max-Q.

Let me put it this way. My prediction about Navi GPUs:
Navi 14: 20 CU GPU, with 128 bit memory bus with GTX 1660 Ti performance levels, for up to 170$ and up to 75W of power.
Navi 10: 40 CU GPU, with 256 Bit memory bus GDDR6, RTX 2080, or higher performance level, for up to 300$ price tag, and 150W TDP.

 

[Yotsugi]

My bet is the same GCN, but with DDR6, 7nm, and some architecturally minor stuff - perhaps better color compression, hardware support for some new video standards, and latest HDMI/display port.

This is the safe bet

There's no safe bets anymore, they've broken the very fundamentals.
Look up inst latencies.

 

[ozzy702]

Let me put it this way. My prediction about Navi GPUs:
Navi 14: 20 CU GPU, with 128 bit memory bus with GTX 1660 Ti performance levels, for up to 170$ and up to 75W of power.
Navi 10: 40 CU GPU, with 256 Bit memory bus GDDR6, RTX 2080, or higher performance level, for up to 300$ price tag, and 150W TDP.

1660TI performance on 75W and 128 bit bus, seems too optimistic on both fronts. 128 bit seems too light given AMD's memory requirements. Even with a 192 bit bus there's a memory performance impact between the GDDR5 1660 and GDDR6 1660ti. Faster GDDR6 may be available by then but it will also be costly so any 128 bit card that shoots for 1660TI performance will likely be heavily bandwidth constrained.

1660ti performance @ 75W, I doubt we'll see that level of power efficiency even on 7nm but it's possible.


My guess based off of what you have listed:


Navi 14: 20 CU GPU, with 128 bit memory bus with GTX 1650Ti (uncut, GDDR6) -1660 (non-TI) performance levels, for up to $200 and up to 110W of power.

Navi 10: 40 CU GPU, with 256 Bit memory bus GDDR6, RTX 2060-2080 performance, $250-$450 price tag, and 175-225W. Obviously this will include cut down models to fill that 2060-2080 range, price/tdp, etc.

 

[tamz_msc]

There's no safe bets anymore, they've broken the very fundamentals.
Look up inst latencies.

You mean stuff like these?
https://forum.beyond3d.com/posts/2066881/
They speak absolutely nothing regarding the high-level capabilities of Navi.

Cannon Lake has the lowest DIV/IDIV latency out of any processor out there, which does zilch in extrapolating its high-level capabilities.

Instructions are added/removed/modified all the time in any ISA. It doesn't usually mean anything as far as the high-level picture is concerned.

 

[Yotsugi]

They speak absolutely nothing regarding the high-level capabilities of Navi.

That they do.

Cannon Lake has the lowest DIV/IDIV latency out of any processor out there, which does zilch in extrapolating its high-level capabilities.

Never apply CPUs to GPUs.

Instructions are added/removed/modified all the time in any ISA

I'm not talking the actual instructions, but their latencies.
Your current GCN1-5 iterations are build about a very specific instruction latency, literally how their SIMDs hide it.

 

[crazzy.heartz]

Has nothing to do with ROPs, but clock speeds of the GPUs.

Difference in performance has zero to do With ROPs, unless you have to be genuinely ROP bound(you are never ROP bound in today's games). Two things affect gaming performance most. Clock speeds and Geometry performance. And those two things are the sole reason why Nvidia GPUs APPEAR to be faster.

Also, if you want more proof for this. Check out Radeon Pro Vega 20 performance compared to GTX 1060 Max-Q.

Let me put it this way. My prediction about Navi GPUs:
Navi 14: 20 CU GPU, with 128 bit memory bus with GTX 1660 Ti performance levels, for up to 170$ and up to 75W of power.
Navi 10: 40 CU GPU, with 256 Bit memory bus GDDR6, RTX 2080, or higher performance level, for up to 300$ price tag, and 150W TDP.

Appreciate the info. That disparity in shader counts for competing AMD/Nvidia GPUs is rather extreme.

If your prediction for either of these cards is true, I'd purchase the same in a heartbeat. That would be dream come true for AMD.

Here's to hoping they have a RV770 moment with Navi.

 

[railven]

Hey Glo is back! Now we can have a really hype train going!

Here is hoping AMD bring it! NV needs a good kick. WTB Competition! Intel 2020! My RTX 2080 Ti is almost a year old, it's obsolete, need to upgrade!

 

[Glo.]

Hey Glo is back! Now we can have a really hype train going!

Here is hoping AMD bring it! NV needs a good kick. WTB Competition! Intel 2020! My RTX 2080 Ti is almost a year old, it's obsolete, need to upgrade!

We actually need competition. For the sake of sane prices, we actually need competition.

P.S. If you have RTX 2080 Ti, don't hold your breath for next gen AMD GPUs . They won't be that fast. At least first two dies.

 

[Ranulf]

His rumors now has Navi anywhere from worse than Vega to better than Turing.

Who would have thought that Navi would be somewhere in there? /sarcasm

To be fair, at the start of part2 he says it is hard to find good info on Navi but he's all over the place on this one. I hope for the best but with the GPU industry these days, I expect nothing but a big meh anymore. That and people endlessly touting how great performance per watt is as they overclock the heck out of their hardware in between talks about saving the earth and going green.

 

[Trumpstyle]

Navi 10 (mid-range) will have 40CU's with 11 Teraflops and performance a bit under Geforce 2060 is my guess.
Navi 11 will be a low-end card with 18-20CU's, why people confusing about Navi 14? that wasn't in the leaked roadmap?

 

[DrMrLordX]

Navi 10 (mid-range) will have 40CU's with 11 Teraflops and performance a bit under Geforce 2060 is my guess.

Interesting, the MS Anaconda console will supposedly have a 14 Teraflop version of Navi. Lockhart: 4 Teraflops.

 

[Glo.]

Navi 10 (mid-range) will have 40CU's with 11 Teraflops and performance a bit under Geforce 2060 is my guess.
Navi 11 will be a low-end card with 18-20CU's, why people confusing about Navi 14? that wasn't in the leaked roadmap?

Why would 2.1 GHz 2560 GCN core GPU(which exactly fits your description) be slower than GTX 1080/RTX 2070?

It makes absolutely zero sense.

 

[Glo.]

My guess based off of what you have listed:


Navi 14: 20 CU GPU, with 128 bit memory bus with GTX 1650Ti (uncut, GDDR6) -1660 (non-TI) performance levels, for up to $200 and up to 110W of power.

Navi 10: 40 CU GPU, with 256 Bit memory bus GDDR6, RTX 2060-2080 performance, $250-$450 price tag, and 175-225W. Obviously this will include cut down models to fill that 2060-2080 range, price/tdp, etc.

GTX 1660 Ti has 1.9 GHz Turbo Clock speeds at 1536 CUDA core, and 288 GB/s bandwidth.

Why 1280 GCN core chip, with, lets say, 1,8 - 2 GHz, 256 GB/s of GDDR6 memory bus, be not able to get close to GTX 1660 Ti performance?

Remember guys. It appears AMD may have changed a lot in the structure of SIMDs. Similar changes nvidia made with Maxwell. And in the first place - it allowed them to increase efficiency, because less cores could do more work, and at the same time, the change in the architecture allowed Nvidia to increase the clocks of the GPUs, becaused reduced the memory movement over the GPU.

Over previous two years there was avalanche of AMD patents related to changes in SIMD structure. Changes in Linux drivers related to Navi are pointing to the reality that they actually manifest in Navi.

So that is first shoe. Second one is the process, itself. If Navi targeted, per AdoredTV latest video, much higher targets for clock speeds than Vega, we may be looking at 2.2 GHz GPUs.

And Navi has to land in Console APUs. What this means: reasonable power consumption. If Navi targeted over 2.0 GHz clock speeds for desktop chips, 1.8 GHz, with massively lower power consumption is reasonable logic, for APU chip.

I can easily see that performance targets have been met, by Navi chips. I can also easily see that power targets have been missed, and the GPUs will run hotter, than initially targeted. I cannot see however small GPU, like 120 mm2 with 128 bit memory bus to consume around 100W of power.

I can easily see that I have missed my prediction on power, and it will be higher, than 75, and 150W's(like 90W and 180W).

P.S. Why I speculate so nonchalantly about it? Because I don't care how well it performs. I will buy any GPU with 75W, no 6 pin connector, and preferably - passive, that has performance bigger than... RX 570. Its time for new computer.