Nvidia Ampere on 10nm: Still Good Enough to Defeat RDNA 2.0?

If Nvidia plays it safe with 10nm Samsung, would it be good enough to defeat “Navi 2x”? Maybe if the die size is big enough…
1) 0:00 New 10nm info reminds me of Maxwell
2) 3:57 Comparing Samsung “8nm” to “TSMC “7nm”
3) 7:44 Nvidia’s Giant Dies vs Better Nodes
4) 9:35 Q4 2020 is going to be fun…

Address Fan Mail to:
Moore's Law Is Dead,
PO Box 10468
Peoria, IL 61612

Broken Silicon [A PC Hardware Podcast] is available on ALL platforms.

All Music by SAHARA:

LINKS!!!!!!:

[Zorro]

The right way to think about "nm" and transistor density is to consider it like maximum possible definition. Basically manufacturer is just saying: "we can pack [x] Mtr in 1 mm² of die area if you ask nicely".

[ApocDevTeam]

I really dislike how absolutely everything is about to be dumped on us all at the same time. Now we have Ampere, Big Navi, Zen 3, X670, PS5, XboxX and Cyberpunk 2077 all being released in Q4? So basically for the entire year all we get is an Intel 10th-gen that literally nobody wants, and a coronavirus that's canceling all the events that I was looking forward to.

[paulschaaf]

I don't care who wins, I just want both of them to be afraid enough to go all out. I want them both to release cards that they can legitimately say, this is the best we can do. Not, this is just barely good enough to maximize profit

[ashleyleskov]

I feel like Nvidia isn't worried about AMD catching up to their level of performance to take the "performance crown, " no; they're worried that they'll have no acceptable excuse to be pricing their new Ampere lineup as high as they can to earn as much profits as possible.


((Edited for -- hopefully -- better wording.))

[carlosnazario]

Im here to support your content and your algorithm

[Ashachi]

Gotta love how AMD has come so far to the point we are asking if people can beat them!


5 years ago we would laugh and dismiss them entirely.

[Cunning_Jester]

I'm stoked for the upcoming competition, dope cards for dope prices.

[emp.splash]

Considering how tightly booked TSMC 7nm is for the rest of 2020, I just couldn't figure out how nVidia could possibly release anything in volume this year. So this rumor about Samsung 10nm makes a whole lot of sense to me.

[juzujuzu]

It would be fun to see AMD go nuts and produce this time a shit ton of 100 CU RDNA2 Navis with 7nm+ EUV. Then select the absolute best silicon, pair it with HMB2, have them watercooled and clocked as high as they can with 300W - 350W range. Then send them to reviewers and rake the benefits for having their own tier GPU for once. This Halo product would change the minds of Nvidia fans, and also the mind of those remaining Intel fans that think AMD just got lucky this time.


If they only win by a small factor, Nvidia fans are going to make excuses and many of those who buy 2080tis will still buy Nvidia, assuming they are better quality. After all, they have been buying Nvidia for so many years that they will rationalize excuses. But if the best AMD GPU is in their own league, and Zen 3 is the best gaming CPU, then they will buy both CPU and GPU from AMD.


I certainly don't want AMD to become monopoly, but Nvidia needs to feel the heat to really start developing. And same goes with Intel of course.

[FAHRENHEITDELIO]

Horizon Zero Dawn Complete Edition for PC is coming to Steam this summer...

[Knee-Lew]

777mm squared...
well, not gonna lie, I'm gonna question about their yield, even if the wafer is produced in a matured node.

[sulphurous]

I'm pretty sure that this partially due to AMD placing dibs on TSMC's node capacity. That and I can recall the two not having the best partnership, to say the least.

[BartoszBielecki]

I think what AMD needs to do is quite simple.

Step 1: release a most powerful GPU in the existence - take the performance crown from RTX Titan or RTX 2080Ti (don't know which one was better for gaming, even if Titan was never meant to be used for games).

Step 2: Gain the recognition of NOT just catching-up.

Step 3: Wait for NVidia to release Ampere that might dethrone them (most likely will with massive die, HBM2e, etc.).

Step 4: Present mid-range cards that will wreck Ampere performance/dolar.

[elr]

not sure what the 7 dislikes are about, but great analysis and makes perfect sense. Competition is always good.

I personally think nothing has changed on AMD's 7nm road map for RDNA. I watched the entire financial analyst day. All 3hrs and 30mins of it! During the Q&A, Lisa Su did answer the question around 7nm+ being removed from the road map and the new slides talking about 7nm. She replied that it was simply AMD aligning to TSMC's new nomenclature for the 7nm finfet. TSMC call all 3 nodes 7nm so AMD is now calling 7nm+ 7nm was pretty much what Lisa Su said.

People also should remember that Lisa Su talked about a Navi 10 refresh. This could literally be a way for AMD to replace cards like the RX 5600 and RX 5500 with better nodes with slightly better performance and power efficiency. This would not be the first time they have done this. I think the RX 590 was on a different node than the RX 580. 12nm vs 14nm if I remember correctly. This refresh could allow AMD to potentially add a few more shaders or CU, the new RDNA 2x CUs that have been optimised for ray tracing calculations. The refresh would also be in line with what David Wang said in 2018 and others re-iterated in 2019, that AMD would not roll out ray tracing till they could push it across their entire product stack. With the consoles now featuring ray tracing, it only makes sense for AMD to roll out lower tier desktop GPUs with the same features. These GPU will probably have more CUs than the consoles especially if BIG Navi is meant to have 80 CUs. There is plenty of room between 40 or 48 CUs and 80 CUs. RX 6600 XT suddenly getting the same performance and CU count as the RX 5700 XT and doing ray tracing with even just a small price increase would be good. And prices are going up. AMD seemed to suggest they are looking at an ~50% margin vs the 40-44% they currently have on RDNA 1.0 and Zen 2.

In addition, is CDNA! Which Lisa Su actually mentioned not only for data centres but also for cloud gaming. Meaning Stadia will likely at some point be getting CDNA GPUs probably on 7nm+ using the prime chips. So what AMD may do is similar to Zen 2 Ryzen 3950X and Threadripper, use the same core design for Big Navi and CDNA but give the best chips to the different SKUs of CDNA and shut off the defective CUs at the low end and pass those to BIG Navi. Exactly what they did with MI50 and Radeon VII.

[tech]

2:09 I owned a gtx 970. Performance was 780ti levels for 330$. Had to buy. ;)

[HAHA.GoodMeme]

didn't realize the density difference between 7nm and 7nm+ was that huge

[excitedbox]

Did you see the PDF I posted yesterday about the Pezy Computing chip? It lists the Nvidia 8192 Core chip that just turned up in Benchmarks and gives some more insight on what we can expect based on what is known about the Pezy Exascaler 3.0.

[scottrobinson]

I need to clarify your math on the theoretical density improvement vs reality.

A 60% density increase would cause a 37% area shrink for the 'same' chip.
If the density goes up by 1.6 but the total number of transistors stays the same (just smaller), then your area is 1/1.6 of its size on the bigger node, meaning a 37.5% area shrink.

Just pointing out that you said a 60% theoretical density increase "didn't even get a 50% area shrink"

[oglack]

Fuck I love your speculation content. Just seeing your mind tick over so many variables is pretty fascinating. I feel like I've learnt a lot in the time I've subbed to your channel.

[GSP-]

The node only helps AMD so much...they need the architecture to deliver...and they've been falling behind on that for years...but now that the console works is essentially done, AMD can finally deliver on the high end PC market.