News Nvidia 4Q22 Earnings

[Hitman928]

First, Nvidia's official financial calendar makes this their 4Q23 earnings, but I am calling it 4Q22 to match the calendar time frame to avoid confusion.

Highlights:

4Q Revenue of $6.05B (-20.8% Y/Y)

Fiscal-year revenue of $27.0 billion, flat from a year ago

NVIDIA’s outlook for the first quarter of fiscal 2024 is as follows:

• Revenue is expected to be $6.50 billion, plus or minus 2% vs. $6.32B consensus
• GAAP and non-GAAP gross margins are expected to be 64.1% and 66.5%, respectively, plus or minus 50 basis points.

Nvidia Non-GAAP EPS of $0.88 beats by $0.08, revenue of $6.05B beats by $30M (NASDAQ:NVDA)

Nvidia press release (NVDA): Q4 Non-GAAP EPS of $0.88 beats by $0.08.Revenue of $6.05B (-20.8% Y/Y) beats by $30M.Fourth-quarter revenue was $3.62 billion, up 11% from a year ago...

seekingalpha.com

More to follow.

 

[coercitiv]

They used too little silicon for competing with the >600mm^2 GA102 which in turn was also on a denser process. Yes, AMD tried to do an extremely area optimized architecture but even so, they used a smaller total area over two different processes and half the 5N area. So it's quite possible AMD did not aim to overcome the top Nvidia chip, but to undercut it in costs. Top end cards costing >1,5K$ or more aren't the top priority in the market, they are good for "halo" effect but the BIAS is so high that no halo would have pushed AMD sales.

AD103 is 379mm2 and offers comparable performance to Navi 31 @ 304mm2 + 225mm2 worth of MCDs. The 7900XTX is likely more expensive to make than the 4080.

 

[leoneazzurro]

AD103 is 379mm2 and offers comparable performance to Navi 31 @ 304mm2 + 225mm2 worth of MCDs. The 7900XTX is likely more expensive to make than the 4080.

That was not my point, my point was that AMD competing with Navi31 vs a way larger die from Nvidia was wishful thinking. About the 4080, at the moment it is correct, more for the memory subsystem than the silicon+packaging in my opinion.

 

[coercitiv]

That was not my point, my point was that AMD competing with Navi31 vs a way larger die from Nvidia was wishful thinking.

And my point was they used too much silicon to be price competitive with 4080, which implies they were aiming for better performance. Your logic isn't providing us with converging results when altering the comparison point from AD102 to AD103.

 

[Stuka87]

AD103 is 379mm2 and offers comparable performance to Navi 31 @ 304mm2 + 225mm2 worth of MCDs. The 7900XTX is likely more expensive to make than the 4080.

The MCDs are relatively cheap to produce. You also have to take into account that AMD was able to use a cheaper memory subsystem because of it. Sure you can single out the price of one component, but its the cost of the combined system that matters. And there is most certainly a cost difference between GDDR6 and GDDR6X.

 

[insertcarehere]

AD103 is 379mm2 and offers comparable performance to Navi 31 @ 304mm2 + 225mm2 worth of MCDs. The 7900XTX is likely more expensive to make than the 4080.

The 4080 isn't a full AD103 die so Nvidia can make do with binned dies for that SKU as well while AMD needs the full N31 die for 7900XTX.

 

[leoneazzurro]

And my point was they used too much silicon to be price competitive with 4080, which implies they were aiming for better performance. Your logic isn't providing us with converging results when altering the comparison point from AD102 to AD103.

My logic is that even if they aimed at higher performance, the amount of silicon used is way too low for providing a direct competition to AD102 even if they met a 3GHz target (there are several tests of 7900XTX boards overclocked just short of 3GHz around proving just that). And AD102 was the original topic, AD103 is something you inserted in the discussion without any relation to the original topic of discussion (that is, Navi 31 was aimed at direct competition with AD102 or not). Does Navi31 compete poorly with AD103? Maybe (and I am not saying the opposite), there are a lot of factors involved there, for sure it isn't from the customer point of view but in any case from the manufacturing costs point of view N31 is quite probably more similar to an AD103 than an AD102.

 

[coercitiv]

AD102 was the original topic, AD103 is something you inserted in the discussion without any relation to the original topic of discussion

The original topic was whether RDNA3 stumbled out the gate or not.

 

[leoneazzurro]

The original topic was whether RDNA3 stumbled out the gate or not.

And I was responding to this

News - Nvidia 4Q22 Earnings

First, Nvidia's official financial calendar makes this their 4Q23 earnings, but I am calling it 4Q22 to match the calendar time frame to avoid confusion. Highlights: 4Q Revenue of $6.05B (-20.8% Y/Y) Fiscal-year revenue of $27.0 billion, flat from a year ago NVIDIA’s outlook for the first...

forums.anandtech.com

Which is a specific question that does not contain your argument. Now, are you tired of trying to discuss uselessly or ?

 

[Mopetar]

It is as much of a professional product as Navi31 or GA103 are, all of these will lead to "Pro" CAD SKUs and GA102 shares the architecture with the GA103/104/106/107. I don't know of any "pro" feature that is GA102 only. And yes, AMD probably missed their clock targets but even if they met them, it is unlikely they could have matched the GA102 with just more than half the equivalent 5N area...

A 4090 is 10%/17%/22% faster than a 7900 XTX in 1080/1440/4K gaming based on the suite of games tested by TPU.

Even a 10% change in clock speed changes the relative value quite considerably. 15% and we're close enough where the argument is about 4K entirely and even then likely a matter of individual titles and how much RT is worth to you.

But even if AMD had 15% better clocks they don't have the wafer capacity to gain much market share. Instead the 7900 XTX would have been priced $200 higher and much closer to the 4090.

If AMD can find out how to solve the problems around using multiple compute dies with a GPU such that it doesn't incur the usual performance hits or problems that traditionally plagued SLI/X-fire setups then they could probably compete heavily on price even if they're down 10% in absolute performance.

Maybe that's not something that ever happens with gaming GPUs, but it does seem more realistic for CDNA products in the future.

 

[leoneazzurro]

A 4090 is 10%/17%/22% faster than a 7900 XTX in 1080/1440/4K gaming based on the suite of games tested by TPU.

Even a 10% change in clock speed changes the relative value quite considerably. 15% and we're close enough where the argument is about 4K entirely and even then likely a matter of individual titles and how much RT is worth to you.

But even if AMD had 15% better clocks they don't have the wafer capacity to gain much market share. Instead the 7900 XTX would have been priced $200 higher and much closer to the 4090.

If AMD can find out how to solve the problems around using multiple compute dies with a GPU such that it doesn't incur the usual performance hits or problems that traditionally plagued SLI/X-fire setups then they could probably compete heavily on price even if they're down 10% in absolute performance.

Maybe that's not something that ever happens with gaming GPUs, but it does seem more realistic for CDNA products in the future.

This is TPU review of the 7900XTX Merc with a 3165Mhz overclock.
It is not enough to reach the 4090, in rasterization.

XFX Radeon RX 7900 XTX Merc 310 OC Review - 3x 8-pin FTW

With the Speedster Merc 310, XFX has engineered an excellent custom-design Radeon RX 7900 XTX. Thanks to an upgraded cooler, the card can easily handle the additional heat output from the three 8-pin power inputs. The additional power is put to good use, to extend the performance lead against...

www.techpowerup.com

Now yes, if AMD uses more dies with fast enough interconnects, then it would be faster.
But this means also use more silicon, that is, the premise is not changed.
What chiplets change is economy of scale. Instead of using a single large die, and having several different dies for every market segment, you'll have smaller dies (higher yield, higher wafer area utilization and thus lower cost per die) which can be used in scalable SKUs, but the equation more silicon=more performance in the end does not change.

 

[maddie]

A 4090 is 10%/17%/22% faster than a 7900 XTX in 1080/1440/4K gaming based on the suite of games tested by TPU.

Even a 10% change in clock speed changes the relative value quite considerably. 15% and we're close enough where the argument is about 4K entirely and even then likely a matter of individual titles and how much RT is worth to you.

But even if AMD had 15% better clocks they don't have the wafer capacity to gain much market share. Instead the 7900 XTX would have been priced $200 higher and much closer to the 4090.

If AMD can find out how to solve the problems around using multiple compute dies with a GPU such that it doesn't incur the usual performance hits or problems that traditionally plagued SLI/X-fire setups then they could probably compete heavily on price even if they're down 10% in absolute performance.

Maybe that's not something that ever happens with gaming GPUs, but it does seem more realistic for CDNA products in the future.

Why is this wafer capacity fable refusing to die?

Even if there were not numerous reports telling us about wafer order reductions, including TSMC refusing to cancel orders by Nvidia and settling for a time extension instead, we still would have no shortage of wafers. Both Nvidia and AMD use TSMC. If Nvidia loses a sale to AMD or the reverse happens, TSMC still has to produce the chip. The various chip sizes only slightly change the total wafers needed, in fact, because AMD uses both 6nm and 5nm for their products, it would be much easier to fab if AMD had 100% of the market.

 

[Mopetar]

Conditions now are not what they were when AMD would have been allocating wafers for RDNA3 production. Doing a respin on N32 has delayed production for months. If N31 had launched in a state where it was competitive with the 4090, AMD would probably buy more wafers if other companies were cutting production. Why bother ramping up production after the fact when you know you fell short?

 

[coercitiv]

Conditions now are not what they were when AMD would have been allocating wafers for RDNA3 production. Doing a respin on N32 has delayed production for months. If N31 had launched in a state where it was competitive with the 4090, AMD would probably buy more wafers if other companies were cutting production. Why bother ramping up production after the fact when you know you fell short?

Your own posts referenced a scenario where N31 launched in a better state, with presumably 15% better clocks under gaming loads. For this scenario you argued that AMD would not have the capacity to gain market share. @maddie responed to this with the following:

• TSMC has capacity to spare due to current market conditions
• TSMC is currently the only one feeding the high-end GPU market

This means that given a highly competitive product stack (close enough in perf, considerably cheaper to make and thus a winner in case of price war), AMD could gain as much market share as the the market would let them. When the market wants to buy one brand, the other brand is forced to lower production, and since there is only one supplier, the winner gets more capacity.

For what is worth, I think your original argument was spot on: AMD has little room to maneuver today after RDNA3 stumbled at launch. IF there's indeed a "bug" and N32 is getting a respin, then IMHO N31 is getting a respin as well and we'll see another dice roll later this year. On the other hand, if this is all that RDNA3 can offer, then we won't see much market share movement as Nvidia will have control over sales by leveraging RT and tunning price points relative to mind share.