Discussion Nvidia Blackwell in Q1-2025

[adroc_thurston]

Minor node update, power consumption going up roughly proportional to performance gains, seems like they got roughly what they could out of the node given the power limit.

Frequency isn't always the function of the dieshrink. See RDNA2.

 

[jpiniero]

I had to, I'm kinda' almost sorry.

Unironically, I think MFG was the reverse version of this meme.

 

[GodisanAtheist]

This is why AMD should always go Halo (hell, paper launch a halo just for the mindshare). At some point the competition trips up really badly and you recover off a bad gen. That's where the wins happen. AMD kept swinging for the fences with their ZEN Cpus and eventually Intel tripped and never really got back up again.

Unfortunately, here, AMD didn't even show up in the big die space Halo space and they've inadvertently given Nvidia a generation to regroup.

Just the old "you miss 100% of the shots you don't take" mentality. AMD looks to have missed this one to recapture some mindshare.

 

[PJVol]

At some point the competition trips up really badly and you recover off a bad gen.

NV PR machine will wipe that "Halo" part from people's minds with just a new fancy bar that is 3x higher than whatever your Halo can go (of course Nx fake frames, dlss-neuro-AI, etc are involved). There's nothing you can do about it in short term, or within one product cycle. You have to win unconditionally.

 

[Keller_TT]

This is why AMD should always go Halo (hell, paper launch a halo just for the mindshare). At some point the competition trips up really badly and you recover off a bad gen. That's where the wins happen. AMD kept swinging for the fences with their ZEN Cpus and eventually Intel tripped and never really got back up again.

Unfortunately, here, AMD didn't even show up in the big die space Halo space and they've inadvertently given Nvidia a generation to regroup.

Just the old "you miss 100% of the shots you don't take" mentality. AMD looks to have missed this one to recapture some mindshare.

That might not be all. Let's see if AMD go for the jugular to take market share with the 30 series and RDNA2 upgraders or suddenly salivate at the chance of doing NV-50/100 strategy.
This is the best chance in a decade for doing a huge market correction if AMD keep true to their purpose and don't go by NV pricing. Just wow the market, up the brand value, and up the Radeon group morale to set the tone for RDNA5.

 

[Saylick]

Frequency isn't always the function of the dieshrink. See RDNA2.

Nvidia's architectures are optimized pretty well already. Not going to get a miraculous 1.5 perf/W boost from Ada without bigger architectural changes under the hood, which don't exist.

Recall that a good chunk of RDNA 2's perf/W gains came from the inclusion of Infinity Cache. Nvidia already got that benefit from Ada.

Without a perf/W boost, only way to get more performance is to throw more silicon at it, assuming clocks are held constant, which they roughly are.

 

[adroc_thurston]

Nvidia's architectures are optimized pretty well already

sort of.

Not going to get a miraculous 1.5 perf/W boost from Ada without bigger architectural changes under the hood, which don't exist.

yeah that's like THE problem. It's a big tock.

Recall that a good chunk of RDNA 2's perf/W gains came from the inclusion of Infinity Cache

A good chunk of RDNA2 perf/W gains came from it huffing crack and paint thinner at once.

Without a perf/W boost, only way to get more performance is to throw more silicon at it, assuming clocks are held constant, which they roughly are.

it's a proper tock, new new shiny SM and this is the end result.

 

[Hitman928]

I am disappoint.

 

[Heartbreaker]

This is why AMD should always go Halo (hell, paper launch a halo just for the mindshare). At some point the competition trips up really badly and you recover off a bad gen. That's where the wins happen. AMD kept swinging for the fences with their ZEN Cpus and eventually Intel tripped and never really got back up again.

Unfortunately, here, AMD didn't even show up in the big die space Halo space and they've inadvertently given Nvidia a generation to regroup.

Just the old "you miss 100% of the shots you don't take" mentality. AMD looks to have missed this one to recapture some mindshare.

I Don't agree at all.

These shots you aren't taking, are NOT free. It costs mega bucks just in up front costs to bring extremely large complex dies to market.

Let's flush 80 Million down the drain hoping NVidia shoots itself in the foot, is not a good strategy.

As you move up the stack, the business case (making all your money back on up front costs) gets weaker and weaker.

At the 5090 level you probably need NVidia volumes and NVidia pricing. Even if AMD beat 5090, they would most likely get both a lower price and a lower volume, and could still have a failed business case.

While it's a little early to start using hindsight on this generation, I'm betting that NVidia still ends up with better performance/area this generation, meaning even this generation, AMD would have still failed.

NVidia's lack of progress this generation was an unwillingness to build larger dies on what is essentially the same process. Except on the lunatic size 5090. There is really no room left for AMD to significantly outsize the 5090 die either.

The time AMD might consider making a run at the 090, is ONLY ifAMD make some kind of big performance boosting breakthrough that the think NVidia is not getting. Then it would be worth a shot, because just trying to out-lunatic NVidia on maximum die size, seems like a costly recipe for failure.

 

[adroc_thurston]

These shots you aren't taking, are NOT free. It costs mega bucks just in up front costs to bring extremely large complex dies to market.

Let's flush 80 Million down the drain hoping NVidia shoots itself in the foot, is not a good strategy.

As you move up the stack, the business case (making all your money back on up front costs) gets weaker and weaker.

At the 5090 level you probably need NVidia volumes and NVidia pricing. Even if AMD beat 5090, they would most likely get both a lower price and a lower volume, and could still have a failed business case.

Halo effect is a business case enough, but they were afraid they're not gonna win.

 

[PJVol]

but they were afraid they're not gonna win.

And quite rightly so, they wouldn't.

 

[adroc_thurston]

And quite rightly so.

?

 

[GodisanAtheist]

I Don't agree at all.

These shots you aren't taking, are NOT free. It costs mega bucks just in up front costs to bring extremely large complex dies to market.

Let's flush 80 Million down the drain hoping NVidia shoots itself in the foot, is not a good strategy.

As you move up the stack, the business case (making all your money back on up front costs) gets weaker and weaker.

At the 5090 level you probably need NVidia volumes and NVidia pricing. Even if AMD beat 5090, they would most likely get both a lower price and a lower volume, and could still have a failed business case.

While it's a little early to start using hindsight on this generation, I'm betting that NVidia still ends up with better performance/area this generation, meaning even this generation, AMD would have still failed.

NVidia's lack of progress this generation was an unwillingness to build larger dies on what is essentially the same process. Except on the lunatic size 5090. There is really no room left for AMD to significantly outsize the 5090 die either.

The time AMD might consider making a run at the 090, is ONLY ifAMD make some kind of big performance boosting breakthrough that the think NVidia is not getting. Then it would be worth a shot, because just trying to out-lunatic NVidia on maximum die size, seems like a costly recipe for failure.

- I am curious, if you're already making GPUs, how much cost actually goes into making the individual parts along the entire line-up.

Would it have cost AMD 80 million to cook up a ~550mm2 die to take on Nvidia's 750mm2 behemoth? Isn't the vast majority of the cost arch R&D, which is going to happen anyway? What are the opportunity costs of brain drain from the department, low morale, doctrinal decay in making moonshot GPUs, etc?

Those are honest questions I don't actually know the answers to, not being a dick.

If a lot of the cost is tied up in the actual production of the chips/wafers/physical stuff... just don't make that many of them, just get enough out there to force the competition to lower their pricing, spend their own money. Business is war, after all.

But if a 64CU RDNA4 die can (potentially) challenge a 4080/5070Ti/5080, I don't see why a 96/128CU RDNA4 die couldn't make it most of the way to or past a 5090.

Maybe we haven't heard the last of AMD this gen, maybe they still have enough oomph for a Hawaii class chip as a mid-gen refresh.

 

[PJVol]

?

Their concerns have obvious grounds

 

[gdansk]

~550mm2 die to take on Nvidia's 750mm2 behemoth?

Wouldn't be big enough. A 384 bit part would end up closer to 5080 again. They'd need a part at least 750mm² themselves. Which makes it pretty obvious why they wouldn't build it. Radeon never? builds dies that big. 600mm² is their preferred limit. So they'd have to justify it and explain why it could work this time.

 

[Heartbreaker]

Would it have cost AMD 80 million to cook up a ~550mm2 die to take on Nvidia's 750mm2 behemoth?

I don't know have much of up fronts costs are shared costs, and how much are per part. I know they layers of masks are very expensive and they are individual. There will be tons of validation/testing work that will be individual.

How would a 550mm2 AMD die take on a 750mm2 Nvidia die unless AMD has a ridiculous advantage, that is clearly NOT in evidence.

I did say, the time to try to take on Halo, is if AMD makes that kind of big breakthrough, so sure if that happen, but there is ZEO sign of that.

To take on a 5090 with the tech as revealed, AMD would probably need to build bigger than 5090... Which is a complete non starter.

 

[Golgatha]

Too big, too pricey.

You misspelled maintain the 75%+ margins on our AI chips.

 

[GodisanAtheist]

Wouldn't be big enough. A 384 bit part would end up closer to 5080 again. They'd need a part at least 750mm² themselves. Which makes it pretty obvious why they wouldn't build it. Radeon never? builds dies that big. 600mm² is their preferred limit. So they'd have to justify it and explain why it could work this time.

I don't know have much of up fronts costs are shared costs, and how much are per part. I know they layers of masks are very expensive and they are individual. There will be tons of validation/testing work that will be individual.

How would a 550mm2 AMD die take on a 750mm2 Nvidia die unless AMD has a ridiculous advantage, that is clearly NOT in evidence.

I did say, the time to try to take on Halo, is if AMD makes that kind of big breakthrough, so sure if that happen, but there is ZEO sign of that.

To take on a 5090 with the tech as revealed, AMD would probably need to build bigger than 5090... Which is a complete non starter.

- NV is throwing a lot of stuff at the 5090 that is clearly intended for AI workloads that is bloating the die size and jacking up the price. Massive 512bit bus + adding GDDR7? Adding in int ops to their dual pumped SPs? Lower precision ops on their tensor cores?

I figure NV will be slamming into a CPU bottleneck even at the 4K mark as far as gaming is concerned, so this over-engineered monstrosity (for gaming anyway) isn't even going to get to stretch its legs until a couple years and all the sales go by.

AMD could lean that up quite a bit methinks and get within striking distance. Nvidia has even left an EXPANSE for AMD to price into between the 5080 and 5090. This isn't AMD selling their rinkadink R770's for $300 bucks a pop anymore. Price the AMD Halo at $1200-$1500 and NV is left in an awkward position where a cut down 5080ti isn't left a lot of room to maneuver between pricing and performance while AMD still makes a reasonable margin on their product.

 

[Keller_TT]

Wouldn't be big enough. A 384 bit part would end up closer to 5080 again. They'd need a part at least 750mm² themselves.

I'm just not made for this sitting near the Swiss Alps. Tears of agony. 575W GPU for gaming at home? All I can think of is my recent visit to the Alpine glaciers. Especially Morteratsch and Mer de Glace (Sea of Ice). Moteratsch has lost one Km of ice cover just since 2014. Rock and mud.
Mer de Glace has disturbing range markers in the descent to show how far the glacier has fallen over the years (over 200 feet since the 90s).

57.5W is where it needs to get to for 4090 performance in 10 years for home, and this GenAI bollocks needs to end. My ex-colleagues at the Paul Scherrer Institut really got short of breath when we were discussing about 0.5kW NV gaming cards.

 

[Saylick]

Wouldn't be big enough. A 384 bit part would end up closer to 5080 again. They'd need a part at least 750mm² themselves. Which makes it pretty obvious why they wouldn't build it. Radeon never? builds dies that big. 600mm² is their preferred limit. So they'd have to justify it and explain why it could work this time.

I think taking N48 and scaling up by 1.5x could take on the 5080 and win by a small margin. If the rumors are true in that N48 gets you 95% of a 4080 in raster, scaling that up by 1.4x already puts you at 4090 levels of performance and VRAM, which itself is ~10% faster than the 5080. RT performance will be a little bit behind, but for $999 it would've been competitive if FSR4 is good enough, I think.

 

[gdansk]

I think taking N48 and scaling up by 1.5x could take on the 5080 and win by a small margin. If the rumors are true in that N48 gets you 95% of a 4080 in raster, scaling that up by 1.4x already puts you at 4090 levels of performance and VRAM, which itself is ~10% faster than the 5080. RT performance will be a little bit behind, but for $999 it would've been competitive if FSR4 is good enough, I think.

It would not achieve halo status nor compete with the 5090. If AMD had made such a part the 5090 would be cut less and there would be a 5080 Ti cut from GB202 to fight with it.

 

[Heartbreaker]

- NV is throwing a lot of stuff at the 5090 that is clearly intended for AI workloads that is bloating the die size and jacking up the price. Massive 512bit bus + adding GDDR7? Adding in int ops to their dual pumped SPs? Lower precision ops on their tensor cores?

The evidence doesn't really seem to bear that out though.

In lower end 50 series parts Die size is about the same as 40 series, for about the same Shader count as last generation, and 5090 expanded proportionally to the SM count increase, so NVidia does not seem to have any die size bloat from the technology update.

Plus this generation AMD is supposed to be adding real AI tensor cores, and beefing up it's RT capability. So AMD may actually have bloat while NVidia is showing none.

So again, no evidence that AMD has any advantage to exploit, so would need a die just as monstrous as 5090's to compete... which I still argue is a complete non starter.

 

[Mopetar]

What an exciting time to be alive. AMD punts and NVidia kicks it right back to them. I think I'm about ready to buy an Intel GPU out of spite.

 

[tajoh111]

This is why AMD should always go Halo (hell, paper launch a halo just for the mindshare). At some point the competition trips up really badly and you recover off a bad gen. That's where the wins happen. AMD kept swinging for the fences with their ZEN Cpus and eventually Intel tripped and never really got back up again.

Unfortunately, here, AMD didn't even show up in the big die space Halo space and they've inadvertently given Nvidia a generation to regroup.

Just the old "you miss 100% of the shots you don't take" mentality. AMD looks to have missed this one to recapture some mindshare.

Zen is way different. Zen does not use a huge complex monolith for their consumer line and is cheap to build. AMD just been delivering at a consistence cadence for their CPU division while keeping power consumption relatively low which is what help it gain consumer confidence. It was not the threadrippers that did it.

Going Halo for Nvidia makes more sense since they have a professional visualization market which provides guaranteed income at high margins.

AMD doesn't have much for this market. As we can see with the possible performance flop of the RTX 5090, going big carries more risk as more things can go wrong. The GTX 480 and Fury X are examples of this. The RTX 5090 is clocked awfully low for the amount of power it uses. This might be a fermi like chip which misses targets in the gaming segments but recovers sales in the professional markets.

Also chip design is more expensive than that. Both companies R and D expenditures seem to line up with this chart.

Here is a newer chart.

I was looking into the software part and it's ridiculously expensive. Someone on reddit says the licenses for single software can be a million dollars per year and indeed use the software mentioned by XPEA mentioned(Cadence and synopsys) and they have hundreds of these licences which are only good for one year. A single big chip can use hundreds of licenses in their design.

So would AMD rather uses these resources on something like instinct or a Halo discrete graphics card? I think we all know the answer. AMD is just following the money.

Spending hundreds of millions on a single halo just doesn't make sense when that money can be spent on datacenter where sales are in billions, not hundreds of millions.

 

[CakeMonster]

I know we've been over this before, but in the light of this release, and Blackwell being rather derivate, and arguably Ada and Ampere as well... What does NVidia have coming up over the next, say two generations? Do we just iterate? Is both node and architecture stagnating? Will that make it easier for others to catch up?