4070TI reviews thread

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

BFG10K

Lifer
Aug 14, 2000
22,709
2,990
126

First official performance leak. A perfect opportunity to highlight the fraud pushed by nVidia with DLSS.

Performance looks fantastic, right? Look again, at the smallprint - 4070 is using frame generation.

I said exactly this would happen as soon as DLSS 3.0 was announced. 2560x1440 is a lie given it's actually ~1080p due to legacy DLSS. Since nVidia got away with it, the next step is frame interpolation lies. It's "faster", yo!

Adjusting for frame generation (i.e. halve the 4070's bars), it's actually barely faster than the 3080 in MFS, and slower in Warhammer 40K. All this from a card that will likely cost more.

Customers not seeing the smallprint and/or not understanding DLSS are having systematic fraud perpetuated on them.


Reviews



 
Last edited:

jpiniero

Lifer
Oct 1, 2010
14,907
5,481
136
103 is the new 104 size-wise, its actually even worse than that, since iirc historically 104 used to be 66 percent or 2/3 of the big 102 chip. Ada103 is not even 60 percent of 102 and 104 is now even less than that.

See that's what I mean about Moore's Law being dead. They did keep the AD102 die size relatively comparable but decently cut the die sizes of the lower parts. I don't think it's out of the question that the size difference could be even bigger with Blackwell if it remains monolithic and use N3E/X.

And yes it's going to be tough to get performance gains if the die sizes are going to get even smaller.
 

Heartbreaker

Diamond Member
Apr 3, 2006
4,264
5,260
136
Moore's law isn't really dead, but if you tell people it is they might just believe you and not be as upset over the outrageous prices.

The old economics of Moores law is Really dead. Each new process gets significantly more expensive, so the savings on die shrinks are no where near 1:1 like they used to be in the good old days.

Area cost is increasing dramatically, while scaling is getting choked off for analog and memory components:
1668499294_245727.jpg


Increasing area costs causes transistor economics to flat line. They are still falling, but they are falling MUCH slower now:
main-qimg-81fb704374c332ad5c5e27e11cbd4ac5-lq
 
Last edited:

Mopetar

Diamond Member
Jan 31, 2011
8,036
6,552
136
See that's what I mean about Moore's Law being dead. They did keep the AD102 die size relatively comparable but decently cut the die sizes of the lower parts.

You're mistaking an incorrect conclusion for the cause.

Moore's law is still chugging along.

1672586538170.png
From: https://en.m.wikipedia.org/wiki/File:Moore's_Law_Transistor_Count_1970-2020.png

Of course if Moore's law isn't dead then you'd have a harder time getting people to swallow your new GPU pricing, so you tell them it is. Most will just take it at face value and swallow the lie along with the higher price tag.
 
  • Like
Reactions: psolord

Heartbreaker

Diamond Member
Apr 3, 2006
4,264
5,260
136
You're mistaking an incorrect conclusion for the cause.

Moore's law is still chugging along.

View attachment 73756
From: https://en.m.wikipedia.org/wiki/File:Moore's_Law_Transistor_Count_1970-2020.png

Of course if Moore's law isn't dead then you'd have a harder time getting people to swallow your new GPU pricing, so you tell them it is. Most will just take it at face value and swallow the lie along with the higher price tag.

That says NOTHING about the cost of transistors, which is the actual problem.

Transistors are still shrinking, on each new process, but the new process price increases almost as much as the density.

In the good old days a new process would cost nearly the same as the old one so you would get double the transistors for the same price. This is no longer the case.
 

psolord

Platinum Member
Sep 16, 2009
2,031
1,225
136
That says NOTHING about the cost of transistors, which is the actual problem.

Transistors are still shrinking, on each new process, but the new process price increases almost as much as the density.

In the good old days a new process would cost nearly the same as the old one so you would get double the transistors for the same price. This is no longer the case.

WHY is the cost of transistors increasing though?

Honest question.

If you tell me that new machines are required, the question will be renewed. WHY are the new machines so much more expensive? I mean you needed new machines before. And just as many engineers.

Maybe they need to be regulated at some point and not let them do whatever the hell they want?
 

Heartbreaker

Diamond Member
Apr 3, 2006
4,264
5,260
136
WHY is the cost of transistors increasing though?

Honest question.

If you tell me that new machines are required, the question will be renewed. WHY are the new machines so much more expensive? I mean you needed new machines before. And just as many engineers.

Maybe they need to be regulated at some point and not let them do whatever the hell they want?

A variety of factors, this covers most of them:
 

GodisanAtheist

Diamond Member
Nov 16, 2006
7,084
7,517
136
A variety of factors, this covers most of them:

TLDR- Transistors are getting way more complicated post planer 28nm and require more equipment and steps to fab chips on Finfet and newer transistor technologies. This means more expensive.

Additionally, with the "smartening" of all the things, the economics of depreciated nodes is completely out of whack. Car manufacturers, smart device manufacturers etc aren't interested in the bleeding edge node, they want older cheaper nodes, but no fabs want to ever build out depreciated tech capacity and now they are having to due to unprecedented demand. This makes things more expensive.
 

Mopetar

Diamond Member
Jan 31, 2011
8,036
6,552
136
Moore's Law is about cost, not about density.

This is just wrong. You can certainly argue that as a consequence of doubling, the cost for some fixed number of transistors has decreased, but that's not Moore's law. Even in his original publication on the observed trend, Moore notes that the decrease in cost only holds when you aren't also increasing the number of transistors that you use beyond a certain point as this decreases yield.

If JHH wanted to he could have said that NVidia is using a bleeding edge node and that it's more expensive and so the cards will also be more expensive. That would have been true and who's going to blame a company for passing on costs to customers? However, saying Moore's law is dead is wrong. Trying to rephrase Moore's law to somehow have been about cost is also wrong and disingenuous on top of that.

So now you've had to distort reality to conform to the lie told by someone else.
 

lopri

Elite Member
Jul 27, 2002
13,211
597
126
Is this one going to be a full die where every unit designed is functional, or is this another cut-down stuff and NV is saving the full dies for something else later?
 

jpiniero

Lifer
Oct 1, 2010
14,907
5,481
136
This is just wrong. You can certainly argue that as a consequence of doubling, the cost for some fixed number of transistors has decreased, but that's not Moore's law

That's exactly what it is. Or was.

Is this one going to be a full die where every unit designed is functional, or is this another cut-down stuff and NV is saving the full dies for something else later?

The 4070 Ti is the full die.
 

jpiniero

Lifer
Oct 1, 2010
14,907
5,481
136

Videocardz got a slide which confirms the release date is Thursday.
 

BFG10K

Lifer
Aug 14, 2000
22,709
2,990
126
Moore's Law is about cost, not about density.
I don't believe it is. First sentence from wikipedia: https://en.wikipedia.org/wiki/Moore's_law

Moore's law is the observation that the number of transistors in a dense integrated circuit (IC) doubles about every two years.

Only talks about transistors, nothing about cost. But I don't think anyone denies R&D, manufacturing and facility costs are going up.

I think there are two additional issues here:

Firstly, all three vendors basically redline higher end parts and factory overclock them to the max, jacking up voltages and power consumption to chase that last 5% benchmark graph. This trend started when all three began following the "turbo" paradigm.

They need to get back IPC/efficiency. For example, at one stage Intel followed the mantra of adding 1% power must increase performance by 2%. This produced the awesome Core 2 lineup. Likewise, Pascal architecture, Radeon 9700 pro, etc. All amazing parts to game on. Cool, quiet, fast, and very reasonably priced.

Now all three are returning to the failed Pentium 4 route which is why we have 4-slot furnace slabs pushing 2x600W power connectors. We also have Zen 4 being "normal" @ 95C.

Secondly, nVidia is now virtually a monopoly in the discrete GPU market, so they can effectively charge whatever they like. They have extremely fat margins because of no competition, not because of any failure from Moore's law.

It's time to break up the company - make the GPU division separate to everything else.
 
Last edited:

gorobei

Diamond Member
Jan 7, 2007
3,719
1,093
136
I don't believe it is. First sentence from wikipedia: https://en.wikipedia.org/wiki/Moore's_law

Moore's law is the observation that the number of transistors in a dense integrated circuit (IC) doubles about every two years.

Only talks about transistors, nothing about cost. But I don't think anyone denies R&D, manufacturing and facility costs are going up.

people really need to read the actual paper.

the doubling is premised on component(of a circuit/transistor) price/cost remaining the same gen to gen from the improvements in manufacturing countering the increasing complexity. cost is always a factor in any product.

i've bolded the relevant sections.
Costs and curves
Reduced cost is one of the big attractions of integrated
electronics, and the cost advantage continues to increase as
the technology evolves toward the production of larger and
larger circuit functions on a single semiconductor substrate.
For simple circuits, the cost per component is nearly inversely
proportional to the number of components, the result of the
equivalent piece of semiconductor in the equivalent package
containing more components. But as components are added,
decreased yields more than compensate for the increased
complexity, tending to raise the cost per component. Thus
there is a minimum cost at any given time in the evolution of
the technology. At present, it is reached when 50 compo-
nents are used per circuit. But the minimum is rising rapidly
while the entire cost curve is falling (see graph below). If we
look ahead five years, a plot of costs suggests that the mini-
mum cost per component might be expected in circuits with
about 1,000 components per circuit (providing such circuit
functions can be produced in moderate quantities.) In 1970,
the manufacturing cost per component can be expected to be
only a tenth of the present cost.
The complexity for minimum component costs has in-
creased at a rate of roughly a factor of two per year (see
graph on next page). Certainly over the short term this rate
can be expected to continue, if not to increase. Over the
longer term, the rate of increase is a bit more uncertain, al-
though there is no reason to believe it will not remain nearly

constant for at least 10 years. That means by 1975, the num-
ber of components per integrated circuit for minimum cost
will be 65,000.

Moore even saw the need for chiplets/disagregation.
Day of reckoning
Clearly, we will be able to build such component-
crammed equipment. Next, we ask under what circumstances
we should do it. The total cost of making a particular system
function must be minimized. To do so, we could amortize
the engineering over several identical items, or evolve flex-
ible techniques for the engineering of large functions so that
no disproportionate expense need be borne by a particular
array. Perhaps newly devised design automation procedures
could translate from logic diagram to technological realiza-
tion without any special engineering.
It may prove to be more economical to build large
systems out of smaller functions, which are separately pack-
aged and interconnected. The availability of large functions,
combined with functional design and construction, should
allow the manufacturer of large systems to design and con-
struct a considerable variety of equipment both rapidly and

economically.
though he was still thinking 2 dimensionally when it came to heat.
 

gorobei

Diamond Member
Jan 7, 2007
3,719
1,093
136
WHY is the cost of transistors increasing though?

Honest question.

If you tell me that new machines are required, the question will be renewed. WHY are the new machines so much more expensive? I mean you needed new machines before. And just as many engineers.

Maybe they need to be regulated at some point and not let them do whatever the hell they want?
moore was dealing with 2d(planar) transistors.

resolution improvements (node shrinks) came quickly because the companies were using equipment from the print/photo imaging sector. eventually they had to start making their own equipment to meet their own standards.

when the resolution required for a node shrink went past the limit of the light used to expose the mask resist chemical, they started having to use tricks.

immersion lithography to get past index of refraction of air.
uv light for shorter wavelength.
but eventually they hit the limit of a single pass.

double and quad patterning meant you used multiple passes offset by .5 or .25 to end up with a usable resist layer to work with. 2 passes means either 2 passes on the same machine or 1 pass on 2 machines. so either double the cost of machines or double the time to produce. time is money so they went with 2x machines.
with quad patterning you need 4 machines. fabs got expensive really fast. but that is only if you stay planar/2d.



with finfet and all the other fancy 3d xtor structures you just upped the number of steps per layer.
layers are made by: [absolute broad strokes version. you can get 50-100 times more complicated]
  1. depositing a resist
  2. exposing the resist with a mask to set it
  3. washing away the unexposed resist
  4. depositing the material to construct the component(drain/gate/via/etc) into the washed away locations
  5. baking/setting the material
  6. washing away the former exposed resist with a solvent
  7. filling the void of the area around the component with some inert filler
rinse and repeat.

with 3d structures you are building a lego structure using liquids. it would be like making a cake decoration tiny castle with openings in the wall using only gelatin as a resist and hard meringue for the walls of the castle.
each layer has a bunch of steps that make temporary walls that have to be torn down and then filled with other walls.

getting the machines to shrink the size of the structures they build is only one part of it. all the steps to build the actual structures is the part that is increasing. more steps -> more layer masks -> more machines required + more operator time -> more more more more cost.
 

lopri

Elite Member
Jul 27, 2002
13,211
597
126
Firstly, all three vendors basically redline higher end parts and factory overclock them to the max, jacking up voltages and power consumption to chase that last 5% benchmark graph. This trend started when all three began following the "turbo" paradigm.
Absolutely hate Intel for starting the turbo trend. AMD is even better at this game now, turning every reliability features into a tool to squeeze every last MHz out of the CPUs out of factory. CPU integrity has become impossible to guage for many users, and they will be surprised to hear where their CPUs true stability lies.
 

BFG10K

Lifer
Aug 14, 2000
22,709
2,990
126
Absolutely hate Intel for starting the turbo trend. AMD is even better at this game now, turning every reliability features into a tool to squeeze every last MHz out of the CPUs out of factory. CPU integrity has become impossible to guage for many users, and they will be surprised to hear where their CPUs true stability lies.
Also thanks to this whole ridiculousness, "hotspots" are now a thing. While it might look like a given part has reasonable thermals by all accounts, some hidden corner is able to easily boil water.

I'm done with high TDP parts. With my 9600K I've set PL1 & PL2 to the base spec of 95W to stop any attempt by my motherboard to "auto boost" it into a furnace. And from now on I'm only buying 65W CPUs and will likewise cap their power limits to 65W.
 
  • Like
Reactions: VirtualLarry

Genx87

Lifer
Apr 8, 2002
41,095
513
126
I guess Nvidia's speech this morning didnt include a 4070Ti announcement?

Edit: NM saw it was announced with Jan 5th as availability date.
 
Last edited:

Mopetar

Diamond Member
Jan 31, 2011
8,036
6,552
136
That's exactly what it is. Or was.

If what you're saying were at all true then Moore's law would be "The cost for a number of transistors decreases by half every 18 to 24 months."

But it isn't and even assuming that cost didn't decrease in an inverse manner, we'd still have built smaller components even if the number of transistors remained relatively stable. Governments and businesses would have still purchased chips with more transistors and therefore computational power.

If it were as you say, I'm sure you'll have no trouble pointing to a large number of sources that state Moore's law is about cost. Maybe some will mention the fact that cost decreases as a side effect, but the law itself is always stated as the number of transistors doubling.

So submit proof of your claim. If it always was as you claim there should be no difficulty providing evidence.
 

jpiniero

Lifer
Oct 1, 2010
14,907
5,481
136
If what you're saying were at all true then Moore's law would be "The cost for a number of transistors decreases by half every 18 to 24 months."

Because Moore was assuming that the cost of manufacturing the wafer would remain flatish. And more transistors is what normally drives better performance.

Economics is what drives the whole thing. Because in most cases there wouldn't be much point in going any denser if you didn't get any cost benefit.

Tangently related, Intel just hiked CPU prices 15%... and they should still continue to be getting reduced costs from depreciation with 10 nm. When you buy from TSMC you aren't getting that benefit right now.
 

Mopetar

Diamond Member
Jan 31, 2011
8,036
6,552
136
Also for what it's worth:

3090 (GA102) - 28.3 billion transistors - $1499
4090 (AD102) - 76.3 billion transistors - $1599

~2.7x transistors for ~6.7% more money.

Even if we use the idiotic and twisted definition of Moore's law that's being pushed, JHH is still wrong even by that definition. I'll grant that the Samsung node used for Ampere wasn't quite as good as what TSMC had at the time, but it's not going to substantially change the results here.