Says the L2 of the 4090 has been cut from 96 MB to 72. That's a pretty big cut and just goes to show they are committed to making the performance gaps be meaningful at the ultra high end, as opposed to Ampere.
From the article:TSMC itself categorizes N4 and the variants as part of the 5 nm family:
TSMC Extends Its 5nm Family With A New Enhanced-Performance N4P Node
TSMC introduces a new 5-nanometer derivative - an enhanced performance N4P node.fuse.wikichip.org
It has the same design rules as N5, which is a pretty good indication that it is just an optimization, not a fundamentally different production process.
Says the L2 of the 4090 has been cut from 96 MB to 72. That's a pretty big cut and just goes to show they are committed to making the performance gaps be meaningful at the ultra high end, as opposed to Ampere.
Yeah, it's just Nvidia making up as much marketing as possible to give the illusion that their GPUs are special. When they say the node is custom for Nvidia, it likely just means they used DTCO, which pretty much everyone uses these days to extract as much out of the node as possible. Nvidia marketing machine go brrrrrr is all.No it's just a variant of TSMC 5 that they market as N4P to confuse people.
Tsk tsk tsk. Artificial segmentation, especially when the node likely yields very well, at its finest.
Says the L2 of the 4090 has been cut from 96 MB to 72. That's a pretty big cut and just goes to show they are committed to making the performance gaps be meaningful at the ultra high end, as opposed to Ampere.
Tsk tsk tsk. Artificial segmentation, especially when the node likely yields very well, at its finest.
A100 is also on a custom N7 node and Nvidia didn't lie after some people posted the die analysis. Hopper/Lovelace are really on a custom N5/N4 process that is basically a preview of N4XTSMC itself categorizes N4 and the variants as part of the 5 nm family:
TSMC Extends Its 5nm Family With A New Enhanced-Performance N4P Node
TSMC introduces a new 5-nanometer derivative - an enhanced performance N4P node.fuse.wikichip.org
It has the same design rules as N5, which is a pretty good indication that it is just an optimization, not a fundamentally different production process.
When you're dropping $2k on a 4090 Ti, what's one more MO-RA?Well, if true I think the 4090 will still be the top card that most enthusiasts consider, given that the Ti will probably draw a lot more power. People will have to draw the line not just for cost, but for practical reasons.
Undervolting and underclocking will be the new overclocking.
However, I worry that they expect us to pay for this fake performance.
Says the L2 of the 4090 has been cut from 96 MB to 72. That's a pretty big cut and just goes to show they are committed to making the performance gaps be meaningful at the ultra high end, as opposed to Ampere.
Yeah, practical reasons for sure. I run my 3090 at 60% TDP regularly because my room becomes an oven when it's full tilt otherwise.Well, if true I think the 4090 will still be the top card that most enthusiasts consider, given that the Ti will probably draw a lot more power. People will have to draw the line not just for cost, but for practical reasons.
I never believed his 160 bit 10gb leak. Also believe Nv and AMD may at times put out disinfo to leakers.Kopite is now saying 12 GB for the 4070, which would be nice if true.
A cut-down version could have 160-bit 10GB, but of course a full chip would be 192bit 12GB.I never believed his 160 bit 10gb leak. Also believe Nv and AMD may at times put out disinfo to leakers.
A cut-down version could have 160-bit 10GB, but of course a full chip would be 192bit 12GB.
RTX 3080 for example is 320bit 10GB.
The rumors say that the 4070 is the full AD104 and that the cut down version will be a 3060 (Ti). But I still worry that the 12 GB version will be the more expensive 4070 Ti version based on a cut down AD103, and that the full AD104 will have 10GB.
Then again, we often see that the cut down version offers best value, so the 4070 Ti may be the new 3080 when it comes to being a sweet spot.
If that ends up being the final spec of the 4070, there probably won't be a 4070 Ti. nVidia could use the cut spec for the mobile 4070 and that might be enough.
What I mean is that this 4070 would be the full AD104. Maybe it has some cut L2 but that doesn't leave much for a theoretical 4070 Ti.
It's supposed to be "win everything, everywhere, all at once"-If history has taught us anything about NV, its that there will be refreshes and stop gap cards and "response to AMD" cards and everything in-between. Once upon a time it seemed like NV had a proper product stack but now it feels more like NV just has a stack of dollar bills. NV is absolutely not concerned with any sort of coherent product stack at this point and is all in on the "win everywhere all at once" train.
If the 3070Ti can exist, then a 4070Ti can exist, no problem. If the 3060Ti can be 10/15% the 3070 but 30% ahead of the 3060, then the 4070Ti can exist, no problem. If a 3060 can have 12gb of ram and a 3080 can have 10gb of ram then the 4070Ti can exist, no problem.
They're just names after all and if it beats the competition by even 5% then that's its raison d'etre, even if its hypothetical and cannot be found on store shelves.
Considering the humongous process advantage this sounds about right. Well, maybe a bit on the high side since these are not AD102 based. They better be close to 2x faster or the perf/w improvements (or "improvements") are disappointing.NVIDIA GeForce RTX 4080 and 4070 get new (lower) power specs, mysterious 20GB model also mentioned - VideoCardz.com
GeForce RTX 4080/4070 with lower power requirement Two hardware leakers report on NVIDIA RTX 40 specs. According to Kopite7kimi, the RTX 4080 and 4070 specs have changed again, but this time the update is only for the power requirements. The previously reported total board power of 420W for RTX...videocardz.com
nVidia has apparently decided to cut the official TDP of the 4080 to 320 W and the 4070 to 280. Guessing this is just a marketing thing though and the FE and most AIBs will be the earlier higher number. Unless they've gotten some intel that RDNA 3 is slower but draws less than they think.
Also there's an even deeper potential cut of AD102 with 112 SMs mentioned, but I have a feeling that is a "Quadro" and not the 4080 Ti. They really shouldn't need to cut that far.
To me it looks more like kopite doesn't know what he's talking about anymore. At this point his "monday: 420W, wednesday: we can expect 350W, saturday: another update now we're at 320W for the 4080" is just annoying. And all of this "there's another change" stuff is just bs although he won't admit it.nVidia has apparently decided to cut the official TDP of the 4080 to 320 W and the 4070 to 280.