Blackwell uses the same 4N process as Ada, and it looks like the RTX 5090 is clocked really low.
Those numbers don't make any sense.
The 5090 has 33% more SMs.The 5080 has 5%, 5070 Ti 6%, and 5070 4% (over 4070 NS). You'd expect the gap between the 4090 and 5090 to be 25-30% more than the other ones.
Blackwell uses what they are calling 4NP. The most dense silicon of the 4/5nm TSMC silicon. It gives maybe 1-2% efficiency boost and about even on performance uplift over ADA. Based on AMD and Intel silicon, Nvidia is still using the most advanced silicon with Blackwell. The highlight for Nvidia is the 30gbps GDDR7 memory.Blackwell uses the same 4N process as Ada, and it looks like the RTX 5090 is clocked really low.
*Scoffs in arrow lake*Blackwell uses what they are calling 4NP. The most dense silicon of the 4/5nm TSMC silicon. It gives maybe 1-2% efficiency boost and about even on performance uplift over ADA. Based on AMD and Intel silicon, Nvidia is still using the most advanced silicon with Blackwell. The highlight for Nvidia is the 30gbps GDDR7 memory.
ITS N4P AHHHHHHHHHBlackwell uses what they are calling 4NP.
Check the diff from N5P/N4 to N4P, it is the same.The most dense silicon of the 4/5nm TSMC silicon. It gives maybe 1-2% efficiency boost and about even on performance uplift over ADA.
Wrong, same stuff, different DTCO.Based on AMD and Intel silicon, Nvidia is still using the most advanced silicon with Blackwell.
And the perf bump seems to be under the bandwidth bump, which is underwhelming.The highlight for Nvidia is the 30gbps GDDR7 memory.
And the perf bump seems to be under the bandwidth bump, which is underwhelming.
My bet is Nvidia lowered the L2 cache size from AD102 to GB102.
Why wouldn't they, with such a massive difference in VRAM bandwidth?
Pretty sure kimi already said 128M L2 for the full die.TPU seems to think it is reduced but the 5090 has a shallower cut, so it has 88 MB vs 72 on the 4090.
If the TPU preliminary spec differences are accurate, it's a combination of factors.Those numbers don't make any sense.
The 5090 has 33% more SMs.The 5080 has 5%, 5070 Ti 6%, and 5070 4% (over 4070 NS). You'd expect the gap between the 4090 and 5090 to be 25-30% more than the other ones.
Which is logical if Moore's Law scaling stops, while memory continues to improve. If you have two identical chips and you have an option to use next gen memory standard you'll use it.And the perf bump seems to be under the bandwidth bump, which is underwhelming.
DF Direct Weekly #196 - CES 2025 Special! - Nvidia, AMD, Intel Highlights + More!
Nice sections on Mega geometry, ReSTIR PT vs Lumen GI and the insane Black State demo. Also a second view on AMD CES PT demo.
View attachment 114875
This is impressive. Honestly the best part of new cards is not the 50 series but the DLSS4 features.
bit dramatic, these are companies their job is sell products/services. I don't care about whatever bull NV says. I see an improvement, I make a note of it and this improvement will help users with older cards-Nvidia marketing is really something else.
"Remember that thing we told you was awesome and you absolutely had to buy our cards to use it? It's trash. Garbage. You should be ashamed of yourself for liking it.
So anyway, here is the new version you can't live without..."
Takes me back to Turing and how Ray Tracing was here, photorealistic lighting. The end of rendering.
Now we're 3 gens in and RT was garbage and you need Ray Reconstruction but what you really want is Path Tracing and...
-Nvidia marketing is really something else.
"Remember that thing we told you was awesome and you absolutely had to buy our cards to use it? It's trash. Garbage. You should be ashamed of yourself for liking it.
So anyway, here is the new version you can't live without..."
Takes me back to Turing and how Ray Tracing was here, photorealistic lighting. The end of rendering.
Now we're 3 gens in and RT was garbage and you need Ray Reconstruction but what you really want is Path Tracing and...
Ray Reconstruction is software that works on all RTX cards.
Basically a free update for everyone, that improves things. How Dare They???
I agree that FG/MFG gating is sketchy, and FG/MFG marketing is beyond sketchy, but not everything NVidia does is automatically sketchy.
Full RT is hard on current hardware and RR "helps" to reduce the performance impact.Ray Reconstruction sort of feels like it's giving up the gig on Ray Tracing.
Ray Tracing inherently adds noise to the frame because of how light rays traverse and end up. It is just one of those things that is accepted as a challenge for an overall more realistic lighting if the scene is set for it. From the days of pixelated graphics to anti-aliasing and texture filtering, and limitations likewise in video capture, every new technique brought a step change even if it created another set of problems to solve.-Ray Reconstruction sort of feels like it's giving up the gig on Ray Tracing. Don't ray trace, just sort of do the semblance of Ray Tracing and then use deep learning to fill in the gaps and get a final result.
RT was sold as "no cheap trick lighting" and it's all just going back to cheap tricks..
Big increase in all resources and they could not get 50%+ proper non-fake perf out of it ...Pretty sure kimi already said 128M L2 for the full die.
Not a good client uarch. Happens.Big increase in all resources and they could not get 50%+ proper non-fake perf out of it ...
It's not.Seen report (rumour) that Rubin is pushed into Q3-Q4 2025 on N3
No, 8Q cadence awaits as usual.Perhaps this whole 50 series will be shortlived...
It's straight up lying nowadays, accepted by the public as being ok. Intel claiming Arrowlake mobile beats competition, secretly hiding that it takes more than twice the power to do so, while Nvidia is pretending copying frames is more performance.-Nvidia marketing is really something else.
Everyone does it these days, remember the days when keynotes didn’t include fudged up graphs or fake frames or ballooning power consumption to cover up performance deficits.It's straight up lying nowadays, accepted by the public as being ok. Intel claiming Arrowlake mobile beats competition, secretly hiding that it takes more than twice the power to do so, while Nvidia is pretending copying frames is more performance.
The single most important reason for more frames is for increased responsiveness. If you don't have it, then you might as well not have it.
If anti-trust agencies think they are actually doing something good for the society, this is what they should fix. Nvidia has so much mindshare that lying through the teeth is acceptable.Everyone does it these days, remember the days when keynotes didn’t include fudged up graphs or fake frames or ballooning power consumption to cover up performance deficits.
Eh, smoother picture is smoother picture. The problem is NV selling interpolation as a more perf technique (it's not).The single most important reason for more frames is for increased responsiveness
Well it's a byproduct of Blackwell client being really weak.remember the days when keynotes didn’t include fudged up graphs or fake frames or ballooning power consumption to cover up performance deficits.
Yes I do think companies should be fined large amounts for deceptive marketing.If anti-trust agencies think they are actually doing something good for the society, this is what they should fix. Nvidia has so much mindshare that lying through the teeth is acceptable.
Forget the whole "too big to fail" deal. Short term pain, long term gain. Stealing without using the literal word.
They've already said it's 48-50 vs 55W.No mention of power consumption and talks of leadership when the comparison is only done against M4 pro and not the M4 Max.
Pipe down champ, Max has twice the membw.I bet the M4 Max uses less power than the Strix Halo and is still more powerful in 3D work.