I agree. The SM improvements just seems to be:Naaa, the SM is just very underwhelming to say the least. I guess this had a higher fmax target?
They didn't showcase L1/shame size either, or any scheduler capacity bumps. No real actual™ SM diagram. Lame.I agree. The SM improvements just seems to be:
- Added back INT to the previous FP-only pipe (meh)
- Added FP4/INT4 to the tensor units (meh)
- Doubled ray-triangle intersection (not sure if the ray-box intersection has improved)
- Support for neural shading (meh)
- Support for SER on the tensor units for neural shading (meh)
Too big, too pricey.Was there a reason they couldn't have used TSMC 3nm for these cards?
You should just be happy Nvidia is allowing people to play games on their AI cards at all and quit complaining.Was there a reason they couldn't have used TSMC 3nm for these cards? This is worse than I expected. 5090 is okayish, but that appears to strictly be because of the massive die and massive power usage. 5080 is pretty bad considering it's a new gen.
Cheating how? Same die size for 4080 and 5080. Some more performance. All at the same low price. And best of all you get 2x more FPS.This ladies & gentlemen, is clear evidence of cheating.
I agree. The SM improvements just seems to be:
- Added back INT to the previous FP-only pipe (meh)
- Added FP4/INT4 to the tensor units (meh)
- Doubled ray-triangle intersection (not sure if the ray-box intersection has improved)
- Support for neural shading (meh)
- Support for SER on the tensor units for neural shading (meh)
That would pair well with GDDR7 not seeming to have the kind of uplift expected. Cutting the cache reduces the effective bandwidth and then GDDR7 picks the slack back up.Considering GB203 has fewer transistors despite having more SMs and the architectural improvements, my guess is Nvidia reduced L2 cache size. Potentially halved it to 32MB. To make up for that they use GDDR7 now.
Was there a reason they couldn't have used TSMC 3nm for these cards? This is worse than I expected. 5090 is okayish, but that appears to strictly be because of the massive die and massive power usage. 5080 is pretty bad considering it's a new gen.
Considering GB203 has fewer transistors despite having more SMs and the architectural improvements, my guess is Nvidia reduced L2 cache size. Potentially halved it to 32MB. To make up for that they use GDDR7 now.
They used a customized node 4N which many said was an improved N5. But it appears it was actually closer to N4. And the new customized node they're using 4NP is eerily close to the previous one (4N). In fact, I'd say it's a very minor revision with same density and doesn't even clock much higher, but should be cheaper by now. So, for Blackwell they're using the (almost) same damn old node. No more "Better than Moore's Law" nonsense.An embarrassing improvement. No wonder the presentation was focused on DLSS 4.
Looking at the transistor density, it also seems last generation they were using something closer to TSMC 4NM rather than 5NM.
Being too AI focused is finally hurting gamers.
Xtor count went up roughly proportional to the die size (~21% more xtors, ~23% larger die). I think the OFA still exists, since it gets used for other applications, but it's no longer leveraged for frame generation.At least they seemed to have done this without really increasing the transistor count and seemingly only removing the Optical Flow Accelerator.
Sit down son, we need to have a talk.
And don't you try to pay with them fake dollar bills, I can smell 'em a mile away, you hear?
I maybe wrong, But at this point, I'm starting to believe AMD RX 9070 may easily match or surpass RTX 5070 in raster. My guess is, 9070 might fall between 5070 & 5080. And if AMD can price it right, they have a clear winner.I was still hopeful Nvidia was sandbagging clock rates and the 5080 would match 4090 in raster.
I maybe wrong, But at this point, I'm starting to believe AMD RX 9070 may easily match or surpass RTX 5070 in raster. My guess is, 9070 might fall between 5070 & 5080. And if AMD can price it right, they have a clear winner.
Xtor count went up roughly proportional to the die size (~21% more xtors, ~23% larger die). I think the OFA still exists, since it gets used for other applications, but it's no longer leveraged for frame generation.
View attachment 115032
no this just smells like an fmax missSo much that NVidia will make the worse generation gains ever, just to keep die size from growing.
I had high hopes for 50 series. But considering many of the DLSS4 features are being backported to 40 series, I think it's better not to upgrade anytime now.If those are accurate it really makes clear, how critical die pricing has become. So much that NVidia will make the worse generation gains ever, just to keep die size from growing.
I’m not so sure. 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.no this just smells like an fmax miss
They're squeezing blood out of stone. And then some. 4NP is a dud.I’m not so sure. 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.