Where is that from? I'm trying to wrap my head around what a T239 Ampere (RTX 20 series) means. 3rd Gen Tensor is Ampere, but what's with the T designation and RTX 20 description.
Where is that from? I'm trying to wrap my head around what a T239 Ampere (RTX 20 series) means. 3rd Gen Tensor is Ampere, but what's with the T designation and RTX 20 description.Switch 2 Techs Specs.
Tensor Core seems from ADA, but 2 RT cores?
Cuz it's literally an Orin derivative.Looks surprisingly similar to Jetson Orin NX/AGX
Wrong, GB202 is the chungusest chungus ever, the first 512b NV part since GT200 too.
no lmao, they'll have to rechop the stack with Supers again.The rest of the product stack can be stagnant or sold on new shiny RT/DLSS technology as opposed to better hardware capabilities.
What I mean by this is that Nvidia doesn't need to release a more cut down GB202 to slot in against some hypothetical AMD GPU that beats GB203. Instead they can charge what they want for the 5090 and let the 5080 be a full die GB203 instead of having to use a more cut down GB202 die.
The 5090 is the only card that looks like a significant upgrade because it's the only one that has to be. The rest of the product stack can be stagnant or sold on new shiny RT/DLSS technology as opposed to better hardware capabilities.
they killed it. that range was addressed by navi41/42.Which is why they probably didn't push that far this time.
no. RTRT is very incremental outside of 2x the tri-rate (not the limiter) and GEMM cores just support new numerics.Didn't the 50 series have big increases in RT and AI hardware?
no. RTRT is very incremental outside of 2x the tri-rate (not the limiter) and GEMM cores just support new numerics.
They didn't. 'modern' ML (xformers) is a Google invention.who innovated their way towards, and invented modern A.I.
What's the limiter?RTRT is very incremental outside of 2x the tri-rate (not the limiter)
Memory latency (cache too, but you'll be hitting mem in any relevant game scene anyway).What's the limiter?
You need far more CPU-like cache and memory latencies to get performant RTRT.I was under the impression that memory played big role, so 70% higher bandwidth would have been very beneficial
Sorta. You can only hide so much latency in your RTRT pipeline.More memory bandwidth increases the performance of the entire pipeline which increases RTRT performance too.
So bigger meaner 256 MB+ cache should help a fair bit then?You need far more CPU-like cache and memory latencies to get performant RTRT.
Rubin this year - might even be Q3, but that is very likely going to be HBM part only like Volta, certainly can't see them releasing 60 series earlier than 2026.Any ideas when their next-gen 3nm GPU parts are gonna be released?
Some Open CL 4090 vs 5090About 37% faster than the average 4090.