Nvidia Blackwell Architecture Speculation

[homerdog]

The first link is 2nd hand sales on cards which were used and returned to the store.
And the second one is for UK where 5070's MSRP is £539 which is roughly $700 in USD which is +$150 to the US MSRP.

I've been saying for some time that it's the USA market in particular which is thoroughly fucked for some reason when it comes to GPU availability, and the fact that US has generally lowest prices on them could actually be one reason for that. Outside of US cards are much easier to find - but they do tend to cost upwards to 2X of USA MSRPs which could explain exactly why they are easier to find.

Yea I dunno what kind of weird shit is going on at walmart but you can't buy a 50 series card for below MSRP lol. The cheapest one on Newegg is $700 and it comes with a big ass CPU cooler...

https://www.newegg.com/Product/ComboDealDetails?ItemList=Combo.4772153

At least you can buy that. All the other 50 series cards are 100% unavailable. I checked some of my distributors (Ingram and D&H) and it's no better there.

Don't know about other countries either. But I wouldn't be surprised if they have better stock if they are way more expensive.

I don't feel that bad anymore about paying $600 for my 4070 two years ago. Looking at ebay it hasn't depreciated at all. In fact most of them (used) are going for >$600.

 

[trinibwoy]

Yeah I don’t know about these price drop claims. All 50 series cards are getting more expensive and just as scarce in the US. The only message consumers are sending is that they’re willing to spend even more money.

I’m almost at the point where I’ll decide to roll with my 3090 another 2 years since it’ll probably play everything fine until then especially with the DLSS upgrade. The 5090 is a beast that will likely top the charts for years but lack of availability and power consumption are dulling my excitement.

 

[TopSpoiler]

Update on the Blackwell's integer performance.
Source

If the number of cores is n, Ada Lovelace can execute n/2 IADD3 instructions per cycle and Blackwell n IADD3 instructions. Ada Lovelace can execute n/2 IMAD and Blackwell n IMAD. But Ada Lovelace can execute n/2 IADD3 and n/2 IMAD per cycle in the same way as Blackwell.

If a benchmark evaluates instructions individually then INT32 performance is 2x. But in practice, where instructions are mixed, INT32 performance is 1x.

 

[raytracingfan]

We can't pretend that 8GB VRAM is a sufficient baseline when Nvidia itself is pushing for VRAM-hungry AI and RT. Project G-Assist requires 12GB VRAM, so it will run on an RTX 3060 but not on a 3060 Ti, 3070, 3070 Ti, 3080 10GB, 4060, 4060 Ti 8GB, 5060 8GB, and 5060 Ti 8GB. With Mega Geometry encouraging developers to store full-detail geometry in the BVH, VRAM requirements for RT will only increase. The ACE small language model used in inZOI takes 1GB VRAM on its own with very limited usage, more robust models and greater usage of the model will take up more. Finally, Nvidia's own driver features like RTX HDR and Nvidia Broadcast have a VRAM cost of their own, and third-party programs commonly used while gaming like Discord can have a VRAM cost if GPU hardware acceleration is on.

Apple is also notoriously stingy with memory, but now that they've joined the AI bandwagon and started promoting Apple Intelligence they've made 16GB the new baseline for Macs. Nvidia, which has been leading the charge on AI, should have already raised its baseline. It's simply incongruous behavior to pack the entire GPU lineup with advanced tensor cores and RT cores, advertise your advantages in RT and AI over the competition, but then give the budget cards - which have always been the most popular - too little VRAM to fully utilize those advantages.

 

[trinibwoy]

We can't pretend that 8GB VRAM is a sufficient baseline when Nvidia itself is pushing for VRAM-hungry AI and RT. It's simply incongruous behavior.

According to HUB Nvidia’s official position is that if you want the cool new tech you gotta buy expensive cards. So they probably don’t see it as incongruous at all.

Daniel Owen has a more sober take on the 5060 Ti 8GB situation. He shows a few games where 8GB is a problem but also acknowledges there are many games where it’s not. Don’t know why it’s so hard for tubers to offer balanced opinions but glad there are a few out there.