The State of the GPU: All Fake MSRPs?

[Watzupken]

Nvidia has an immense backlog of Blackwell cards for the server market - currently an expected delivery time of over 12 months. The biggest problem for Nvidia is allocating their production line at TSMC for consumer gpus when they need to use it to produce server gpus enmasse. Wafers of lower quality will be set aside for consumer gpus - and then produced in batches between filling the backlog for server gpus.
Fact is we probably won’t see a «normal» supply of consumer gpus until the initial AI boom is over - which could take at least a couple of years. There’s too much money in AI at the moment

I feel that this is not the main reason for the miserable supply of consumer GPUs. My take is that (1) GDDR7 supply is not keeping up since it is new, and, (2) there may be a wider problem with Blackwell dies which resulted in missing ROPs. Nvidia claimed that the missing ROPs issue should only impact 0.5% of the population, but what they did not say is that most of them would have been detected before it leaves the factory or foundry. Nvidia claimed that it will not impact laptop GPUs, but they asked laptop makers to check every single one of them before sending them to the end client. So I feel it is a wider issue that they are trying to make it sound fine, at least not 0.5%.

 

[Eldrach]

I feel that this is not the main reason for the miserable supply of consumer GPUs. My take is that (1) GDDR7 supply is not keeping up since it is new, and, (2) there may be a wider problem with Blackwell dies which resulted in missing ROPs. Nvidia claimed that the missing ROPs issue should only impact 0.5% of the population, but what they did not say is that most of them would have been detected before it leaves the factory or foundry. Nvidia claimed that it will not impact laptop GPUs, but they asked laptop makers to check every single one of them before sending them to the end client. So I feel it is a wider issue that they are trying to make it sound fine, at least not 0.5%.

Considering 1 single GB200 NVL72 uses -> 72 fully unlocked Blackwell GPU's - I'm inclined to believe it's more a matter of production capacity. The GB200 doesn't use GDDR7 - but HBMe 3
I do know Samsung has had some issues with the memory...But it looks like the world is just too hungry for AI servers now. A single GB200 NVL72 has a price of 3 million USD.
For 1 single rack server they would have to make 1500 RTX 5090 for that amount of revenue - using only 72 GPU's to get there for the GB200 NVL72. There's no doubt AI racks is just printing cash for Nvidia

 

[OortCloud]

To fully grasp just how poor the 4 and 5 series GPUs have been and how disappointing the AMD release is too...

I bought a 3080 FE at launch nearly 5 years ago for $699! In the 18 game average the 9070 XT is still only 25% faster. They are similarly priced and more often, given the MSRP nonsense, far more expensive. That's 5 years of zero progress. I don't understand why reviewers are giving the 4 and 5 series the scores they are and the AMD GPUs aren't much better - especially the non XT.

 

[Eldrach]

To fully grasp just how poor the 4 and 5 series GPUs have been and how disappointing the AMD release is too...

I bought a 3080 FE at launch nearly 5 years ago for $699! In the 18 game average the 9070 XT is still only 25% faster. They are similarly priced and more often, given the MSRP nonsense, far more expensive. That's 5 years of zero progress. I don't understand why reviewers are giving the 4 and 5 series the scores they are and the AMD GPUs aren't much better - especially the non XT.

Well - 3080 FE was a "top tier card" - with inflation that 700 dollar card would be around 1000 dollars now - which means you need to compare it to the 5080 and the effective speed of the 5080 is 77% faster than the 3080 - that is, if you could find the 5080 at MSRP ..Which you can't at the moment and probably not for a while....but STILL..I heard the guy in the leather jacket say 1000usd msrp!

 

[OortCloud]

Well - 3080 FE was a "top tier card" - with inflation that 700 dollar card would be around 1000 dollars now - which means you need to compare it to the 5080 and the effective speed of the 5080 is 77% faster than the 3080 - that is, if you could find the 5080 at MSRP ..Which you can't at the moment and probably not for a while....but STILL..I heard the guy in the leather jacket say 1000usd msrp!

I know it was more expensive (there's a site that tells you exactly how much, and $699 in August 2020 is $840 today), so not that far from the prices these cards are actually going for.
I was trying to make a point about how little progress has been made, how stagnant the GPU market has become and how poor the value offered by modern 'mid tier' cards is especially when you disregard their MSRPs and look at the prices AMD and NVidia always knew they would be sold at.

 

[Eldrach]

I know it was more expensive (there's a site that tells you exactly how much, and $699 in August 2020 is $840 today), so not that far from the prices these cards are actually going for.
I was trying to make a point about how little progress has been made, how stagnant the GPU market has become and how poor the value offered by modern 'mid tier' cards is especially when you disregard their MSRPs and look at the prices AMD and NVidia always knew they would be sold at.

Yeah, in one regard you are absolutely correct - we are starting to see physical limits of the current sillicone level. We have maybe two more iterations of regular sillicone before we hit the absolute wall (1,2 - 1,4nm) which will give us maybe 30-40% more raster performance. After that we're at the point where we need to find new production techniques or as we see now - they add new types of enhancements on the table to try to get around the limitations of the actual GPU. I think the Framegen stunt they are trying to pull now is pretty bad - it's basically an advanced "image smoother" and not really any form of advancement. Neural rendering might actually be progress - when we actually see it implemented in any game, of which there are currently 0.

 

[DonquixoteIII]

Hmmm... Manufacturer's SUGGESTED Retail Pricing...

 

[Resonator]

Screw Nvidia and AMD - has anyone actually seen an Intel Arc B580 for sale? I've been looking in the usual online places for US since launch day, and have only found the insanely overpriced Gunnir and Weeliao models.

 

[Resonator]

I've read that May is the ETA for ASRock Intel Arc B580 GPUs at least, so I'm hoping that we'll see other AIB boards around that time as well. Just FYI, this info was from centralcomputer.com and they are listing the two ASRock cards at close to MSRP.

 

[Resonator]

I really don't get the mentality behind the "I must have it now" thought process. The only people that are really forced to buy immediately are people who have a hardware failure. Everyone else is free to wait until the hardware they want becomes available at a price they are willing to pay.

Totally agree. I'm waiting on an Intel Arc B580 myself and am currently using a RX 580 I picked up new for $90 as a temp fill in on my new build. It's running all my GOG games just fine.

 

[godrilla]

A large RDNA 4 chip like the 4090/5090 would be a beast. I hope they keep tweaking the software as well because you can see that there's still performance on the table and buggy games that perform badly on RDNA 4.

The cost of not having a high end gpu is not competing with the new Rtx pro 6000 gpu. I can't find anything articles discussing what AMD has to succeed the Radeon pro w7900 with rdna 3 to compete. While Nvidia is dedicating Blackwell gaming gpus for 4x the premium vs 5090.

 

[Loadedaxe]

This may explain why MSI isn't producing the AMD cards this year. I don't think its in everyone's interest for AMD to control the pricing they receive from rebates when different variants are made.

This is a costly game AMD is playing and could cause more AIBs to jump ship. They want their profit and rightly so, controlling how much they are going to make on their investment is not good business.

Nvidia, well, we all know they do not care about gamer's anymore, at least until the Ai bubble bursts.

 

[WhiteLeaff]

The cost of not having a high end gpu is not competing with the new Rtx pro 6000 gpu. I can't find anything articles discussing what AMD has to succeed the Radeon pro w7900 with rdna 3 to compete. While Nvidia is dedicating Blackwell gaming gpus for 4x the premium vs 5090.

The N48 is small compared to the 5090 chip. AMD can use a dual GPU configuration like Pro-duo was, add a ton of memory and voila you have a superior product with lower BoM. This works well for the professional and AI market.