Ada RTX 6000: why did Nvidia cheap out on the cooling solution and VRAM?

[80251]

Considering Nvidia is selling this product for $6800 why didn't they engineer a better cooling solution and/or use GDDR6x instead of GDDR6? I can see how a larger cooling solution might not work if it's designed to fit in a 2RU rack but you would think they could engineer something better than an early 2000's era blower for nearly $7k.

Interestingly enough the Nvidia L40S isn't even available yet which has the same specs as the Ada RTX 6000 with higher clock speeds.

 

[Vya Domus]

Double the memory density means more heat, so they went with the cheaper and more reliable option, also there are probably no double density GDDR6X modules available.

They could engineering something better for the cooler but there is no point in doing so, again, they go with safe and reliable option, something that you toss into a system and do not care about it until you change it.

 

[londiste]

48GB of VRAM on 384-bit bus still means clamshell mode as both GDDR6 nor GDDR6X are available only in chips up to 2GB. GDDR6X is much more capricious about the memory bus, it consumes considerably more power and thus emits more heat. The last one is especially a problem due to clamshell mode because there are VRAM chips on the backside of PCB. RTX3090 had that same configuration and it was somewhat problematic in those aspects, something you would not want to see in a professional card.

As for cooling what exactly is "better"? Blower is a very safe cooling, you can stick it into pretty much anywhere and expect it to work. Or in some worst case scenario survive. These are professional cards so stacking a case full of them is not unusual either, both dual-slot and blower will come in handy in these scenarios. And Nvidia's blowers are really good at what they are build to do.

 

[AnotherReader]

An additional benefit is that blowers work better in multi-GPU setups. Unlike gaming computers, workstations can often have multiple GPUs for rendering or other work that can be parallelized. An open air cooler would be sub-optimal in that case.

 

[80251]

Considering the way these dual slot, blower videocards are jammed into 2U enclosures it's no wonder they run at such slow clock speeds:

https://www.gigabyte.com/Enterprise/GPU-Server/G292-Z40-rev-100

But you would think they would be cheaper than a 4090, not 400% more expensive, unless twice the GDDR6 memory adds that much to the BOM. Even though the RTX 6000 and L40 series have more shaders/TMUs/ROPs/L2 cache I'll bet they're not as performant as a 4090 because of the relatively slower core clocks and slower memory.

 

[Deleted member 182555]

those are new 20gbps gddr6 they go a little less than the 21/24gbps gddr6x...

 

[Assimilator]

Considering Nvidia is selling this product for $6800 why didn't they engineer a better cooling solution and/or use GDDR6x instead of GDDR6? I can see how a larger cooling solution might not work if it's designed to fit in a 2RU rack but you would think they could engineer something better than an early 2000's era blower for nearly $7k.

Interestingly enough the Nvidia L40S isn't even available yet which has the same specs as the Ada RTX 6000 with higher clock speeds.

It's engineered exactly as well as it needs to be for its target market. As are AMD's offerings for the same marketing.

 

[DemonicRyzen666]

what do mean cheap out?
I'm pretty sure there is more copper inside the workstation cards blower, than are inside of any other consumer card that's a blower too.

 

[Dr. Dro]

Double the memory density means more heat, so they went with the cheaper and more reliable option, also there are probably no double density GDDR6X modules available.

They could engineering something better for the cooler but there is no point in doing so, again, they go with safe and reliable option, something that you toss into a system and do not care about it until you change it.

Well, there are suitable memory chips now. It's the power consumption and reliability concerns. Hot hardware is unreliable hardware.

But you would think they would be cheaper than a 4090, not 400% more expensive, unless twice the GDDR6 memory adds that much to the BOM. Even though the RTX 6000 and L40 series have more shaders/TMUs/ROPs/L2 cache I'll bet they're not as performant as a 4090 because of the relatively slower core clocks and slower memory.

The margins on these things are insane, BOM is not a concern... it probably costs Nvidia $400 to build one of these, and that's probably shooting it high, too.

 

[80251]

@Dr. Dro, more than a 1000% markup!!!!! Ngreedia really lives up to their moniker don't they?

So would Nvidia's driver paradigm prevent someone from using a much cheaper(!) 4090 in the same capacity as a L40S or RTX 6000?

 

[claes]

You’re forgetting about R&D… It’s not like some guy was sitting on a toilet and thought “hey I have a great idea for a GPU architecture” and they started producing them that week.

i’m not saying that nvidia isn’t making a boatload on GPUs, but even 100% markup is probably hyperbole.

 

[Dr. Dro]

@Dr. Dro, more than a 1000% markup!!!!! Ngreedia really lives up to their moniker don't they?

So would Nvidia's driver paradigm prevent someone from using a much cheaper(!) 4090 in the same capacity as a L40S or RTX 6000?

It's as @claes said, these segments cover R&D which is extremely expensive, and a lot of the money is also spent on software maintenance. Driver engineers don't work for freer. AMD does the same with the Radeon Pro and Instinct lines, and Intel offers Arc Pro - the certification they go through is extensive - and expensive.

 

[80251]

Wouldn't all R&D expenses and software/driver maintenance as well be tax deductible? Even if they only made a penny per videocard their R&D would already be covered.

 

[claes]

Wouldn't all R&D expenses and software/driver maintenance as well be tax deductible?

I don’t believe so?

Even if they only made a penny per videocard their R&D would already be covered.

I think you’re grossly underestimating the cost of R&D… nvidia spent $2 billion on it in the past quarter

 

[80251]

It's still strange to me that an Ada RTX 6000 w/an inferior cooling solution and roughly the same performance as an RTX 4090 sells for 3x the price -- although I'm very glad the price-points for both videocards isn't reversed -- otherwise I could only afford to be upgrading every decade. The Ada RTX 6000 doesn't even come w/NVLINK support!!

https://www.tomshardware.com/news/rtx-4090-beats-rtx-6000-ada-in-content-creation-performance