• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

NVIDIA 4000 Series

My understanding is when these are made some cores are faulty and can’t make a full fat chip so they end up disabling the damaged parts and making a cut down variant. There is also the marketing side of things. Trust me, they have all the data to make decisions to maximise profits, you don’t need to worry about that.

There will be some perfect dies but not a lot, on the other side a few chips will have a defect somewhere important and be scrap but the vast majority will have like 80% of the cores available
 
There will be some perfect dies but not a lot, on the other side a few chips will have a defect somewhere important and be scrap but the vast majority will have like 80% of the cores available

Yea, as usual we won’t get the top dog straight away and by the time it comes the 4000 series will be past half its life cycle. Either that or make it available earlier at a huge premium for those chasing epeen. But I doubt they would do this, as it makes more sense to let people buying the xx90 class feel like they have the best for a while as relatively speaking they represent a much larger segment of the market.
 
Fascinating, but one does not change architecture overnight - so if we're only hearing about this now, RTX4000 would have been changed already months ago and leakers somehow didnt know?

If true then the performance gains could be higher than expected; Nvidia is really worried about AMD.

Indeed yes. Will be interesting to see how this develops. Nvidia only just scraped by with the performance crown this generation, so I think they have every reason to be worried about the next generation.

Either way, whoever releases the best product gets my money, that includes efficiency. I'd rather compromise a little bit of performance for efficiency.
 
Indeed yes. Will be interesting to see how this develops. Nvidia only just scraped by with the performance crown this generation, so I think they have every reason to be worried about the next generation.

Either way, whoever releases the best product gets my money, that includes efficiency. I'd rather compromise a little bit of performance for efficiency.
I wouldn't discount Nvidia, my 1070 was a very good card, but for my target if 300W max out is sounding like AMD will be the most efficient. Looking forward to seeing what both launch, some good competition would be welcome.
 
Apparently Lovelace has been cancelled. Nvidia 4000 will now be based on Hopper

One does not simply move project time lines without consequence.

If true, and depending on how many months they've brought it forward by this could be a very unpolished launch.
 
One does not simply move project time lines without consequence.

If true, and depending on how many months they've brought it forward by this could be a very unpolished launch.
Not like the last few generations of launches have been anything to write home about from both parties. :p
 
Nvidia has released the whitepaper for Hopper https://resources.nvidia.com/en-us-tensor-core

* Hopper looks like its quite similar to Ampere, but the A100 didn't get the double FP32 cores in the SM last time, now they do so there is much more FP32 performance, Hopper H100 has roughly triple Teraflop performance vs A100 in FP32 and FP16

* Apart from the FP32 changes, the only other significant new item is the 4th generation tensor cores, there are a number of improvements to them which improve their performance anywhere from 2x to 6x depending on the type of data processing and maths being done. There is also more of them, now has 576 Tensor cores

* The L2 cache is larger, 60mb

* Nvidia has not yet finalised clock speeds, so performance may change

* The cards use PCIE5, not 4 like some rumors have said

* 80 billion transistors, double Ampere. Built on custom TSMC 4nm process
 
May 23rd for potential announcement:


Better get listing them 30xx series cards before the cheeky offers in MM :D :cry: ;)
 
May 23rd for potential announcement:


Better get listing them 30xx series cards before the cheeky offers in MM :D :cry: ;)
What's that? You to sell me your 3080? I'll offer you £250 and a bag of chips :cry:
 
The design looks able to use faster GDDR7 memory, but might be in the later offerings not initial release.
Maybe if nvidia does go with Hopper they will delay the launch till early next year and go with GDDR7 which would be a substantial power saver over GDDR6X.
 
Or they'll have the non ti series have gddr6x then go with 7 for the ti boards.
I'd think they would go with GDDR7 for the whole stack if it's available as the extra bandwidth and lower power usage would mean nvidia could use a lower bus speed and cheaper power delivery components which would offset the VRAM costs.
 
whats 20w or even 30w power saving when the GPU has 600w TDP
Given the 3070ti uses 70w more than the non ti despite only having 8gb VRAM and a modest clock and Cuda increase I'd bet that the difference would be much higher than 30w especially for cards that use 16gb+.
 
256 bit bus with GDDR7 would have a bandwidth of 1047, with GDDR6X the bandwidth drops to 608 which I'm not sure will be enough for cards like the 4080 that are supposed to be much faster than a 3090, it also gives nvidia the option of 16gb VRAM.

A 4070/4060 could use a 192 bit bus and get a bandwidth of 768 with GDDR7 and come with 12gb VRAM etc.
 
Back
Top Bottom