Discussion Nvidia Blackwell in Q1-2025

[adroc_thurston]

Deepseek bypasses CUDA to use PTX close to metal programming that makes it far more efficient. Thus needing less HW to do the same thing. Thats what hit Nvidias stock.

DeepSeek's AI breakthrough bypasses industry-standard CUDA for some functions, uses Nvidia's assembly-like PTX programming instead

Dramatic optimizations do not come easy.

www.tomshardware.com

PTX is Nvidia low-level IR ISA, lmao. SASS is like a level below, but it's unreadable garbage.
Granted, nothing stops anyone from writing amdgcn kernels either, so.

 

[SolidQ]

IgorLabs seems post early RTX 5080 review, and then took it down fast.
People who read there, was like 7% over 4080s in 4k

 

[adroc_thurston]

People who read there, was like 7% over 4080s in 4k

Lmao rip.
Wallahi may AMD plaster some GDDR7 onto Navi48 for kicks.

 

[Kepler_L2]

AMD pushing N48 to 400W as we speak

 

[Thunder 57]

Lmao rip.
Wallahi may AMD plaster some GDDR7 onto Navi48 for kicks.

Not gonna happen.

 

[branch_suggestion]

AMD pushing N48 to 400W as we speak View attachment 115817

Holy crap Blackwell sucks. 0% PPC gain for a new uArch with new shiny memory is hilarious.

 

[adroc_thurston]

Not gonna happen.

Wdym not gonna happen, it'll take at best a stepping.

Also holy crap Blackwell sucks. 0% PPC gain for a new uArch with new shiny memory is hilarious.

It makes RDNA3 seem competent and that's a dire statement.
At least N33 had the decency to chop 33mm^2 iso config.

 

[branch_suggestion]

Lmao rip.
Wallahi may AMD plaster some GDDR7 onto Navi48 for kicks.

Hopefully just memory oc and power limits will be enough.
Commodity 20Gbps could maybe hit 24Gbps on some samples.

 

[adroc_thurston]

Hopefully just memory oc and power limits will be enough.
Commodity 20Gbps could maybe hit 24Gbps on some samples.

24G is 'samping' status aka doesn't exist.
GDDR7, OTOH.

 

[SolidQ]

AMD pushing N48 to 400W as we speak

Wonder what happens slap GDDR7 and push it to 400W
Interesting how close UDNA to 5090.
They never miss an opportunity to miss an opportunity (c)

 

[branch_suggestion]

It makes RDNA3 seem competent and that's a dire statement.
At least N33 had the decency to chop 33mm^2 iso config.

When was the last time NV dropped the ball this badly?
Fermi1 was quick but broken, bumpgate was prior but that was just bad engineering... Rankine?

 

[adroc_thurston]

When was the last time NV dropped the ball this badly?
Fermi1 was quick but broken, bumpgate was prior but that was just bad engineering... Rankine?

Tesla2 was fodderized by the baby known as RV770 and this is the closest we'll ever get to that.
Pottery, rhymes, George Lucas etc.

Just genuinely baffled at NV delivering a new SM that's just like, nothing. Like they could've plastered the refreshed RTRT hardware onto Ada and this would be the result.

 

[moonbogg]

Just saw the supposed pricing for MSI 5090s here and its downright bizzarre. Like only one model, Ventus, at MSRP, as expected. But the remaining 4 models, all straight up at 3000 EUROs or more (obviously including VAT). So no 200 to 600 EUROs about MSRP options, like in the past, straight up 600 for second cheapest variant. I am speechless.

Scalpers have been camping out for days already and have already listed cards for resale on ebay before even getting the cards. It actually makes sense for manufacturers to realize this and simply charge a price that's closer to what gamers are willing to actually pay. The cards need to be around $3500+ to even things out on the demand side IMO.

 

[SolidQ]

When was the last time NV dropped the ball this badly?

feels like story repeated and it's similar GeForce FX which also 5th series

 

[branch_suggestion]

Tesla2 was fodderized by the baby known as RV770 and this is the closest we'll ever get to that.
Pottery, rhymes, George Lucas etc.

Just genuinely baffled at NV delivering a new SM that's just like, nothing. Like they could've plastered the refreshed RTRT hardware onto Ada and this would be the result.

Yeah I guess their last 512b part is the best analogue.
GT200(b) was a slow, fat thing milking the last of G80's glory.

 

[adroc_thurston]

Yeah I guess their last 512b part is the best analogue.
GT200(b) was a slow, fat thing milking the last of G80's glory.

It was an actual incremental uarch update, but it also got stomped by ATi's ALUspammy wunderwaffen.
This time, it's cache magicks that haunt Nvidia like witches from PMMM.

 

[SolidQ]

add games screens

Spoiler

 

[adroc_thurston]

add games screens

Spoiler

Lmao it's literally the bandwidth uptick alone.
0% from the uarch.

 

[branch_suggestion]

It was an actual incremental uarch update, but it also got stomped by ATi's ALUspammy wunderwaffen.
This time, it's cache magicks that haunt Nvidia like witches from PMMM.

The sad part is the consumer loses due to the ATi curse.

 

[branch_suggestion]

I think we know why there is no GB204 now.
Would've been the most irrelevant die ever made.

 

[poke01]

this is the worst GPU arch EVER. NV is bleeding design talent, simply put.

This makes RDNA3 look like a okay gen

 

[adroc_thurston]

NV is bleeding design talent, simply put.

Hell no, they have a talent retention cheat code (RSUs worth millions of dorra). Some microarches are just duds.

this is the worst GPU arch EVER

There have been worse. Even from Nvidia.

 

[poke01]

There have been worse. Even from Nvidia.

It clocks low, is more power hungry and uses GDDR7 and yet has no uplifts. Now we know why the price is the same, its the same perf

 

[adroc_thurston]

It clocks low

Not really? Same as Ada more or less.

is more power hungry

Not really?

and uses GDDR7 and yet has no uplifts

Well GDDR7 is the uplift, see 5080.

Now we know why the price is the same, its the same perf

$50 off on 5070/5070ti at least.

 

[Kepler_L2]

Regressions in L1 bandwidth and L2/mem latency: