Discussion Nvidia Blackwell in Q1-2025

[igor_kavinski]

Probably installed with the Uno Reverse card, ASUS style.

Must've been unusually high quality to survive reverse polarity and not explode immediately after being turned on.

 

[Heartbreaker]

Ooh, the GB10 chip that's in the DGX Spark is claimed to be using TSMC N3 for both the CPU and GPU. Also the GPU die is derived from GB100 and not the gaming chips.

Still very lacking in the memory bandwidth tho - 301 GB/sec.

https://videocardz.com/newz/nvidia-details-gb10-blackwell-superchip-successful-collaboration-between-nvidia-and-mediatek

It will be interesting see a matchup with Strix Halo.

 

[jpiniero]

It will be interesting see a matchup with Strix Halo.

It says the TDP is 140 W so not really a laptop chip. (Course tell that to Intel)

 

[Heartbreaker]

It says the TDP is 140 W so not really a laptop chip. (Course tell that to Intel)

There are lots of strix Halo boxes using about that much power.

Here is a guy comparing 3 Halo boxes where he says it can use a peak of 120 watts, peak 140.

 

[fastandfurious6]

It says the TDP is 140 W so not really a laptop chip. (Course tell that to Intel)

in 2025 even 13'' thin laptops can do 140W total sustained without throttling

16'' 2kg class can do upwards of 200W

 

[Fallen Kell]

It will be roughly a Nvidia L4 combined with 20 core ARM, just with the latest gen 5 cuda cores instead of gen 4. The L4 does 30 TFLOPs FP32, while the GB10 is said to have 31 TFlops FP32 and they both have the same memory bandwidth, just that the GB10 may have up to 128GB unified memory, vs the L4 having 24GB dedicated memory (the unified memory is shared for both the ARM CPU cores and the GPU).

 

[Mopetar]

in 2025 even 13'' thin laptops can do 140W total sustained without throttling

16'' 2kg class can do upwards of 200W

Doesn't sound like anything I'd want to put on my lap. 😛

It also sounds like the battery should be specified in Wm (watt minutes) instead of Wh (watt hours).

 

[jpiniero]

It will be roughly a Nvidia L4 combined with 20 core ARM, just with the latest gen 5 cuda cores instead of gen 4. The L4 does 30 TFLOPs FP32, while the GB10 is said to have 31 TFlops FP32 and they both have the same memory bandwidth, just that the GB10 may have up to 128GB unified memory, vs the L4 having 24GB dedicated memory (the unified memory is shared for both the ARM CPU cores and the GPU).

The raw specs are similar to the 5070, even though they claimed it was using the GB100 version.

Without the memory bandwidth of course.

 

[regen1]

RTX Hair in Indiana Jones und der Große Kreis

ComputerBase hat RTX Hair in Indiana Jones und der Große Kreis im Test mit Benchmarks, Analysen und einem Video.

www.computerbase.de

ComputerBase' take on RTX Hair(in Indian Jones)

 

[SolidQ]

ComputerBase' take on RTX Hair

looks like Sh*t with on

 

[Tup3x]

looks like Sh*t with on

I only see video compression artifacts that likely smooth out the finer detail. Without it hair seems to be less dense so compression doesn't hurt as much.

 

[igor_kavinski]

A company trying to find ways to justify the high price tag of its overpriced unreliable GPU with buggy drivers. Yeah guys. Keep enjoying staring at the hair instead of actually playing the game and oh, don't worry if you smell something weird. It's part of the experience. If something burns, take it as an opportunity to reward this company for making more innovative products and buy the GPU again. Remember, you need to trust them and keep letting them try, try again. One day they will make a true banger of a product with zero issues!

 

[Mopetar]

looks like Sh*t with on

NVidia HairSortaWorks.

 

[Ranulf]

This all seems vaguely familar, some sort of deja vu but I can't quite put my finger on it.

Spoiler: Bonus reference

 

[ToTTenTranz]

looks like Sh*t with on

I'm actually playing Indiana Jones maxed out on a RTX 5090 and enabling RTX Hair didn't make it look like the picture on the left at all.
Though I have to say I didn't notice any substantial difference in how the hair looks, either.


If anything, the biggest change I noticed on the RTX Hair update is how it brought a bug in the denoising pipeline that creates noisy backgrounds whenever the game enters a cutscene. Now that's super frustrating.

A company trying to find ways to justify the high price tag of its overpriced unreliable GPU with buggy drivers. Yeah guys. Keep enjoying staring at the hair instead of actually playing the game and oh, don't worry if you smell something weird. It's part of the experience. If something burns, take it as an opportunity to reward this company for making more innovative products and buy the GPU again. Remember, you need to trust them and keep letting them try, try again. One day they will make a true banger of a product with zero issues!

Agreed. The only reason I own a 5090 is because its purchase is justifiable as AI processing hardware. There's no such alternative from AMD, unfortunately.

 

[igor_kavinski]

Agreed. The only reason I own a 5090 is because its purchase is justifiable as AI processing hardware. There's no such alternative from AMD, unfortunately.

That's true. AMD can't even be bothered to do Vulkan properly for RDNA4. It's not detected in LM Studio yet both Intel ARC and Nvidia cards (even the 1080 Ti) have no issue being detected by LM Studio.

 

[Saylick]

I didn’t see a Rubin thread, so I’ll post this here:

NVIDIA Rubin CPX Accelerates Inference Performance and Efficiency for 1M+ Token Context Workloads | NVIDIA Technical Blog

Inference has emerged as the new frontier of complexity in AI. Modern models are evolving into agentic systems capable of multi-step reasoning, persistent memory, and long-horizon context—enabling…

developer.nvidia.com

It looks like Nvidia are pre-announcing what I’m guessing is the professional version of what will likely be the same die as the RTX 6090, especially since it employs GDDR7 memory. I also count what appears to be a 512-bit memory interface:

 

[Kepler_L2]

I didn’t see a Rubin thread, so I’ll post this here:

NVIDIA Rubin CPX Accelerates Inference Performance and Efficiency for 1M+ Token Context Workloads | NVIDIA Technical Blog

Inference has emerged as the new frontier of complexity in AI. Modern models are evolving into agentic systems capable of multi-step reasoning, persistent memory, and long-horizon context—enabling…

developer.nvidia.com

It looks like Nvidia are pre-announcing what I’m guessing is the professional version of what will likely be the same die as the RTX 6090, especially since it employs GDDR7 memory. I also count what appears to be a 512-bit memory interface:
View attachment 129892

288 SMs? Looks like an additional row on each side vs GB202

 

[ToTTenTranz]

288 SMs? Looks like an additional row on each side vs GB202View attachment 129893

So this is a rasterizer going close to reticle size limit like GB202, but on N3P?

 

[jpiniero]

I was going to say that nVidia is claiming it's a new product... but I can see the resemblance to Blackwell Gaming.

So this is a rasterizer going close to reticle size limit like GB202, but on N3P?

The price whining is gonna be fun. Gotta love nFlation.

 

[adroc_thurston]

So this is a rasterizer going close to reticle size limit like GB202, but on N3P?

Yeah.

 

[Saylick]

288 SMs? Looks like an additional row on each side vs GB202View attachment 129893

So we're likely looking at a ~40% to 60% uplift gen-on-gen. Yes, I understand that there likely will be clock speed improvements due to N3 FinFlex, but performance doesn't also scale perfectly with SMs either.

 

[adroc_thurston]

but performance doesn't also scale perfectly with SMs either.

That's been especially the problem with lotsa SM parts from NV.
They gotta work that out or this will be the most useless gaming toy in eons.

 

[Saylick]

That's been especially the problem with lotsa SM parts from NV.
They gotta work that out or this will be the most useless gaming toy in eons.

I bet it won't prevent it from being the best selling Nvidia flagship, courtesy of some new exclusive feature that Nvidia's developers are secretly inserting into future games as we speak.

 

[adroc_thurston]

I bet it won't prevent it from being the best selling Nvidia flagship, courtesy of some new exclusive feature that Nvidia's developers are secretly inserting into future games as we speak.

Features are whatever, just that SM scaling might be bad outside of doing GEMM.