GPU rumors 2021

[smatovic]

Is it btw possible to use iris GPU in laptop for LC0 ?
...

It "should" be possible, probably you need to install the onnx runtime environment first:

https://duckduckgo.com/?q=Intel+onnx+ru ... =h_&ia=web

I myself have no experience with onnx.

--
Srdja

[smatovic]

AMD's Zen 6-based desktop processors may feature up to 24 cores
https://www.tomshardware.com/pc-compone ... o-24-cores

--
Srdja

[...]
AMD's server Zen 5 "Turin" is already on TSMC 3nm:
[...]

AMD Zen 5 Ryzen and Zen 5 Epyc "Turin" are fabricated on TSMC N4X node, Zen 5c "Turin "Dense" on TSMC N3E .

https://en.wikipedia.org/wiki/Zen_5
https://en.wikipedia.org/wiki/Epyc#Fift ... rin_Dense)

So we can expect with Zen 6 and likely switch to 3nm or 2nm significant core-count increase for the former ones.

https://en.wikipedia.org/wiki/5_nm_process#Nodes
https://en.wikipedia.org/wiki/3_nm_proc ... cess_nodes

Zen 6 initially on track for 2025, now maybe 2026-2027?

TSMC 2nm process to enter volume production in H2 2025.

https://en.wikipedia.org/wiki/2_nm_proc ... cess_nodes

Intel skipped 2nm node and bets on 18A, Samsung is in with SF2.

--
Srdja

[smatovic]

Nvidia announces Blackwell Ultra and Rubin AI chips
https://www.cnbc.com/2025/03/18/nvidia- ... hips-.html

Next gen Nvidia Rubin (2026) will be probably a bigger step for end users in regard of core count cos of smaller fab process (form. called tick-tock cycle).

And Nvidia DGX Spark and Station announced, Nvidia supercomputers in a desktop:

NVIDIA Announces DGX Spark and DGX Station Personal AI Computers
https://nvidianews.nvidia.com/news/nvid ... -computers

--
Srdja

[Vinvin]

Video here :

Nvidia announces Blackwell Ultra and Rubin AI chips
https://www.cnbc.com/2025/03/18/nvidia- ... hips-.html

Next gen Nvidia Rubin (2026) will be probably a bigger step for end users in regard of core count cos of smaller fab process (form. called tick-tock cycle).

And Nvidia DGX Spark and Station announced, Nvidia supercomputers in a desktop:

NVIDIA Announces DGX Spark and DGX Station Personal AI Computers
https://nvidianews.nvidia.com/news/nvid ... -computers

--
Srdja

[towforce]

And Nvidia DGX Spark and Station announced, Nvidia supercomputers in a desktop:

NVIDIA Announces DGX Spark and DGX Station Personal AI Computers
https://nvidianews.nvidia.com/news/nvid ... -computers

My reading of that article is that the Spark will be a petascale computer, which was achieved in the supercomputer world 16 years ago (link). That was too long ago to fairly make the comparison. After all, the Exynos W1000, the SOC in the Galaxy Watch 7, will handily beat any supercomputer from the 1970s.

[smatovic]

[...]
My reading of that article is that the Spark will be a petascale computer, which was achieved in the supercomputer world 16 years ago (link). That was too long ago to fairly make the comparison. After all, the Exynos W1000, the SOC in the Galaxy Watch 7, will handily beat any supercomputer from the 1970s.

Nvidia Spark is former DIGITS:

viewtopic.php?p=975064#p975064
viewtopic.php?p=975073#p975073

and as you mentioned, it depends on the precision of the FLOP, as far as I got it, DIGITS is 1 petaflop in FP4, what is used in nowadays LLMs/Transformer models.

Would be interesting to know what precision Lc0 uses nowadays, in the past it was FP32 and FP16, INT8 quantization did not work well with CNN architecture, Idk about BT-series networks with Transformer/Attention architecture.

--
Srdja

[smatovic]

[...]
as far as I got it, DIGITS is 1 petaflop in FP4, what is used in nowadays LLMs/Transformer models.
[...]

And, the mentioned petaFLOPS achieved by modern GPUs/TPUs are in matrix compute, not vector (SIMD), not scalar (ALU/FPU).

--
Srdja

[smatovic]

[...]
Next gen Nvidia Rubin (2026) will be probably a bigger step for end users in regard of core count cos of smaller fab process (form. called tick-tock cycle).
[...]

...contrary to Intel's former tick-tock cycle Nvidia releases every ~two years* a new microarchitecture with new (external) fab process (die shrink), and in-between an optimization step of microarchitecture and fab process.

https://en.wikipedia.org/wiki/Tick%E2%80%93tock_model

https://en.wikipedia.org/wiki/List_of_e ... hitectures

*Tesla microarchitecture lasted from 2006 to 2010.

--
Srdja

[smatovic]

Apple reportedly readies Baltra processors for AI servers
https://www.tomshardware.com/pc-compone ... ai-servers

If Apple develops its own AI server chip, we might see a decent Mac Pro for running neural networks in future.

--
Srdja

[smatovic]

And here one post for the misc category....

Golden Asus ROG RTX 5090 Astral with Jensen Huang's autograph shown off at GTC
https://www.tomshardware.com/pc-compone ... off-at-gtc

Nvidia Breakfast Bytes are now available at Denny's if you want to experience the 'breakfast of geniuses'
https://www.tomshardware.com/pc-compone ... f-geniuses

This gilded ROG Astral RTX 5090 Dhahab Edition features actual 24k gold and could set you back $10,000
https://www.tomshardware.com/pc-compone ... -usd10-000

...golden GPUs, these days....

--
Srdja