GPU rumors 2021

[smatovic]

....

Yes, 7nm with DUV lithography, there is another thread:

China boosts in silicon...
viewtopic.php?f=7&t=72566&p=954071#p954071
viewtopic.php?f=7&t=72566

--
Srdja

[smatovic]

....on Chinese GPUs:

Moore Threads
https://en.wikipedia.org/wiki/Moore_Threads

Biren Technology
https://en.wikipedia.org/wiki/Biren_Technology

Zhaoxin
https://en.wikipedia.org/wiki/Zhaoxin#Discrete_GPU

--
Srdja

[towforce]

....

Yes, 7nm with DUV lithography, there is another thread:

China boosts in silicon...
viewtopic.php?f=7&t=72566&p=954071#p954071
viewtopic.php?f=7&t=72566

--
Srdja

Thanks for that link: I rarely look at the Programming And Technical Discussions forum, even though I know that this choice causes me to sometimes miss interesting discussions.

[smatovic]

These Chinese makers, micro-soldering...

Chinese workshops recondition Nvidia's old flagship gaming GPU for AI — RTX 2080 Ti upgraded to 22GB for $499
https://www.tomshardware.com/pc-compone ... -dollar499

Besides selling the altered graphics card, the Goofish seller also offers an upgrade service for $120, which includes labor and materials, for GeForce RTX 2080 Ti owners who want to send in their graphics card for the 22GB modification.

Modder Upgrades A GeForce RTX 2080 Ti 11GB To 44GB Of VRAM But There's A Caveat
https://hothardware.com/news/modder-upg ... s-a-caveat

--
Srdja

[smatovic]

Draude-people know, size matters

NVIDIA Blackwell Architecture and B200/B100 Accelerators Announced: Going Bigger With Smaller Data
https://www.anandtech.com/show/21310/nv ... aller-data

Nvidia officially unveils Blackwell GPU architecture at GTC
https://www.fudzilla.com/news/graphics/ ... ure-at-gtc



it's about the server class, consumer Blackwell might come in 2025.

Nvidia uses advanced TSMC 4N node, so they increased transistor count (x2.5) by using two tiles, multi-chip, AI throughput increased by FP4 and FP6 Transformer Engines.

--
Srdja

[towforce]

NVIDIA Blackwell Architecture and B200/B100 Accelerators Announced: Going Bigger With Smaller Data...

Using this technology, they also announced the DGX B200 supercomputer, which will be at the exascale at the reduced precision used for machine learning. It will be able to train trillion parameter NN models, which should result in a reasonably good knowledge of chess (or other subject matters).

"Nvidia tells us that one of the standout features of the GB200 Superchip is its ability to deliver up to 30 times the performance of Nvidia’s current leading H100 Tensor Core GPU for large language model inference tasks. This remarkable improvement pushes the boundaries of AI supercomputing and will enable the more efficient development and deployment of more sophisticated AI models." - link.

[Modern Times]

This remarkable improvement pushes the boundaries of AI supercomputing and will enable the more efficient development and deployment of more sophisticated AI models."[/i] - link.

Skynet is not far away.

[smatovic]

Straggler...

Amazon Graviton 4 with 96 cores and SVE2 vector-unit is available in the cloud:
https://en.wikipedia.org/wiki/AWS_Graviton#Graviton4
[...]

Looked up, now I am not sure if Graviton 4 is SVE or SVE2, Arm v9 should be v2 IIRC, but according to Wikichip, both Graviton 3 and 4 are 2x256b SVE:

Amazon Debuts 4th Gen Graviton
https://fuse.wikichip.org/news/7633/ama ... -graviton/

AWS R8g Instances: Memory-Optimized Instances Powered by AWS Graviton4 Processors
https://www.linkedin.com/pulse/aws-r8g- ... ukwu-ujzqf

--
Srdja

[smatovic]

- CPU-GPU coherent memory
With Nvidia moving into the CPU realm we have a tight coupling of CPU-GPU arch
for HPC incoming. IBM dropped NVLink support in their POWER10 series, so all
HPC-GPU vendors will come up with a solution for coherent memory between CPU
and GPU, maybe an open standard like CXL over PCIe, maybe something proprietary
like NVLink and Infinity Fabric, unknown if and how this descents to the gamer
gpu market.

--
Srdja

...now with AI as silicon arms-race, we see Amazon, Microsoft and Google as cloud provider with own, ARM based, CPUs and own AI inference/training chips rising, a coupling of cloud provider+CPU+AI chips, interesting times.

--
Srdja

[Werewolf]

For Lc0 it looks like the 4090 will still be king for the next 12 months then.