GPU rumors 2021

[smatovic]

[...]

I can agree. I think the vector unit per CPU core is meant to stay, the mat-mul unit per server CPU core is meant to stay, and the NPU in end user CPU is meant to stay.

--
Srdja

[Werewolf]

As a general thought: graphics cards are not ideal for AI

Which is why GPUs are not actually used for that.
"By tradition, Nvidia still calls the H100 a graphics processing unit, but the term is clearly on its last legs: just two out of the 50+ texture processing clusters (TPCs) in the device are actually cable of running vertex, geometry, and pixel shader maths required to render 3D graphics."
https://www.datacenterknowledge.com/dat ... it-matters

If you compare the performance, you'll notice that the H100 rocks in FP16, relevant for AI, while the 4090 beats the H100 in FP32, used for gaming. The FP64 is interesting as well, used e.g. for scientific high precision simulations, where the H100 leaves the 4090 in the dust. Also, the H100 doesn't even support DirectX11/12. Performance in TFLOPS.
4090: FP16: 82.58, FP32: 82.58, FP64: 1.29.
H100: FP16: 248.3, FP32: 62.08, FP64: 31.04.
https://www.techpowerup.com/gpu-specs/g ... 4090.c3889
https://www.techpowerup.com/gpu-specs/h ... 6-gb.c4164

Would be great to see Lc0 nps comparison between the 4090 and the H100

[smatovic]

...Intel's new Xeon 6 with 500W:

https://fudzilla.com/news/59760-intel-f ... p-on-cores

...Nvidia's upcoming RTX 590 with 600W?:

https://www.tomshardware.com/pc-compone ... 12-bit-bus

...AI super chips with up to 1KW:

https://en.wikipedia.org/wiki/Nvidia_DGX#Accelerators
https://en.wikipedia.org/wiki/List_of_A ... nct_series

...and the solution is to build 5x1GW nuclear reactors for data centers:

https://arstechnica.com/tech-policy/202 ... port-says/


Oh boy.

--
Srdja

[smatovic]

[...]
According to WP Blackwell and Zen 5 are fabricated on TSMC 4NP, an advanced 5nm process, maybe Nvidia Rubin and Zen 5+/6 will be bigger steps for us end users in regard of core-count resp. transistor-count (3nm fab process, "tick-tock-cycle").
[...]

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

AMD Turns The Screws With “Turin” Server CPUs
https://www.nextplatform.com/2024/10/10 ... rver-cpus/

[...]
Notable is that Intel uses tiles resp. slices contrary to AMD's chiplets, AFAIK up to four slices are coupled to one single chip, where AMD uses up to eight chiplets.
[...]

And uses 16x8-cores resp. 12x16-cores chiplets with one IO-die and 12 memory channels:

http://www.nextplatform.com/wp-content/ ... -specs.jpg

--
Srdja

[smatovic]

AMD puts grip on GPU server roadmap...

AMD launches Instinct MI325X Accelerator for AI workloads
https://fudzilla.com/news/graphics/5986 ... -workloads

AMD updates Instinct roadmap with MI350 CDNA 4 Series and next-gen MI400 Series
https://fudzilla.com/news/graphics/5986 ... 400-series

AMD Announces Flurry of New Chips
https://www.hpcwire.com/2024/10/10/amd- ... new-chips/

...these new server boxes with 8 server GPU modules, gee, each GPU consumes about ~1 KW power, and computes ~1 petaFLOPS in FP16 (matrix), it was said, to reach human brain scale via artificial neural networks you will need ~ 40 petaFLOPS of compute - interesting times ahead.

--
Srdja

[Werewolf]

AMD's Turin EPYCs look good.

The story I've heard though about the Nvidia 5090 is that you won't get one for 12 months because Elon & co have bought them all...

[towforce]

Apologies if the usage of graphics cards for drawing images (their original purpose) is off-topic, but a new technique has been created by Nvidia which enables moving around in a 3D scene to be a lot more beautiful and realistic:

[smatovic]

Nvidia is building a CPU
https://fudzilla.com/news/60039-nvidia- ... ding-a-cpu

...seems like a logical next step, they already have the Grace CPU for AI running...

But bringing Arm into the PC world is a different story—something that has had mixed success. Without an x86 license, Nvidia could have a significant hurdle to clear to make its CPUs fully compatible with the existing PC game library.

According to DigiTimes, Nvidia could announce its PC CPUs as early as September 2025, with a full launch potentially following in March 2026. Much depends on whether the outfit can get games to work on the thing.

AMD says RDNA 4 GPUs are coming in early 2025 — RX 8000 will deliver ray tracing improvements, AI capabilities
https://www.tomshardware.com/pc-compone ... pabilities

Intel Arc Xe2 Battlemage GPUs rumored to arrive next month — ahead of AMD RDNA 4 and Nvidia Blackwell
https://www.tomshardware.com/pc-compone ... -blackwell

Nvidia GeForce RTX 50-series reportedly launching 'soon' — but all indicators continue to point to a January 2025 reveal
https://www.tomshardware.com/pc-compone ... 025-reveal

We can probably expect the RTX 50-series to still be at least partially revealed at CES 2025 — CEO Jensen is on tap for the keynote, after all. Will that be an official launch, or an announcement of an upcoming launch like last year's 40-series Super cards? Our money would be on the latter, meaning a late January release.

In the meantime, there's no reason for Nvidia to rush things. It already dominates the GPU market and continues to sell GPUs. The RTX 4090 remains a prime example of this, an extremely expensive card that has recently increased in price due to diminishing supply coupled with a still-high demand from AI customers and perhaps even a few remaining 'normal' consumers.

--
Srdja

[towforce]

Nvidia is building a CPU
https://fudzilla.com/news/60039-nvidia- ... ding-a-cpu

...seems like a logical next step, they already have the Grace CPU for AI running...

But bringing Arm into the PC world is a different story—something that has had mixed success. Without an x86 license, Nvidia could have a significant hurdle to clear to make its CPUs fully compatible with the existing PC game library.

According to DigiTimes, Nvidia could announce its PC CPUs as early as September 2025, with a full launch potentially following in March 2026. Much depends on whether the outfit can get games to work on the thing.

This is a problem: people are not willing to spend time learning how to build out better solutions to their computing needs.

* for most people, their best choice of computing device would be a Chromebook - but most people pay more than they need to for a product that doesn't meet their needs as well

* Linux has the Wine compatibility layer for running Windows programs - but it requires the device to be using an x86 CPU

* There are flavours of Linux which are similar to Windows in look and feel, but I'm not sure how far you can go before Microsoft inflicts legal difficulties on you

* related to this: you could simulate Windows rather than have a compatibility layer (there are lots of precedents for simulating an entire computer, including the OS, on another device with a completely different CPU), but there are pitfalls: (1) copyright claims from Microsoft (2) the simulation might run slower than your old Windows with x86 device, which is not what you want when you're running games. If the new CPU is a lot faster than the x86 it's replacing, though, then maybe we could live with this problem.

[towforce]

For running x86 programs on ARM CPUs, there are QEMU and KVM, of course (IIRC QEMU is the emulator that enabled the project that made the great dedicated chess computers of old available under Windows).