Pierre

I build SaaS

[Pierre @serphacker]

True. Angular is an opinionated framework where everything works out of the box. No library fatigue, solid architecture foundation.

Sit down and code.

[Jordan Leeper @const_let_var]

@mjackson @Enea_Jahollari Modern Angular absolutely gets it right. No decision fatigue on bundler, testing framework, routing, built in SSR. Everything works and it’s fast. Devex is fantastic. Typescript built in and not negotiable. When they migrated from webpack to Vite, I had absolutely 0 work to do.

[Pierre @serphacker]

@pupposandro in my case improvement would be minor for the price (5000$ ?), I already own one I bought for 2300€.

I prefere to wait for DDR5/DDR6 with at least 600Gb/s bandwidth and 256GB RAM (exactly 2x specs of a 395 strix). That's not crazy considering we already have DRR7 at 1.7 Tb/s.

[Pierre @serphacker]

@vision_ia Rien de révolutionnaire, le Apple M5 pro fait déjà mieux en IA, le strix halo (2025) fait pareil (ou à peine moins bien). C'est un DGX spark au format portable, et le DGX spark n'a déjà rien de ouf, alors rajoute les contraintes d'un portable... C'est un gros flop leur truc.

[Pierre @serphacker]

@Beunwa J'utilise depuis quelques années, fonctionne bien, thème/launcher et applis de base (gallery) un peu austère/simpliste mais pas génant (on peut changer). J'apprécie d'avoir un Android debloated par défaut. Preque aussi simple qu'un Android si tu utilises le playstore (j'évite).

[Pierre @serphacker]

@dsampaolo Je préfère Fedora pour les laptops/desktops. Debian stable je garde ça pour les serveurs.

Là j'ai un laptop sorti en début 2025 (HP g1a 395+), le driver de la caméra a été mergé dans kernel 7.0 il y a seulement quelques semaines...

[Pierre @serphacker]

@pupposandro So their "new era of PC" is a slightly better 2025 strix halo laptop... That pretty much sums up the state of consumer hardware... Maybe we will have more interesting stuff from Intel with Crescent Island AI GPU announcement

[Pierre @serphacker]

@digitalix Unfortunately it will have low memory speed, it will be slightly better than a strix halo, like a DGX Spark. A M5 is probably faster. Not the hardware we are waiting for...

[Pierre @serphacker]

@0xSero Gemma 4 24B (MoE), tends to lose its mind at 50k/60k+ context (even at Q8). I prefered Gemma 4 too first (less thinking, concise response), but now I only use Qwen3.6, it's just better (and it works with big context). Don't know about 31B dense, but I think it has the same issue

[Pierre @serphacker]

@pupposandro In 2/3 years, we will probably have the same performance for 2x, 3x (10x ?) cheaper. 6000 Pro is very tempting, but spending $130k would be a very bad short/mid-term investment. Don't FOMO. There are new models and software improvements every week. Hardware will follow. Hodl.

[Pierre @serphacker]

Models are evolving too fast anyway, the gap between small and big models is narrowing.

Difference between a Qwen 30B MoE (run on small GPUs) and a MiniMax 230B MoE (require 25k$ dual RTX 6000) is not worth the price !

Of course the later is better. Worth 25k$ of hardware ? No.

[Pierre @serphacker]

Worst time to buy expensive hardware for AI.

There will be much more options in the upcoming years: DDR6, new GPUs (a lot of teasing recently), NPU/TPU, AMD Medusa Lake, Intel/NVidia Serpent Lake...

[Espen JD @Snixtp]

The RTX Pro 6000 I got for $9500 in April is now $12k (no tax)

[Pierre @serphacker]

@pupposandro Fun build and it's unmatched in raw speed/RAM per $ at 5k$

For me, not worth the power usage (1500w-1800w gpu alone), the heat, the noise, and above all, the risk of 4 USED gpu. I don't want this at home...

There are no alternatives at this price, though. 6000 is 300w but 10k$

[Pierre @serphacker]

@YannDecoopman Pour de l'IA local ?

[Pierre @serphacker]

So much knowledge in these 3 classics. Very good pick. A must-read for any programmer.

[Matt Pocock @mattpocockuk]

Pump the classics into my veins

[Pierre @serphacker]

@txtnurcahyo @sudoingX Can you run dense 27B ?

[Pierre @serphacker]

@pupposandro @LottoLabs (I own a strix 128G)

[Pierre @serphacker]

@pupposandro @LottoLabs If you already have 3090s, not worth it (unless for energy bill).

The best model you can run on 2x3090 (Qwen 3.6 27B) is better than anything you could run on your 96G.

You can't run dense model on strix (slow af). Strix users need to wait for a good 80B-120B MoE beating 27B.

[David Fowler @davidfowl]

https://t.co/Y4ALRmrJCN

[Sandro @pupposandro]

2.5x faster than llama.cpp on Strix Halo.

We just shipped DFlash + PFlash for the AMD Ryzen AI MAX+ 395 iGPU (gfx1151, 128 GiB unified memory).

Qwen3.6-27B Q4_K_M, end-to-end on the same silicon:

▸ Decode: 26.85 tok/s, 2.23x faster (DFlash + DDTree, budget 22)
▸ Prefill 16K: 20.2s, 3.05x faster (PFlash)
▸ Wall clock, 16K prompt + 1K gen: 58s vs 147s

~100 GiB still free in the box. 122B and 139B MoE class is next.

Massive thanks to @smpurkis0 for the contribution 🙏