Val Kantchev

👨🏻‍💻 Software Developer 🔰 Investor 🔥@daylohq @avblocks

[SpaceX @SpaceX]

Teams are go for launch with a $135 price per share for the SpaceX IPO → https://t.co/HlsF4GxRbI

[Steve Hilton @SteveHiltonx]

🚨RACE CALLED—Change will be on the ballot this November!

Thank you to every single one of you in this great state. I am running for governor to be of service to you—to make California the greatest place in the world and to restore the California Dream.

Change is Coming ☀️👊

[EarthCam @EarthCam]

Months of work, condensed into seconds. Watch the transformation of the Lincoln Memorial Reflecting Pool as crews drained, repainted, and refilled one of America's most iconic landmarks ahead of a busy summer season in Washington, D.C.

[Val Kantchev @valkantchev]

FINRA Rule 4210 that removes the pattern day trader designation and $25,000 minimum equity requirement took effect on June 4, 2026.

The pattern day trader designation is going away

- Old rule: Clients who placed four or more day trades within five business days were designated as "pattern day traders" (PDT) and subject to enhanced margin requirements.

- New rule: The PDT designation is eliminated. Broker-dealers are no longer required to track the number of day trades placed in a margin account or apply unique margin requirements to those designated as pattern day traders.

[Val Kantchev @valkantchev]

@ivanfioravanti At half the cost NVIDIA RTX 4090 will give you close to 4000 t/s prompt and 120 t/s. Just buy a GPU. Still the best solution.

[Val Kantchev @valkantchev]

AMD basically lost the AI race... So now they are like: but but actually we have the most affordable gaming GPUs.

Who cares?

[Val Kantchev @valkantchev]

@AMD @AMDRyzen @jackhuynh This is complete and utter marketing BS. Get the memory bandwidth up to 1000 GB/s and fix the despicable horrible ROCm software which is still 3 times slower than CUDA on comparable hardware.

[Val Kantchev @valkantchev]

@AMD 50 tokens / second is ridiculously slow. Not a breakthrough.

[Val Kantchev @valkantchev]

@neogoose_btw You can also do it with forward slashes

[Val Kantchev @valkantchev]

@NVIDIARTXSpark Slow UMA machines including also this one are the biggest scam in the world right now. DO NOT BUY!

[Val Kantchev @valkantchev]

Your best and cheapest option for running local LLM is actually NVIDIA RTX 3090 or 3090ti - 1000 GB/s (1 TB/s) memory bandwidth.

No other consumer UMA machine currently ships with 1000 GB/s (1 TB/s) memory bandwidth as of mid-2026.

[Val Kantchev @valkantchev]

If you want to use for local coding:

DO NOT BUY:
M4 Mac Mini, 120 GB/s
AMD Strix Halo, 256 GB/s
Nvidia DGX Spark, 273 GB/s
M5 32 Core MacBook Pro, 460 GB/s
Intel Arc Pro B70 GPU, 608 GB/s
M5 40 Core MacBook Pro, 614 GB/s

BUY:
Nvidia RTX 3090 GPU, 936 GB/s
Nvidia RTX 5090 GPU, 1,792 GB/s

[Val Kantchev @valkantchev]

@The_Only_Signal Anything below 1000 GB/s will feel super slow for agent coding.

[Val Kantchev @valkantchev]

@The_Only_Signal RTX 4090 1,792 GB/s same as 5090 but 24GB instead of 32GB

[Val Kantchev @valkantchev]

@cyrilismyname @AMDRyzen 215–273 GB/s. For comparison a RTX 3090 Ti is 1000 GB/s - 4x faster.

[Val Kantchev @valkantchev]

Actual Benchmarks (Ryzen AI Max+ 395 / 128GB, similar to 400 series). From detailed context-scaling tests on real hardware:
• Good MoE models (e.g. Nemotron-3-Super-120B-A12B Q4, Qwen3.5 122B-A10B, gpt-oss-120B):
• Prefill speed at ~100k+ context: ~120–230 t/s (often drops toward the higher end of the range as KV cache fills).

• Time for 100k tokens:
• At 200 t/s average → ~8.3 minutes.
• At 150 t/s → ~11 minutes.
• At 120 t/s → ~14 minutes.

Short-context is much faster (often 300–700+ t/s), but it degrades significantly as context grows due to KV cache pressure and the ~215–273 GB/s bandwidth limit.

[Val Kantchev @valkantchev]

@AMDRyzen So same slow memory. Coding requires typical contexts of 60k to 100k tokens.

Wait Time for 100k tokens( this is per Grok):

• At 200 t/s average → ~8.3 minutes.
• At 150 t/s → ~11 minutes.
• At 120 t/s → ~14 minutes.

Do not waste your time and money on this machine.

[Val Kantchev @valkantchev]

@stevibe All Unified Memory machines are USELESS for coding because the memory is too slow. Do not buy!

[Elon Musk @elonmusk]

T-zero is currently 5:30 Texas time