Pinned Tweet
Do not pay $12,000 for 2 x DGX Sparks.
For the same price you can buy 4 x 170HXs and a computer to run them in, for FAR better performance.
Should I post complete builds at various price points?
For the same price you can buy 4 x 170HXs and a computer to run them in, for FAR better performance.
Should I post complete builds at various price points?
46
14
311
27,992
Morrowmake retweeted
⚔️DGX Spark vs CMP 170HX: Deepseek Edition⚔️
Four 170HXs and 2 DGX Sparks cost the same. But the performance gap is huge.
The 170HXs are 2-4x faster across the board. 🔥
If you’re comfortable with the tradeoffs, the 170HXs are still a good option.
Four 170HXs and 2 DGX Sparks cost the same. But the performance gap is huge.
The 170HXs are 2-4x faster across the board. 🔥
If you’re comfortable with the tradeoffs, the 170HXs are still a good option.
RELEASE: Deepseek 4.1 Flash 1017 on 4 CMP 170HX 🔥📢 ⚡️
1 user: 162t/s (prose), 287t/s (code), 319t/s (structured)
8 users: 432t/s (prose), 728t/s (code), 765t/s (structured)
Cold prefill: 3800t/s
Full recipe link below 👇
4
2
27
1,624
Is there interest in a pp4 version of the Deepseek v4.1 Flash 1710 recipe? Let me know.
@Morrowmake 非常感谢！以后会有类似glm5.3 flash那样的一个PP模式版本么？
7
0
8
687
⚔️DGX Spark vs CMP 170HX: Deepseek Edition⚔️
Four 170HXs and 2 DGX Sparks cost the same. But the performance gap is huge.
The 170HXs are 2-4x faster across the board. 🔥
If you’re comfortable with the tradeoffs, the 170HXs are still a good option.
Four 170HXs and 2 DGX Sparks cost the same. But the performance gap is huge.
The 170HXs are 2-4x faster across the board. 🔥
If you’re comfortable with the tradeoffs, the 170HXs are still a good option.
RELEASE: Deepseek 4.1 Flash 1017 on 4 CMP 170HX 🔥📢 ⚡️
1 user: 162t/s (prose), 287t/s (code), 319t/s (structured)
8 users: 432t/s (prose), 728t/s (code), 765t/s (structured)
Cold prefill: 3800t/s
Full recipe link below 👇
4
2
27
1,624
Morrowmake retweeted
RELEASE: Deepseek 4.1 Flash 1017 on 4 CMP 170HX 🔥📢 ⚡️
1 user: 162t/s (prose), 287t/s (code), 319t/s (structured)
8 users: 432t/s (prose), 728t/s (code), 765t/s (structured)
Cold prefill: 3800t/s
Full recipe link below 👇
1 user: 162t/s (prose), 287t/s (code), 319t/s (structured)
8 users: 432t/s (prose), 728t/s (code), 765t/s (structured)
Cold prefill: 3800t/s
Full recipe link below 👇
12
4
60
5,207
RELEASE: Deepseek 4.1 Flash 1017 on 4 CMP 170HX 🔥📢 ⚡️
1 user: 162t/s (prose), 287t/s (code), 319t/s (structured)
8 users: 432t/s (prose), 728t/s (code), 765t/s (structured)
Cold prefill: 3800t/s
Full recipe link below 👇
1 user: 162t/s (prose), 287t/s (code), 319t/s (structured)
8 users: 432t/s (prose), 728t/s (code), 765t/s (structured)
Cold prefill: 3800t/s
Full recipe link below 👇
12
4
60
5,207
Doomers keep saying the AI induced collapse is right around the corner.
But so far all AI has done is exponentially increase the amount of stuff being created, requiring more human oversight and direction than ever.
There is now more demand for human expertise than before AI.
But so far all AI has done is exponentially increase the amount of stuff being created, requiring more human oversight and direction than ever.
There is now more demand for human expertise than before AI.
Wake up.
Every profession deemed white-collar, the middle class dream, is about to be decimated by AI, at a much faster pace than anticipated, with monstrous social consequences.
2
0
5
453
Dear @NVIDIAAI @JensenHuang: the CMP 170HX is a GA100, the same chip as the A100, with ~1.6 TB/s of HBM2e. It's deliberately handicapped in firmware.
We're in the middle of a compute shortage. Flip the switch, unlock them officially, and thousands of cards go back to work.
Do it for the people.
We're in the middle of a compute shortage. Flip the switch, unlock them officially, and thousands of cards go back to work.
Do it for the people.
Dear @nvidia, a friendly ask from a guy running 4 of your 3090s and 1 of 3080 Ti.
The 30 series is 6 years old. You dont make these cards anymore. Maybe its time to hand them over to the community completely.
My post about unlocking the 3090 got 270K views in a day, and the replies are the real story. People asking about the 4090, 3080, 3060 and 2080 Ti. People asking how to do it on Windows. Others listing what the community already patches on these cards: the NVENC session limit, P2P between GPUs.
Thousands of people are digging through your old cards with kernel modules, 1 setting at a time, because they love this hardware and want everything it can give.
Publish what is limited on consumer Ampere and Turing. Ship an official driver switch for older generations. Or just open it up.
I understand the business side. Some people might keep their old cards longer instead of buying new ones. But it would be a huge gesture. You talk about local AI on stage, you ship open models, you sell DGX Spark for home. This would show the care goes all the way down to a used 3090 under someone’s desk.
Give them to us. We will do the rest: benchmarks, guides, models tuned for these cards.
@NVIDIAAI
7
5
60
6,285
I would never have said this even 3 months ago, but the M5 Ultra is now good value for inference.
If you're only interested in smaller models then 2 x 16GB v100s is good too.
But honestly at current pricing not much is appealing to me.
If you're only interested in smaller models then 2 x 16GB v100s is good too.
But honestly at current pricing not much is appealing to me.
3
0
21
2,541
Morrowmake retweeted
@ItsmeAjayKVe 64GB Ampere class NVIDIA cards. PCIE 2.0 x4 link sounds limiting but we’re getting great speeds with pipeline parallelism anyways. 6,000 tok/s prefill and 200+ tok/s decode is pretty nice.
github.com/AlexBrandes/cm…
Or you can mod them to x16 and do tensor parallelism. Suggest checking out @Morrowmake’s work on that.
github.com/AlexBrandes/cm…
Or you can mod them to x16 and do tensor parallelism. Suggest checking out @Morrowmake’s work on that.
0
2
10
411
Extremely unfortunate that EXL3 basically halves the speed of the CMP 170HX. I've toyed with the idea of trying to create a quantisation that's better suited to these GPUs while retaining the benefits of EXL3. It might be achievable but it's a big undertaking.
8
0
24
1,542
Is this why 3090 prices have been flying?
Expecting the low end of the market to reprice higher again. The entry level AMD cards, B70s and 170HXs likely see a bump.
Expecting the low end of the market to reprice higher again. The entry level AMD cards, B70s and 170HXs likely see a bump.
My 3090s just got 1.9x FASTER at BF16 tensor math!!! Same cards, same driver, same 300W, one Linux kernel module.
bf16 tensor pipe: 82 -> 155 TFLOPS
cuBLAS BF16 GEMM 8192: 64 -> 99 TFLOPS
TF32: 41 -> 79 TFLOPS
NVIDIA ships every 3090 with fp32-accumulate tensor math at half speed. Tht: FMLA32 = 1/2. gpu-unlock by Zagitta sets it to full before the NVIDIA driver loads. After a cold boot all 4 of my cards say FMLA32 = full.
Real inference got less than I predicted. Strata, Qwen3.8 Flash-Next on 2 cards, prefill:
4K: +12%
10K: +12%
90K: 3205 -> 3536 tok/s, +10%
Before installing I said +14 to 19% at 90K, so I was wrong. The GEMM phases got 23% shorter, but they are only about 40% of Strata prefill, and the int8 experts, copies and CUDA core kernels didnt get any faster.
Outputs did not change. With the same expert cache size, my log-prob dump over 1996 teacher-forced positions is the same file byte for byte as the stock run before the module.
Next is Strata itself, now its everything around the GEMMs that makes prefill slow. How I set it up is in the replies.
2
0
12
1,263
At this point I have to assume they’re handing out resets like candy because demand has fallen off. These models are awful. I don’t care how they benchmark, I’d use GLM5.3 Flash over Astra at this point.
Day 17/ Bet you thought we forgot, but no, it’s reset day today.
Also good time to remind you to get Daybreak Blue access, it only takes 15 mins and you don’t want to have to do it in the middle of an incident.
0
0
5
486
Are there societies where people use their incomes to buy GPUs that reduce their quality of life?
In functional societies, people use their income to improve their quality of life. In dysfunctional societies, a significant portion of that income goes into compensating for the government's failures.
1
0
2
395
Soon: Deepseek v4.1 Flash on 4 CMP 170HXs ⚡️⚡️
Single stream: up to 313t/s
Concurrency 8: up to 759t/s
Cold prefill: 3900t/s
KV Cache: 1 million+ tokens
Releasing tomorrow. More gains to come.
Single stream: up to 313t/s
Concurrency 8: up to 759t/s
Cold prefill: 3900t/s
KV Cache: 1 million+ tokens
Releasing tomorrow. More gains to come.
11
4
60
4,208
Grok @bot will be a superior product to Anthropic's. The reason is simple.
Grok has never refused any sensible request. It will sign up, browse for me, enter captchas for me. It is MY agent.
Anthropic's will be riddled with 'safety' refusals instead of letting people choose.
Grok has never refused any sensible request. It will sign up, browse for me, enter captchas for me. It is MY agent.
Anthropic's will be riddled with 'safety' refusals instead of letting people choose.
Btw, this is less of a competition between Claude Duet and Grok Bot, and more of a battle between Claude and Grok models.
0
0
2
451
I was a child when I first read the Hacker's Manifesto.
It is a simple piece, written by a junior high school student. But there is a certain timeless wisdom to it. It captured an essence of what the internet used to be.
It was the great equaliser. Here you could meet minds e strictures of geography, race, religion, sex, or petty prejudices. Every individual could be judged on what they said and did, and not how they happen to be behind the screen.
Today the internet has become the most poisonous vector for spreading and popularising hate based on these differences. For promoting thinking that destroys the individual and reduces them to their membership in some group or another. I yearn for better days.
It is a simple piece, written by a junior high school student. But there is a certain timeless wisdom to it. It captured an essence of what the internet used to be.
It was the great equaliser. Here you could meet minds e strictures of geography, race, religion, sex, or petty prejudices. Every individual could be judged on what they said and did, and not how they happen to be behind the screen.
Today the internet has become the most poisonous vector for spreading and popularising hate based on these differences. For promoting thinking that destroys the individual and reduces them to their membership in some group or another. I yearn for better days.
0
0
3
267