Wednesday, February 4th 2026

Tenstorrent Cuts 20 Cores From Already-Shipping "Blackhole" P150 Cards

Tenstorrent, a startup focused on designing high-performance AI accelerators and led by the renowned computer architect Jim Keller as CEO, has announced significant hardware updates to its existing Blackhole P150 accelerators, which include the P150a and P150b models. In the latest documentation change, the company notes that its Blackhole P150 accelerators will now work with about 14.3% fewer cores than originally advertised. In the official documents, the P150 accelerators are now shipping with 120 working "Tensix" cores instead of the previously advertised 140 cores. The reason for this change is unknown, as the company provided a vague explanation: "To present a unified interface to metal and other system software, firmware v19.5.0 and later will change the core count on all existing cards to 120. Typical workloads show a non-material (~1-2%) performance difference."

The Blackhole P150 accelerators featured 140 "Tensix" cores and 32 GB of GDDR6 memory, operating at up to 300 W in an actively cooled form factor designed for desktop workstations, and the P150a model includes four passive QSFP-DD 800G ports. However, as the number of cores is reduced by approximately 14%, TeraFLOPS take a nosedive as well. In the older documents for the 140-core SKUs, the BLOCKFP8 8-bit floating point performance was listed at 774 TeraFLOPS, while the new 120-core version reduces that number to 664 TeraFLOPS at the same precision level. Why this sudden change is happening is still a mystery. However, the HPC community with a lot of knowledge in the industry suggests a few reasons.
The first one is that the Blackhole P150 accelerators are heavily limited by their thermal capacity, which stands at 300 W only. This means that a 140-core version might not be reaching its full potential easily, limiting the card's performance target. A second and also plausible reason is that the accelerator is aiming to increase its silicon yield, where it is much easier to get a chip with 120 working cores instead of all 140. As the company is preparing to scale its accelerator design to multi-chip systems, getting the silicon yield within an acceptable range is the most important factor. Finally, it could also mean that the shipping silicon versions are not entirely ready, and some argue that this represents the equivalent of shipping experimental silicon.
Source: FelixCLC
Add your own comment

12 Comments on Tenstorrent Cuts 20 Cores From Already-Shipping "Blackhole" P150 Cards

#1
igormp
TIL that Felix is not with Tenstorrent anymore either, huh
Posted on Reply
#2
tenstorrentcustomer
I have a p150a and now cannot upgrade to the latest firmware without getting my hardware materially downgraded. So in the unlikely event that Tenstorrent fixes e.g. the "fan speed is a RNG" issue, I will not be able to upgrade. If Tenstorrent wanted future p150a cards to have 120 Tensix cores instead of 140, the honorable thing would have been to introduce a seperate SKU for that (p130a?), not downgrade hardware of existing customers.

"I have altered the deal, pray I don't alter it further."
Posted on Reply
#3
AleksandarK
News Editor
tenstorrentcustomerI have a p150a and now cannot upgrade to the latest firmware without getting my hardware materially downgraded. So in the unlikely event that Tenstorrent fixes e.g. the "fan speed is a RNG" issue, I will not be able to upgrade. If Tenstorrent wanted future p150a cards to have 120 Tensix cores instead of 140, the honorable thing would have been to introduce a seperate SKU for that (p130a?), not downgrade hardware of existing customers.

"I have altered the deal, pray I don't alter it further."
Interesting! How many cards do you have? Is the software good? What do you use it for if not secret :)
Posted on Reply
#4
tenstorrentcustomer
AleksandarKInteresting! How many cards do you have? Is the software good? What do you use it for if not secret :)
I have one card (and also owned Grayskull, the first gen), I primarily use it as a validation platform for my own (possibly dogshit) theories on how to make a good compiler for AI dataflow architectures. The hardware is quite interesting in that it is much more flexible than MIT's Eyeriss, offering possibilities for novel schedules that go beyond simplified schemes like "row stationary." So yeah, the hardware is cool and the ISA is so clean that the main problem to solve is "how to get data where it needs to be/when does it need to get there?" and the actual computation can almost be treated as an afterthought.

Sadly, the software stack is where it all falls apart, as their software pretty much fails entirely to capitalize on the novel possibilities offered and instead treats the hardware like: "what if we made something that PyTorch wants to interface to, and then struggled really hard to recover a sense of mechanical sympathy from that?" Hot take, possibly unfair, but there is a reason why big tech is taking out huge loans to buy AI hardware, and very little (if any) of that makes it into Tenstorrent's pocket, you know?
Posted on Reply
#5
Dirt Chip
Just think the extinction level outrage if say NV/AMD/Intel would do such retroactive trick…
Axing cores of an already lunched product with bios update is a new level.
Posted on Reply
#6
nguyen
I bet Jim Keller is going to complain about the CUDA moat cutting cores on his product
Posted on Reply
#7
Vayra86
So, 20 cores sucked into a black hole there. Nice.
Posted on Reply
#8
Tomorrow
This sounds incredibly scummy. Intel axed AVX-512 on already shipped 12th gen Alder Lake chips with a microcode update and that was already scummy, but disabling over 14% functional compute units takes the cake. I wouldn't be surprised if this causes lawsuits or even a class action lawsuit. Not to mention damaging Tenstorrent's image.

How can any potential buyers be sure that buying their future hardware will not being retroactively downgraded too?
Tenstorrent will lose a ton of business because of this.
Posted on Reply
#9
mrnagant
Run some benchmarks. Board power across the stack before and after was 300W. Maybe they just hit power wall and there was actually really no difference at all. Could potentially help workloads run faster, have better timing or be more accurate maxing out all the AI cores vs having them run at unbalanced utilization rates if power limited proper full 100% utilization.

I would doubt it has anything to do with yields. It is likely manufactured on a 12nm or 16nm process.
Posted on Reply
#10
tenstorrentcustomer
mrnagantRun some benchmarks. Board power across the stack before and after was 300W. Maybe they just hit power wall and there was actually really no difference at all. Could potentially help workloads run faster, have better timing or be more accurate maxing out all the AI cores vs having them run at unbalanced utilization rates if power limited proper full 100% utilization.

I would doubt it has anything to do with yields. It is likely manufactured on a 12nm or 16nm process.
Why speculate when several of us have p150a:s and can confirm they run stable at highest AICLK running stress benchmarks like Tenstorrent's own tt-burnin? Furthermore, the reason for this change is actually documented: github.com/tenstorrent/tt-zephyr-platforms/commit/b0ff3c7c2627cb285a696b5789871ddad686ed4e "New P150x cards will be shipped with Bin 3 parts, which lack 2 functional tensix columns", so this is purely a binning/yield issue, where Tenstorrent wants to pretend the new cards are just as good as the old cards, by crippling the old.
Posted on Reply
#11
Roph
Don't give Nvidia ideas :roll:

When 6000 series cards ship with missing ROPs, they'll just cut the normal ones down to the same ROP count as the defective ones through a forced driver update.

.... Then sell you your own ROPs back to you, with a monthly subscription
Posted on Reply
#12
Vayra86
RophDon't give Nvidia ideas :roll:

When 6000 series cards ship with missing ROPs, they'll just cut the normal ones down to the same ROP count as the defective ones through a forced driver update.

.... Then sell you your own ROPs back to you, with a monthly subscription
Way too much effort. They just let you rent the whole thing and log in to use it. For all you know there's really a 1060 6GB running code.

Hey... they already do that.
Posted on Reply
Add your own comment
Nov 20th, 2026 16:27 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts