Last night’s Olmo-core 3 note was a training-systems ticket — open MoE capacity on B300. Yesterday morning’s Everspin CXL PERSYST was persistent media. Voice agents still need a different board: real-time text-to-speech is numerically harsher than token LLMs, and most “FP8 everywhere” playbooks break the waveform. On October 1, 2026, Tenstorrent and Smallest.ai announced a production on-prem stack that runs Lightning V2 natively on Blackhole hardware, backed by a measured L40S bake-off and an arXiv paper. This morning is that waveform inference economics board.
Primary: Tenstorrent newsroom · Smallest.ai blog · arXiv:2604.03279v2 · Galaxy Blackhole specs. Numbers below are Smallest.ai’s / Tenstorrent’s unless noted. Treat cost gains, concurrency, and DNSMOS deltas as vendor measurements under the stated configs, not a claim that every TTS stack will move the same way on your rack.
Spec sheet (named numbers)
| Lightning V2 on Tenstorrent | |
|---|---|
| Partnership announced | October 1, 2026 (tech post Sep 28; paper arXiv:2604.03279v2) |
| What it is | Diffusion TTS co-optimized for Tenstorrent NoC + distributed SRAM; available on Blackhole today (usage-based) |
| Precision coverage | >95% layers LoFi; >80% BlockFloat8 (BFP8) |
| Model / traffic cuts | ~2× model size; 1.8× memory-transfer volume; 4× compute cut (acoustic) / 8× (vocoder) |
| L40S baseline | $9,000; concurrency 3; latency 300 ms → cost gain 1× |
| P150 | $1,400; concurrency 1; latency 250 ms → 2.6× cost gain |
| P100 | $1,000; concurrency 1; latency 250 ms (carried from P150; single-chip, no QSFP) → 3.6× |
| Fleet model (550 × 5 s streams) | 11× L40S ~$100K vs 27× P100 ~$27K or 27× P150 ~$37K (~3–4×) |
| Quality | DNSMOS 3.872 (L40S) vs 3.801 (P150), Δ −0.071; normalized WER 0.009 |
| Layer probe (~6B MACs) | ~60 μs (L40S) vs ~31 μs (P150); authors project 8–12× cost-normalized headroom if more layers match |
| Galaxy Blackhole (partner rack) | 32× Blackhole ASICs; 23 PFLOPS Block FP8; 6.2 GB SRAM @ 2.9 PB/s; 1 TB GDDR6 @ 16 TB/s; 10× 400 GbE/ASIC → 32 TB/s fabric; up to 56× 800 GbE QSFP-DD; 6U air-cooled; 8–10 kW avg; list $160,000 |
| P150 card peek | 120 Tensix cores; 180 MB SRAM; 32 GB GDDR6 @ 512 GB/s; 664 TFLOPS BLOCKFP8; 300 W TBP ( docs ) |
That is a different purchase sentence from “another open TTS checkpoint.” This one is: can I put real-time voice on Blackhole without paying L40S fleet CapEx, and what do they measure when they try.
Why TTS breaks the FP8 playbook
LLMs tolerate small logit noise — softmax and the next token often absorb it. Diffusion TTS generates a continuous waveform over many denoising steps. A rounding error does not just flip a token; it perturbs phase, harmonics, and low-energy fricatives. Smallest.ai’s paper is blunt about the metric trap: L40S vs AMD EPYC CPU end-to-end PCC was only ~0.72 with identical inputs, yet the audio was perceptually fine; a different layer with PCC rounded to 1.0 still caused audible breakage that took more than a month to isolate. Tensor similarity is not a hearing test.
Their answer is selective, not blanket, precision: keep sensitive layers higher fidelity; push LoFi / BFP8 where perceptual checks allow; exploit Tenstorrent’s explicit dataflow (reader / unpack / compute / writer per Tensix core, software-managed SRAM circular buffers, NoC multicast of weight tiles) so single-stream TTS does not pay a full DRAM round-trip every hop the way a batched GPU SM path often does.
What the bake-off actually shows
Single device. L40S wins raw concurrency (3 vs 1) at 300 ms. P150/P100 win latency (250 ms) and sticker price. Normalize by sustained stream capacity and Smallest.ai’s table lands 2.6× / 3.6× cost gains on P150 / P100.
Fleet extrapolation. For 550 overlapping 5-second generations at full utilization, they count ~50 audio-seconds/s per L40S → 11 GPUs ≈ $100K, versus ~20 audio-seconds/s per P150 → 27 cards ≈ $37K (or $27K on P100). That is the ~3–4× headline. It assumes 100% TTS-stage utilization in an AST→LLM→TTS pipeline and listed device prices — your discount schedule and idle mix will move the dollars.
Quality. DNSMOS drops 0.071; normalized WER between L40S and P150 outputs is 0.009. Authors call the DNSMOS gap minor perceptual variation. Bring your own listening panel before you quote “no audible artifacts” into a compliance deck.
Headroom. One ~6B MAC layer at ~31 μs on P150 vs ~60 μs on L40S is a software-limited probe, not a full-graph guarantee. The 8–12× cost-normalized projection is an extrapolation if more layers get the same treatment — mark it as such.
Where it sits next to other boards
| Purchase | Product | What it buys |
|---|---|---|
| Open MoE trainer (last night) | Ai2 Olmo-core 3 | B300 capacity tax fix; training tok/s |
| Persistent CXL media (yesterday AM) | Everspin PERSYST | Byte-addressable MRAM on CXL |
| On-prem TTS on Blackhole (this post) | Lightning V2 × Tenstorrent | Measured L40S cost/concurrency bake-off; BFP8/LoFi co-design; Galaxy rack path |
| Decision-model flood (Oct 1) | Clef / Decider / Strands | Structured yes/no/choice probs — different lane |
| Personal GB10 desk box (today) | DGX Spark 64GB | Local agent SKU + Sync Cluster — not this rack story |
Do not collapse this into “Tenstorrent is cheaper than NVIDIA.” The claim is narrower: this diffusion TTS, under these concurrency/latency points, on P100/P150, vs L40S list economics — then a Galaxy Blackhole path for denser on-prem voice.
Caveats
- Vendor bake-off. Latency, concurrency, DNSMOS, and dollar gains are Smallest.ai’s. Independent reproductions on your stack are not this post.
- P100 latency is carried. Paper states Lightning V2 is single-chip and does not use QSFP; P100 latency equals measured P150.
- Fleet math is a model. 550 streams × 5 s audio, 100% TTS utilization, list prices — change any knob and CapEx moves.
- Not full-model low precision. Some layers still refuse LoFi/BFP8 without audible damage; coverage caps exist.
- Compiler headroom. Authors say program configs are sub-optimal; 8–12× is projected, not shipped system gain.
- Galaxy vs card. Partnership markets Galaxy Blackhole servers; the L40S tables are P100/P150 card-class. Do not paste 23 PFLOPS rack peak onto the single-stream TTS latency row.
- Lightning V3 is future. Blog says V3 already surpasses V2 in quality/efficiency and is planned for the same co-design path — not this ship.
How I would evaluate
- Read §3–4 of arXiv:2604.03279v2 for the PCC case and Table 3 before quoting 4×.
- Reproduce the single-device concurrency/latency points on a P150 (or QuietBox) with your own 5-second prompts — watch p95, not only mean.
- Run a blinded listening test next to DNSMOS; treat PCC as a debug aid, not a ship gate.
- Price a 550-stream equivalent with your L40S and Blackhole quotes, idle fraction, and power — CapEx tables without OpEx are incomplete for on-prem voice.
- Keep this board separate from Olmo-core training and from the Oct 1 decision-model flood — this is waveform inference economics on Tensix.
Bottom line
On October 1, 2026, Tenstorrent and Smallest.ai put Lightning V2 on Blackhole as a production on-prem voice path, with a paper that pushes >95% LoFi and >80% BFP8 through a numerically fragile diffusion TTS and still reports DNSMOS 3.801 vs L40S 3.872 (Δ −0.071) and normalized WER 0.009. The durable systems story is the cost table — 250 ms / concurrency 1 on $1,000–$1,400 P100/P150 vs 300 ms / concurrency 3 on a $9,000 L40S, extrapolating to roughly 3–4× lower accelerator CapEx for a 550-stream model — plus the honest finding that PCC can look perfect while the ear fails. Galaxy Blackhole gives the dense rack SKU (32 ASICs, 23 PFLOPS Block FP8, $160K list) if you outgrow cards. Code path and paper are public; treat projected 8–12× headroom and Lightning V3 as future work.
Create-only. Specs from Tenstorrent’s Oct 1 partnership announcement, Smallest.ai’s Lightning V2 technical post, arXiv:2604.03279v2, and Tenstorrent Galaxy / Blackhole documentation. Update when third-party reproductions, multi-chip Lightning, or Lightning V3-on-Blackhole numbers ship.
Sources: Tenstorrent partnership · Smallest.ai blog · arXiv:2604.03279v2 · Galaxy Blackhole · Blackhole PCIe specs.
Anpoo Sivanadi, Staff Software Engineer