Tenstorrent Blackhole

Blackhole is an AI accelerator chip from Tenstorrent, and Galaxy Blackhole is the server the company builds around it. Tenstorrent lists the Galaxy Blackhole as a machine holding 32 Blackhole ASICs and rates it at 23 PFLOPS of block FP8 performance.^[1]

A card built from many small cores

Tenstorrent's documentation lists 240 Tensix cores on a Blackhole card, 1.5 MB of SRAM for each Tensix core, and 64 GB of GDDR6 for the card as a whole.^[2] Tensix is the name Tenstorrent gives the repeating compute block in its designs; a card is essentially a large array of them, with memory around the edge.

That arrangement is a design choice worth spelling out. Instead of concentrating arithmetic in a handful of very wide compute units, the chip spreads it over hundreds of modest cores, each with memory of its own close by. Keeping operands in nearby SRAM avoids moving them repeatedly out to off-chip memory and back, which is where much of the time and energy in matrix-heavy AI workloads goes. The 64 GB of GDDR6 per card supplies the capacity that the small local memories cannot hold.

Block FP8, the format in which the server's throughput is quoted, is a low-precision number format: values are grouped into blocks that share a scaling factor, so the hardware spends few bits per number while keeping a usable dynamic range. Fewer bits per value means more arithmetic per unit of power and memory bandwidth, at some cost in accuracy that training and inference software manages with scaling and accumulation.

The Galaxy Blackhole server

Beyond the 23 PFLOPS of block FP8 across 32 chips,^[1] Tenstorrent publishes a memory hierarchy for the system: 6.2 GB of accelerator SRAM running at 2.9 PB/s, and 1 TB of GDDR6 at 16 TB/s.^[1] The two tiers do different jobs. The SRAM tier is small and extremely fast; the GDDR6 tier is far larger and slower. The ratio between them tells you how much of a workload has to fit in the fast tier to run well.

Networking is where the design is most distinctive. Each Blackhole ASIC carries 10 400-gigabit Ethernet links, which in the 32-chip server add up to 32 TB/s of accelerator fabric.^[1] For linking servers into larger clusters, Tenstorrent lists up to 56 800-gigabit QSFP-DD ports, giving 11.2 TB/s of cluster scale-out.^[1] Ethernet is a widely deployed standard interconnect rather than a link designed only for accelerators, and here both the chip-to-chip fabric and the scale-out network are built on it.

Power is documented at 8 to 10 kW average and 12 kW maximum for the system, configurable up to 14.5 kW.^[1] Those figures matter as much as the FLOPS number: they set what a datacenter can actually place and run, and they are why accelerators of this class are discussed together with their power and cooling requirements rather than as cards alone.

One programming model from one chip to thousands

Tenstorrent says its Tensix Neo architecture scales from a single chip to thousands under one programming model across an Ethernet mesh.^[1] The other numbers are what support that claim: a card with 240 cores, a server with 32 cards, and a cluster of servers are meant to be described in the same terms, so that software written for one chip does not have to be rewritten for a rack.

Read together, the published figures describe a machine with an unusually explicit hierarchy: SRAM per core, GDDR6 per card, Ethernet links per ASIC, QSFP-DD ports per server, and a power envelope given as a range rather than a single point. Blackhole is therefore easiest to understand not as one peak number but as a set of ratios — memory bandwidth to capacity, network bandwidth to compute, performance to power — that decide which workloads run well on it.