Intel's Crescent Island Offers 480GB of Memory and No Bandwidth Figure

Intel has published the capacity of its Crescent Island inference card, 160GB as the reference design and up to 480GB from board partners, and has not published the figure that decides what that capacity is worth: memory bandwidth. The estimates in circulation run from roughly 650 GB/s to 1.5 TB/s, a spread of more than 2x on the one number that caps how fast a model can generate tokens.

Intel gave the details at Hot Chips 2026. ServeTheHome reported on Aug. 24 that Sumit Mohan, whom it lists as chief enterprise AI systems architect, and Hong Jiang, an Intel Fellow, presented the part. HWBusters followed on Aug. 25 and says customer sampling is slated for the second half of 2026.

What Intel disclosed about Crescent Island

The card is an Xe3P part built for inference rather than training. The figures below are the ones both write-ups agree on.

Spec	Intel's figure
Memory	160GB LPDDR5X (reference), up to 480GB from partners
Compute	32 Xe3P cores, 256 XMX matrix engines
Cache	32MB unified L2
Host link	PCIe Gen5 x16
Power	350W, air-cooled
Precision	FP4 and MXFP4 up to FP64

The card in that photograph is the Arc Pro B70, an earlier Intel part; no photograph of Crescent Island hardware was available under a free licence. Intel's slides give a software list that includes vLLM, SGLang, llm-d and NVIDIA Dynamo, plus Triton, SYCL and oneDNN, with day-0 drivers promised at launch. They do not give a price, a benchmark against a named competitor, or a bandwidth figure.

The bandwidth number nobody has

ServeTheHome's Patrick Kennedy writes that he wishes Intel "went into more detail" and guesses "650-700GB/s? More?" HWBusters cites a VideoCardz estimate of about 1.5 TB/s, which assumes LPDDR5X-9600 on a very wide bus. Neither number comes from Intel.

HWBusters frames the trade in four words: "Intel is selling room." The next line is "Everyone else is selling bandwidth." Kennedy supplies the comparison from memory, flagging it with "IIRC": the AMD Instinct MI350P at "like 3.6-4TB/s" and the NVIDIA RTX Pro 6000 Blackwell at "1.6-1.79TB/s". Taking his figures at face value, the highest Crescent Island estimate sits just under the low end of the NVIDIA card and below half of the AMD one. The lowest sits at roughly 40 percent of the NVIDIA figure.

Intel does contrast the 160GB reference card with a 96GB GPU in its slides, which Kennedy tentatively identifies as the RTX Pro 6000 Server. Capacity is the axis Intel chose.

What 160GB buys when bandwidth sets the pace

Token generation reads the model's weights once per step. For a dense model that fills the card, the ceiling on tokens per second is bandwidth divided by capacity: how long one full sweep of memory takes. The arithmetic below is ours, not Intel's, and it assumes every byte is read per token, no KV cache, and perfect bandwidth utilisation, so real numbers land lower.

Assumed bandwidth	160GB model	480GB model
650 GB/s	4.1 tokens/s	1.4 tokens/s
700 GB/s	4.4 tokens/s	1.5 tokens/s
1,500 GB/s	9.4 tokens/s	3.1 tokens/s

1. 650 GB/s (ServeTheHome guess, low)4.1 tokens/s
2. 700 GB/s (ServeTheHome guess, high)4.4 tokens/s
3. 1,500 GB/s (VideoCardz estimate)9.4 tokens/s

Source: The Terminal arithmetic: bandwidth estimates from ServeTheHome (Aug. 24) and HWBusters (Aug. 25), capacity from Intel's reference card

A dense model is the worst case for this card, and Intel's argument is that it is no longer the typical case. One Intel slide, as reported by ServeTheHome, says the weights a frontier model holds grew about 7x from Llama 2 70B to Kimi K2 1T, while the bytes read per token fell about 4x. That is the mixture-of-experts effect: hold a lot, read a little. It is the case for buying capacity over bandwidth, and it is a fair one.

It has a limit visible in Intel's own numbers. At 4 bits a weight, a 1-trillion-parameter model needs about 500GB before any KV cache, which is above the 480GB ceiling. Our arithmetic again; it means a model of Kimi K2's size spans more than one card, and the only scale-up path ServeTheHome lists is PCIe Gen5 x16 over an open switch fabric.

Speculative decoding is the vendor-supplied part

Intel also claims 2.9 to 4.9 tokens retired per verification pass with an 8-token draft tree, across frontier mixture-of-experts designs. That is a vendor-supplied figure, and neither write-up gives the batch size, context length or acceptance conditions behind it. If it held on this hardware, it would multiply the ceilings above by the same 2.9 to 4.9: a 160GB model at 1.5 TB/s would move from 9.4 to roughly 27 to 46 tokens per second. Held to the standard we applied to ClickHouse's Postgres storage benchmark, a figure without its conditions is a claim, not a measurement.

Supply is the argument; price is the missing variable

The strongest point for LPDDR5X is availability. HWBusters argues that skipping HBM sidesteps a supply bottleneck, and it lists Ponte Vecchio, Falcon Shores and Gaudi as the Intel accelerators this part follows. Intel has not given a price, and HWBusters notes that no independent performance numbers exist.

Cheap, capacity-rich memory is also the logic behind small-model experiments on commodity gear, such as the Anthropic red-team test that ran Gemma4-26B-A4B across 15 $55 Android TV boxes. The difference is that Intel is asking for a data-centre budget.

What would change this read

Three numbers would settle it. A published bandwidth at or above 1.5 TB/s makes the card a credible home for mixture-of-experts serving at modest batch sizes. A figure at 700 GB/s or below makes it a capacity box for offline and low-concurrency jobs. And a measured tokens-per-second figure on a named model, with batch size, context length and precision given, replaces every estimate in this piece. Sampling in the second half of 2026 is when those could appear.