On July 23, 2026, AMD and Cerebras Systems announced an AI inference platform that combines AMD Helios with the Cerebras Wafer-Scale Engine (WSE). Helios processes prompts and long context, while WSE handles low-latency decoding and token generation. What's new about this partnership is that it incorporates WSE into the same inference path as Helios. As NVIDIA also moves in the same direction with Vera Rubin and Groq 3 LPU, the axis of competition has shifted from the raw speed of individual chips to the ability to orchestrate heterogeneous engines as a single service.
Prefill goes to Helios, decode goes to WSE
Large language model inference has two phases with different computational characteristics: prefill, which reads in the input all at once, and decode, which generates output one token at a time. According to the joint announcement, Helios handles prompts and large context windows with high throughput. WSE takes on decoding, which depends heavily on memory bandwidth, and returns responses with low latency. Rather than forcing both processes onto the same type of accelerator, this design assigns each to hardware suited for it.
Helios packs 72 Instinct MI455X GPUs into a single rack along with 6th Gen EPYC CPUs, Pensando networking, and ROCm. Per rack, it reaches peak FP4 compute of 2.9 exaFLOPS and peak FP8 of 1.4 exaFLOPS. It also carries 31TB of HBM4 and 1.7PB/s of memory bandwidth, giving it the capacity to host the model itself and its context as the side that processes large volumes of requests and long inputs in parallel.
What Cerebras brings to the table with its current WSE-3 is 44GB of on-chip SRAM placed right next to the compute units. Its bandwidth reaches 21PB/s, and a single chip integrates 900,000 cores and 4 trillion transistors. While SRAM has less capacity than HBM4, it makes it easier to minimize the time spent repeatedly moving data from external memory. For token generation, which involves continuous sequential processing, this proximity translates directly into response speed. Note that whether the joint service will actually use WSE-3 has not been specified.
For coding assistants and real-time agents, even high aggregate throughput doesn't improve the user experience if generation speed per user is slow. Conversely, a low-latency engine alone struggles to accept many long inputs simultaneously. The division of labor between Helios and WSE is an attempt to advance both of these requirements separately.
What the "up to 5x" figure is actually being compared against
AMD and Cerebras state that combining the two engines yields up to 5x the "tokens per second per kilowatt." However, this 5x figure is neither a comparison against NVIDIA products nor a result measured in a live customer service. It comes from modeling the two companies conducted in July 2026 using the Kimi 2.6 1T model, under conditions of comparable interactivity, and the comparison baseline is a configuration using Cerebras WSE alone for processing.
What can be read from these conditions is a design-level expectation: adding Helios lets WSE focus on decoding, increasing the overall system's throughput per watt. It does not mean "token generation becomes 5x faster." Time to first token, per-user generation speed, and concurrent connection counts have not been disclosed, so it's not yet possible to measure at what level low latency and high throughput can actually be reconciled.
Furthermore, the announcement does not specify how Helios and WSE are connected. What data is transferred between the two engines and which software routes requests between them also remains unclear. If data movement increases, the time saved on WSE could be lost on the networking side. To translate the up-to-5x figure into commercial service value, real-world figures for latency under congestion and power consumption will be needed, in addition to average values.
The GPU-and-SRAM division of labor that NVIDIA also chose
In March 2026, NVIDIA announced an inference platform combining Vera Rubin NVL72 with Groq 3 LPX. Groq 3 LPX packs 256 Language Processing Units (LPUs) into a single rack, with a total of 128GB of on-chip SRAM, 40PB/s of SRAM bandwidth, and 640TB/s of scale-up bandwidth. The structure—where the GPU handles massive HBM capacity and compute performance while the LPU, using SRAM as its main memory, supplements low-latency processing—closely resembles the AMD and Cerebras configuration.
On the other hand, NVIDIA has disclosed its division of labor at a finer granularity. The Rubin GPU handles prefill and decode attention, while Groq 3 LPX takes on feed-forward network (FFN) and Mixture of Experts (MoE) expert computation. NVIDIA Dynamo controls both, passing intermediate data back and forth each time a token is generated. By contrast, the AMD and Cerebras announcement stops at the broad outline of Helios handling input and WSE handling decode.
The materials needed to determine which is faster are not yet in place. While WSE-3 has 44GB of SRAM on a single chip, Groq 3 LPU connects many chips together to expand the model's working area. However, AMD and Cerebras have not disclosed how many WSE units the joint service will use. Until results emerge using the same model, the same input/output lengths, and matched concurrent connection counts, lining up chip counts and stated bandwidth figures side by side won't allow for a comparison of operating costs.
H2 2026: the completeness of the integration will be put to the test
Cerebras plans to deploy Helios in its own data centers and offer the joint service first through Cerebras Cloud in the second half of 2026. AMD says Helios is already in production, and OpenAI is expected to bring Helios online starting in Q4 2026. That said, the launch month, supported models, and pricing for the Cerebras offering have not been announced. Whether users will be able to deploy it in their own environments or whether it will be exclusive to Cerebras Cloud also remains undetermined.
This partnership also introduces a shift in AMD's product strategy. In a post-announcement briefing, Lisa Su told The Register that AMD intends to collaborate with multiple companies whose technologies complement Helios, and to further advance workload division going forward. While AMD builds out its own GPUs, CPUs, and networking, it is also incorporating external specialized engines for use cases such as low-latency decoding. That choice extends Helios's role beyond a fixed-configuration rack, positioning it as a platform to which compute resources can be added according to the use case.
Whether this becomes a commercial success can be judged by generation speed measured with a common model, latency under congestion, and price per token. If Cerebras Cloud discloses these figures in the second half of 2026 and manages to absorb the data movement that occurs between Helios and WSE, AMD will be able to present an alternative path to the heterogeneous inference approach that NVIDIA already demonstrated with Groq 3 LPX. The first real basis for judgment will emerge once the joint service actually becomes available.