Intel this morning is kicking off the second day of their Vision 2024 convention, the corporate’s annual closed-door enterprise and customer-focused get-together. While Vision just isn’t usually a hotbed for brand new silicon bulletins from Intel – that’s extra of an Innovation factor within the fall – attendees of this yr’s present usually are not coming away empty handed. With a heavy deal with AI occurring throughout the business, Intel is utilizing this yr’s occasion to formally introduce the Gaudi Three accelerator, the next-generation of Gaudi high-performance AI accelerators from Intel’s Habana Labs subsidiary.
The newest iteration of Gaudi will likely be launching within the third quarter of 2024, and Intel is already transport samples to clients now. The {hardware} itself is one thing of a blended bag in some respects (extra on that in a second), however with 1835 TFLOPS of FP8 compute throughput, Intel believes it’s going to be greater than sufficient to carve off a chunk of the expansive (and costly) AI marketplace for themselves. Based on their inside benchmarks, the corporate expects to have the option beat NVIDIA’s flagship Hx00 Hopper structure accelerators in at the very least some vital giant language fashions, which is able to open the door to Intel grabbing a bigger piece of the AI accelerator market at a vital time within the business, and a second when there merely isn’t sufficient NVIDIA {hardware} to go round.
The upcoming launch of Gaudi Three additionally comes amidst a change in how Intel is positioning its AI accelerator merchandise – one which has seen the Gaudi lineup elevated to Intel’s flagship server accelerator. Traditionally downplayed in favor of Intel’s GPU Data Center Max merchandise (Ponte Vecchio), Habana Labs and Gaudi have gained a brand new respect inside Intel following the cancellation of Rialto Bridge in favor of the 2025 launch of Falcon Shores. In brief, Intel doesn’t have every other new AI accelerator silicon popping out in addition to Gaudi 3, so Intel goes to battle with the chip that it has.
Intel 2023 Data Center Silicon Roadmap
Which is to not knock Habana Labs or Gaudi Three upfront. Intel thinks they will win right here on efficiency; if they will, that’s a giant deal. But it is a product that’s clearly been uplifted from a side-project beneath the Intel umbrella to a front-and-center processor. So the scope of Gaudi 3’s skills, {hardware}, and how much markets Intel is chasing, is narrower than we’ve seen with a few of their different flagship merchandise.
| Gaudi Accelerator Specification Comparison | |||||
| Gaudi 3 | Gaudi 2 | Gaudi (1) | |||
| Matrix Math Engines | 8 | 2 | 1 | ||
| Tensor Cores | 64 | 24 | 8 | ||
| Clockspeed | ? | ? | ? | ||
| Memory Clock | 3.7Gbps HBM2e | 3.27Gbps HBM2e | 2Gbps HBM2 | ||
| Memory Bus Width | 2x 4096-bit | 6144-bit | 4096-bit | ||
| Memory Bandwidth | 3.7TB/sec | 2.45TB/sec | 1TB/sec | ||
| VRAM | 128GB (2x 64GB) |
96GB | 32GB | ||
| FP8 | 1835 TFLOPS | 920 TFOPS? | N/A | ||
| BF16 | 920 TFLOPS? | 230 TFLOPS? | ? TFLOPS | ||
| Interconnect | 200Gb Ethernet 24 Links (600GB/sec) |
100Gb Ethernet 24 Links (300GB/sec) |
100Gb Ethernet 10 Links (120GB/sec) |
||
| Processor | Gaudi 3 | Gaudi 2 | Gaudi (1) | ||
| Transistor Count | 2x (Loads) | ? | ? | ||
| TDP | 900W | 600W | 350W | ||
| Manufacturing Process | TSMC 5nm | TSMC 7nm | TSMC 16nm | ||
| Interface | OAM 2.0 | OAM 1.1 | OAM | ||
Diving in to the {hardware} itself, let’s check out Gaudi 3.
Gaudi Three is a direct evolution of the Gaudi 2 {hardware}. Habana Labs has settled on an structure they like and take into account profitable, so Gaudi Three isn’t orchestrating a large overhaul of their structure (that can include Falcon Shores). The flip facet to that, nevertheless, is that there’s not a ton to speak about right here in…
