Intel Xeon Scalable Gets a Rebrand: Intel "Xeon 6" with Granite Rapids and Sierra Forest Start a New Naming Scheme

[erek]

"Intel has also teased an exciting feature in its forthcoming Granite Ridge processors-support for the MXFP4 data format. This new precision format, backed by the Open Compute Project (OCP) and major industry players like Nvidia, AMD, and Arm, promises to revolutionize performance. It could reduce next-token latency by up to 6.5X compared to fourth-gen Xeons using FP16. Additionally, Intel stated that Granite Ridge will be capable of running 70 billion parameter Llama-2 models, a capability that could open up new possibilities in data processing. Intel claims that 70 billion 4-bit models run entirely on Xeon in just 86 milliseconds. While Sierra Forest is slated for this quarter, Intel has not provided a specific launch timeline for Granite Ridge, stating only that it will arrive "soon after" its E-core counterpart. The Xeon 6 branding aims to simplify the product stack and clarify customer performance tiers as the company gears up for these major releases."

Source: https://www.techpowerup.com/321345/...s-and-sierra-forest-start-a-new-naming-scheme

 

[BrainEater]

I like the scalable stuff for sure. There are still connectors on my mobo I have never made functional. (supermicro X11DPI-nt, Xeon gold 6144's)
The supermicro X13 series Motherboards are sooooo sweet, I look forward to the new Xeon 6 ones.

 

[toast0]

Sweet, my budget server is now 3 branding cycles behind (it's the last gen before the 26xx series)

 

[bwang]

A much needed rebranding. The precious metals based scheme was utter chaos, with at one point something like 100 finely segmented SKUs. Conroe to Haswell at least had feature parity across the entire stack (except for the two socket-filler SKUs each generation with no turbo or HT), but once Skylake rolled around Intel marketing decided that every possible opportunity to segment SKUs should be taken advantage of. I'm really hoping this generation they keep things simple with unlocked features through the stack and tunable TDP rather than selling the same part as three different SKUs with different TDPs and socket counts.

 

[Lakados]

A much needed rebranding. The precious metals based scheme was utter chaos, with at one point something like 100 finely segmented SKUs. Conroe to Haswell at least had feature parity across the entire stack (except for the two socket-filler SKUs each generation with no turbo or HT), but once Skylake rolled around Intel marketing decided that every possible opportunity to segment SKUs should be taken advantage of. I'm really hoping this generation they keep things simple with unlocked features through the stack and tunable TDP rather than selling the same part as three different SKUs with different TDPs and socket counts.

We’re still going to get a bunch of SKU’s traditionally they make up the different binnings of the silicon. It allows them to sell as much of the product as they can. Even assuming a greater than 80% perfect yield rate those failures manifest in unusable cores, unstable cache, etc… so they burn out the bad bits so it’s not there and re assign it a different model accordingly.

Bigger chips means more types of errors which means more binnings which gives more SKU’s.
Smaller chips brings smaller errors means fewer binnings means fewer SKU. So we’ll likely have less SKU’s then we are used to, but it’s still going to be a lot.