Intel is doing the spec improve by opinion weblog factor once more, this time it’s Xeon reminiscence speeds. SemiAccurate likes the tech of the improve however the format is… meh.
We will’t hyperlink the unique as a result of it isn’t up when that is being written, however it’s principally three bullet factors completed over 2+ pages. The bullet factors are that present Xeon 6 and 6+ (no clue what these are literally) will quickly assist DDR5-8000. Then comes the tease for MRDIMM-8800 within the upcoming 6900P CPUs in Q1/27. Lastly we’ve got the enjoyable indisputable fact that Intel considers excessive core counts to be 72-128 cores. In the event you heard that too, it’s AMD engineers laughing. Lets check out these items in a bit extra element.
The reminiscence pace improve on the 6/6+ collection from 6400 to 8000MT/s is an efficient factor, interval. Positive it tends to necessitate top quality PCB materials and plenty of layers however that’s the case for everybody, AMD and ARM included. The 25% pace bump is welcome, Intel’s wrapping it in AI BS isn’t, it’s only a easy pace improve albeit a tough one to engineer. 1DPC is right here to remain, why didn’t Intel and AMD leap on board IBM’s reminiscence buffer prepare, it’s a higher means.
With MRDIMMs going from 8000 to 8800MT/s we additionally go from MRDIMM Gen 1 to Gen 2. There are some things right here, first off that on condition that MRDIMMs are successfully rated at ~2x the clock pace of the chips, the present 8000MT/s gadgets run at a lot decrease clocks. The online result’s a few 35-40% enchancment in precise transfers. Shifting from 8000 to 8800MT/s is a decrease share bump than DDR5 going from 6400 to 8000MT/s. That mentioned it’s nonetheless welcome.
Extra troubling is the declare that the 8800MT/s components are ‘Gen 2’. Why? As a result of there isn’t any data on what Gen 2 is. Did they bump the clocks? Is it a spec change? Suitable? Will you want new CPUs? New boards? Who is aware of? This is the reason you don’t do spec bulletins by way of fluff weblog posts, folks like me ask questions and mock you. See the subsequent paragraph for extra mocking.
Right here is the extra mocking. Intel calls Granite Rapids CPUs within the 72-128c vary ‘excessive core counts’ which they appear to be utilizing individually from HCC gadgets from what we will inform. Intel has unquestionably higher reminiscence assist however is a bit mild on reminiscence channels. AMD is at present at 192 2-thread cores per socket, the small Intel socket caps out at 86c with 8-channel reminiscence whereas be massive socket has as much as 128c with 12-channel reminiscence. AMD has one socket with 12-channel reminiscence and one set of boards and associated infrastructure. 128c vs 192c isn’t a battle and when Venice debuts actually quickly we can have 256c with 16-channel reminiscence too. Typically it’s best to simply go away issues with the abbreviations, see HCC for extra.
So in the long run, noxious ‘as a result of AI’ fluff apart, we’ve got a 25% bump in DDR5 speeds from 6400 to 8000MT/s. It is a good factor when it is available in September, attainable infrastructure upgrades apart. Identical with the MRDIMM increase, 8000 to 8800MT/s is smaller however nonetheless welcome subsequent 12 months. Once more the query of attainable infrastructure upgrades and incompatibility with ‘Gen 2’ apart. All this, and the mocking might have been averted with a technical temporary, even a brief one.S|A
