Twitter user YuuKi_AnS has shared several screenshots exposing the information about Intel's upcoming Sapphire Rapids-SP 56-core processor. While there are no benchmarks, we get a sneak peek at the chip's specifications.
Like Intel's mainstream Alder Lake processors, Sapphire Rapids will feature the chipmaker's Golden Cove cores and Intel 7 node, previously known as the 10nm Enhanced SuperFin process. The leaked Sapphire Rapids-SP processor has 56 cores and 112 threads. On paper, it looks underwhelming since AMD's EPYC 7003 (Milan) chips span up to 64 cores and 128 threads. Nonetheless, it'll be exciting to see Golden Cove go up against Zen 3 in the data center space.
The 56-core Intel chip also has 112MB of L2 cache and 105MB of L3 cache. The highest tier EPYC 7003 has up to 256MB of L3 cache. However, Intel is also readying Sapphire Rapids with up to 64GB of HBM2e memory; meanwhile, AMD has prepared its Milan-X chips with 512MB of 3D V-Cache.
The Sapphire Rapids-SP processor from YuuKi-AnS is an engineering sample (ES), so don't take the clock speeds seriously. Thus far, the 56-core part operates with a 1.9 GHz base clock and a 3.3 GHz boost clock. According to the report, the single-core boost clock escalates up to 3.7 GHz.
The Sapphire Rapids-SP chip (ES2 QYFS) was on Intel's C741 (Emmitsburg) platform with 1TB of DDR5 memory with CL40-39-38-76 timings. The Socket E (LGA4677) motherboard housed the processor.
The processor has a 350W PL1 rating if the information is accurate and a 420W PL2 rating. However, the actual enforced power limit from inside the BIOS is at a whopping 764W. The EPYC Milan 7763, the flagship Milan SKU, has a 280W TDP. The Twitter user didn't share what type of cooling the Sapphire Rapids-SP processor had at its disposal. However, the report revealed that the Sapphire Rapids-SP chip hit 99 degrees Celsius.
Sapphire Rapids was supposed to come out in the second quarter of the year, but delays have pushed it to the third quarter. That'll be a big problem for Intel since AMD's EPYC 7004 (Genoa) lineup may arrive around the same timeline. Unlike Milan, Genoa will leverage TSMC's groundbreaking 5nm process node and wield AMD's Zen 4 cores. Genoa will also bump EPYC chips' core count from 64 to 96, meaning Sapphire Rapids will contend with a 96-core, 192-thread Genoa monster. As a result, Sapphire Rapids will have its hand full.
Also how many watts should 56 cores running 512 at 3.3ghz burn? That's 7.5w/core or 60w for an 8 core. And does that include the 64Gb HBM2e?
sorry, I meant need a level 2 charger
A far more useful metric than the watts/core is performance per core. Or, if you prefer, watts/core for the same performance per core. With 27% higher frequency we can theoretically have 27% higher performance if IPC were equal. Then comes IPC which is a per workload metric. In workloads like Cinebench R23, Golden Cove cores have a 17-17.5% better IPC than Zen 3 cores. So in that workload we have a total of 48.6% (=1.27x1.17) higher performance per core in that workload. Factoring this we have 5W/core for the same performance, not 7.5. So not so far off from the 4.38W/core of Epyc. Then there are other things to consider like the fact that this is an engineering sample and final silicon will be better. It might as well get down to 4-4.5W/core for the same performance. Zen 4 Epyc will be launching next year. By that point (or within 3-4 months) Intel will be launching SPR's replacement. SPR's delayed launch won't affect (or at least not much) the timeline of its replacement.
The 7H12 is not the latest 64c Epyc. The 7H12 was released in September 2019, is a high frequency CPU, and is a Rome uArch design. Whereas the 7763 is a Milan CPU that was released in March 2021. The 7763 has a 2.45GHz base clock, 3.0GHz all core boost, and a 3.5GHz single core boost.
AMD does not base their TDP around the base clock speed. AMD uses total package power and their server CPUs will top out a 280W regardless of instruction set being used. Intel on the other hand bases their TDP around the base clock speed. This is why we see their CPUs say "120W TDP" but under load will draw 300W and can continue to do that if the Tau is set to infinity.
That SPR engineering sample has a 1.9GHz base clock and a 3.3GHz boost clock. From the article "Thus far, the 56-core part operates with a 1.9 GHz base clock and a 3.3 GHz boost clock. According to the report, the single-core boost clock escalates up to 3.7 GHz."
If that max draw is 420W under AVX-512 then at best the CPU will be running at the base clock of 1.9GHz. Previous versions of Intel CPUs that ran AVX-512 instructions could not run at their base clock with all cores running in AVX-512. It wasn't uncommon to see up to 500MHz reduction in clock speed from the base clock when all cores were running AVX-512. Intel seems to have gotten around some of this clock speed reduction by making the TDP in the PL2 being MUCH higher.
Again the frequency of this SPR engineering sample is only a 1.9GHz base clock. That puts the SPR 29% slower in the frequency which would negate your "17% higher IPC." When normalized for score per GHz in Cineabench R23, the Golden Cove with its best setup (Win 10 + DDR4) is 13% faster than Zen 3 in the single threaded benchmark. Golden Cove @ 2008 = 386.2 per GHz vs Zen 3 @ 1636 = 340.8 per GHz. https://www.anandtech.com/bench/product/2674?vs=2894 That higher performance won't matter if the clock speeds aren't high enough. This is the same reason that Ice Lake Xeon isn't always faster than Cascade Lake Xeon.
SPR has been delayed to Q3 this year. As of January 2022, Zen 4 was scheduled to come out end Q3 or beginning Q4 2022. The odds of Intel releasing Emerald Rapids all of 6ish months after SPR is released almost nil.