By: Adrian (a.delete@this.acm.org), April 7, 2021 12:29 am
Room: Moderated Discussions
anonymous2 (anonymous2.delete@this.example.com) on April 6, 2021 4:28 pm wrote:
> > Good:
> > The cheaper SKUs are no longer crippled by disabling 1 of the 2 FMA units.
> > No clock frequency penalty for AVX-256 and light AVX-512 instructions.
> > For heavy AVX-512 instructions, the clock frequency 80% of normal frequency.
>
> I would rephrase those 3 to "AVX-512 is a bit better" with the caveat "most things and most people
> don't care". Maybe AVX-512 is useful for futureproofing but it doesn't look that way yet.
>
> > There are a few SKUs with good price/performance, e.g. Gold 6312U.
>
> Really? The costs aren't that low and the performance isn't that good.
>
> This really seems like a bit of a dud release to me.
>
> I guess Intel has to do something, they have to keep moving and releasing product
> which is good for everyone, but it's stunningly uninteresting and uncompetitive.
>
The AMD Epyc SKUs with the best price/performance ratio are 7313P (16 cores, 128 MB L3 cache, $913), 7443P (24 cores, 128 MB L3 cache, $1337) and 7453 (28 cores, 64 MB L3 cache, $1570).
The Xeon Gold 6312U (24 cores, 36 MB L3 cache, $1450) can be considered a direct competitor of 7443P.
So Intel is only slightly more expensive, but the main differences versus AMD are:
1. The clock frequency in AVX-512 mode is only about 2/3 of Epyc
2. The cache size is less than 1/3 of Epyc
3. The number of floating-point multipliers is double for Xeon
4. The multi-thread memory throughput for straightforward code is better for Xeon
5. With few active threads, the memory throughput is worse for Xeon (this behavior has been exhibited by all recent generations of Intel server CPUs, with only a few active threads it is not possible to saturate the memory channels, unlike in Intel client CPUs or in AMD CPUs)
There is no doubt that in general-purpose applications Epyc 7443P wins, because of higher clock frequency and bigger cache memory.
Nevertheless, in applications that depend on FP multiplication throughput or on other special AVX-512 features, Xeon 6312U wins.
Since the difference in price and performance is not large for these 2 SKUs, a buying decision can also be made based on local availability of the processors or motherboards, which is probably how I will decide when I will upgrade one of my servers later this year.
While among the cheaper Ice Lake Server SKUs are a few that are competitive, like in this 6312U example, the more expensive SKUs cannot compete in price or performance with AMD Milan so all those who will buy them will do that only due to other reasons.
> > Good:
> > The cheaper SKUs are no longer crippled by disabling 1 of the 2 FMA units.
> > No clock frequency penalty for AVX-256 and light AVX-512 instructions.
> > For heavy AVX-512 instructions, the clock frequency 80% of normal frequency.
>
> I would rephrase those 3 to "AVX-512 is a bit better" with the caveat "most things and most people
> don't care". Maybe AVX-512 is useful for futureproofing but it doesn't look that way yet.
>
> > There are a few SKUs with good price/performance, e.g. Gold 6312U.
>
> Really? The costs aren't that low and the performance isn't that good.
>
> This really seems like a bit of a dud release to me.
>
> I guess Intel has to do something, they have to keep moving and releasing product
> which is good for everyone, but it's stunningly uninteresting and uncompetitive.
>
The AMD Epyc SKUs with the best price/performance ratio are 7313P (16 cores, 128 MB L3 cache, $913), 7443P (24 cores, 128 MB L3 cache, $1337) and 7453 (28 cores, 64 MB L3 cache, $1570).
The Xeon Gold 6312U (24 cores, 36 MB L3 cache, $1450) can be considered a direct competitor of 7443P.
So Intel is only slightly more expensive, but the main differences versus AMD are:
1. The clock frequency in AVX-512 mode is only about 2/3 of Epyc
2. The cache size is less than 1/3 of Epyc
3. The number of floating-point multipliers is double for Xeon
4. The multi-thread memory throughput for straightforward code is better for Xeon
5. With few active threads, the memory throughput is worse for Xeon (this behavior has been exhibited by all recent generations of Intel server CPUs, with only a few active threads it is not possible to saturate the memory channels, unlike in Intel client CPUs or in AMD CPUs)
There is no doubt that in general-purpose applications Epyc 7443P wins, because of higher clock frequency and bigger cache memory.
Nevertheless, in applications that depend on FP multiplication throughput or on other special AVX-512 features, Xeon 6312U wins.
Since the difference in price and performance is not large for these 2 SKUs, a buying decision can also be made based on local availability of the processors or motherboards, which is probably how I will decide when I will upgrade one of my servers later this year.
While among the cheaper Ice Lake Server SKUs are a few that are competitive, like in this 6312U example, the more expensive SKUs cannot compete in price or performance with AMD Milan so all those who will buy them will do that only due to other reasons.