3 days to Epyc Zen3 launch--where is the price war b/w Xeon and Epyc???

[mjlitola]

I have been intently following the prices of both Intel Xeon and AMD Epyc processors in the eve of Zen3 launch of Feb 15th. Xeon has an ability to have a maximum of 4P in one Supermicro X12 or X11 motherboard (they are said to be quite hard to boot and quite expensive Supermicro X11QPH+ costs 1268 euros VAT 0%). Epycs have more cores (max 64 cores per processor) but they are limited to 2P motherboards such as Gigabyte MZ72-HB0 for 772,73 euros VAT 0%. The Zen3s are launching Monday 15th of February 2021, so is there any feeling of PRICE WAR between Intel and AMD. The prices were collected from Alibaba.com and Ebay.de.

 

[cdabc123]

No need. The products hit a reasonable price point already. Intel is Intel so they have 0 incentive in lowering xeon prices and amd may be facing supply issues making it pointless for them to do the same.

 

[mjlitola]

I would think that AMD's chip shortage is not as bad as people are saying. If we estimate that an average Xeon price is ab. USD 1000 (in the server market it was USD 538 in 2015), then Intel sells ab. 5 million Xeon chip per year (with USD 5,91 Billion server sales https://www.thestreet.com/investing...orting-lower-server-cpu-sales-8-key-takeaways). In comparison AMD shipped 1 million Ryzen 5 chips in Q4/2020, so we might easily see some price war. It is better to sell some chips with an average of USD 538 than almost nothing at all with USD 1000 a pop.

 

[cdabc123]

I would think that AMD's chip shortage is not as bad as people are saying. If we estimate that an average Xeon price is ab. USD 1000 (in the server market it was USD 538 in 2015), then Intel sells ab. 5 million Xeon chip per year (with USD 5,91 Billion server sales https://www.thestreet.com/investing...orting-lower-server-cpu-sales-8-key-takeaways). In comparison AMD shipped 1 million Ryzen 5 chips in Q4/2020, so we might easily see some price war. It is better to sell some chips with an average of USD 538 than almost nothing at all with USD 1000 a pop.

This market is not as free as you believe. Intel can charge what they want and companies will still buy xeons in mass. Amd can make their product slightly more competitive then Intel as far as price/prefermance and sell as many as they can fab (as they are doing)

I don't belive we will see much of a price war

 

[mjlitola]

It is true that server market price elasticity of demand is fairly inelastic (according to 2015 paper server market price elasticity is -0.51 http://www.qyjohn.net/wp-content/uploads/2015/10/price-elasticity-enterprise.pdf ); it means that any price increase of 1 percent, the effect of demand is only 50% of that > demand is lowered by 0,5%. However, the previous prices of server processors (i.e. Intel Skylake) the price of Xeon 8180 was ab. USD 13000, where in today's prices you can get ab. 2x more powerful part for 15% of that price (that is USD 1995) > When the price differences (and computing power differences in AMD's favor) are so great > EVEN AN INELASTIC PRICE HAS TO TAKE INTO ACCOUNT THE COMPETITOR'S PRICES. Example was given in wikipedia of inelastic prices of cigarettes, i.e. a smoker will smoke his/her brand no matter the price--even the cigarettes have a price elasticity of demand of -0,3; so the cigarettes are even more inelastic in their demand as server chips > every price will adjust to the changes in their demand. Intel would be out of business, if it was not so in their server processor market.

 

[bwang]

1. Hyperscale customers don't pay list price for CPUs
2. Passmark is not a valid metric of server performance (Passmark isn't really a valid metric for anything)
3. Price-per-unit-Passmark is not a valid measure of anything - datacenters are built to volumetric and power constraints, not optimized for performance per dollar
4. For enterprise parts, the cost of the CPU is a small part of overall costs. If the parts are going into a datacenter, storage, infrastructure, staff, and power all add up to much more than the cost of the CPU. If the parts are going into a workstation, the salary of the employee using the workstation is much more than the cost of the workstation.

Of course, for hobbyists, all of this is untrue - hobbyists usually build to a fixed budget, have no ROI considerations, and frequently aren't using their workstations to make money. But hobbyists also don't dictate the pricing of datacenter CPUs, not significantly at least.

 

[learners permit]

I'm fairly certain supply constraints are a damper for any real competition at this particular time.

 

[mjlitola]

1. Hyperscale customers don't pay list price for CPUs

In article https://www.servethehome.com/intel-is-serving-major-xeon-discounts-to-combat-amd-epyc/ Patrick Kennedy claims that Intel is matching the price of Xeon to correspond AMD Epyc prices and this comes well below a thousand parts ordered. I agree with you that life cycle costs of a processor are more relevant than price. For example if you run Epyc 7763 (64c TDP 280W) vs Xeon Platinum 9282 (56c TDP 400W) your 3-year running cost of electricity of that power difference is 0,12 KWh x 24h/day x 365 days/year x 3 years x USD 0,12 per KWh = USD 380 in AMD's favor even if Intel matches the price of Xeon Platinum 9282 with that of Epyc 7763.

2. Passmark is not a valid metric of server performance (Passmark isn't really a valid metric for anything)
3. Price-per-unit-Passmark is not a valid measure of anything - datacenters are built to volumetric and power constraints, not optimized for performance per dollar

I agree, but since I do not find widely used benchmark reported results with various Epyc and Xeon parts I am out of luck. I agree that for example data mining (if you do it from public domain like facebook [which puts up time pauses as a deterrant for mining]), it is (not the speed of the internet connection nor speed of the processor, but) the pure number of cores and number of your VMs that is doing the waiting that does the job > I would say that Epyc 7763 has more cores than Xeon Platinum 9282

4. For enterprise parts, the cost of the CPU is a small part of overall costs. If the parts are going into a datacenter, storage, infrastructure, staff, and power all add up to much more than the cost of the CPU. If the parts are going into a workstation, the salary of the employee using the workstation is much more than the cost of the workstation.

I would think that almost all of the parts (in exclusion of motherboard) i.e. processor and rack cooling system, hard drives, server racks with their attached hotswappable drive bays and ECC DDR4 memory are interchangeable between Xeon and Epyc. I am talking about server and not work station parts, so I do not go to salary. I would think that an employee would know how to use a same program in the same operating system was it Linux, Win10, Mac osx etc.

Of course, for hobbyists, all of this is untrue - hobbyists usually build to a fixed budget, have no ROI considerations, and frequently aren't using their workstations to make money. But hobbyists also don't dictate the pricing of datacenter CPUs, not significantly at least.